VM.TensorFlow Preview

Dragon/python/dragon/vm/caffe/net.py

@@ -550,7 +550,7 @@ class Net(object):
         The implementation of `Net_outputs(pycaffe.py, L81)`_.
         The implementation of `Net_outputs(pycaffe.py, L81)`_.
 
 
         """
         """
-        return self._net_outputs
+        return list(self._net_outputs)
 
 
 
 
     def replace(self, A, B):
     def replace(self, A, B):
@@ -571,8 +571,8 @@ class Net(object):
         --------
         --------
         >>> import dragon.ops as ops
         >>> import dragon.ops as ops
         >>> data, label = ops.LMDBData()
         >>> data, label = ops.LMDBData()
-        >>> net.Replace(net.blobs['data'].data, data)
-        >>> net.Replace(net.blobs['label'].data, label)
+        >>> net.replace(net.blobs['data'].data, data)
+        >>> net.replace(net.blobs['label'].data, label)
 
 
         """
         """
         self._swap_tensors[A] = B
         self._swap_tensors[A] = B

Dragon/python/dragon/vm/tensorflow/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+# Framework
+from dragon.vm.tensorflow.framework.framework_lib import *
+
+# Session
+from dragon.vm.tensorflow.client.client_lib import *
+
+# Ops
+from dragon.vm.tensorflow.ops.standard_ops import *
+
+# Bring in subpackages.
+from dragon.vm.tensorflow.ops import nn
+
+# Import the names from training.py as train.Name.
+from dragon.vm.tensorflow.training import training as train
+
+#from .utils.gradients import *
+
+# Export modules and constants
+from dragon.vm.tensorflow.layers import layers
+
+

Dragon/python/dragon/vm/tensorflow/client/client_lib.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.client.session import Session
+
+from dragon.vm.tensorflow.client.session import get_default_session
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/client/session.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import warnings
+
+from dragon.core.tensor import Tensor
+import dragon.vm.theano as theano
+
+from dragon.vm.tensorflow.protobuf import config_pb2
+from dragon.vm.tensorflow.training.optimizer import Optimizer
+from dragon.vm.tensorflow.ops.variables import VariablesInitializer
+from dragon.vm.tensorflow.framework import ops
+
+
+_TRANSACTIONS = {}
+
+
+class Transaction(object):
+    def __init__(self, functions):
+        self.functions = functions
+
+    def run(self, feed_values=None):
+        for i, function in enumerate(self.functions):
+            if i == 0 and feed_values is not None:
+                function(*feed_values, return_outputs=False)
+            else: function(return_outputs=False)
+
+_default_session = None
+
+class BaseSession(object):
+    """
+    Construct a BaseSession.
+    """
+    def __init__(self, target='', graph=None, config=None):
+        if graph is None:
+            self._graph = ops.get_default_graph()
+        else:
+            raise NotImplementedError('Session can only use the default graph yet.')
+
+        self._opened = False
+        self._closed = False
+
+        if config is not None:
+            if not isinstance(config, config_pb2.ConfigProto):
+                raise TypeError('Config should be a tf.ConfigProto, but got {}'.format(type(config)))
+            self._config = config
+            self._add_shapes = config.graph_options.infer_shapes
+        else:
+            self._config = None
+            self._add_shapes = False
+
+        self._session = None
+
+    def list_devices(self):
+        """Lists available devices in this session.
+
+        ```python
+        devices = sess.list_devices()
+        for d in devices:
+          print(d.name)
+        ```
+
+        Each element in the list has the following properties:
+         - `name`: A string with the full name of the device. ex:
+              `/job:worker/replica:0/task:3/device:CPU:0`
+         - `device_type`: The type of the device (e.g. `CPU`, `GPU`, `TPU`.)
+         - `memory_limit`: The maximum amount of memory available on the device.
+              Note: depending on the device, it is possible the usable memory could
+              be substantially less.
+        Raises:
+          tf.errors.OpError: If it encounters an error (e.g. session is in an
+          invalid state, or network errors occur).
+
+        Returns:
+          A list of devices in the session.
+        """
+        # with errors.raise_exception_on_not_ok_status() as status:
+        #     if self._created_with_new_api:
+        #         raw_device_list = tf_session.TF_SessionListDevices(
+        #             self._session, status)
+        #     else:
+        #         raw_device_list = tf_session.TF_DeprecatedSessionListDevices(
+        #             self._session, status)
+        #     device_list = []
+        #     size = tf_session.TF_DeviceListCount(raw_device_list)
+        #     for i in range(size):
+        #         name = tf_session.TF_DeviceListName(raw_device_list, i, status)
+        #         device_type = tf_session.TF_DeviceListType(raw_device_list, i, status)
+        #         memory = tf_session.TF_DeviceListMemoryBytes(raw_device_list, i, status)
+        #         device_list.append(_DeviceAttributes(name, device_type, memory))
+        #     tf_session.TF_DeleteDeviceList(raw_device_list)
+        #     return device_list
+
+    def close(self):
+        pass
+
+    @property
+    def graph(self):
+        return self._graph
+
+    @property
+    def graph_def(self):
+        return ''
+
+    @property
+    def sess_str(self):
+        return ''
+
+    def as_default(self):
+        pass
+
+    def run(self, fetches, feed_dict=None, options=None, run_metadata=None):
+        try:
+            if options is not None:
+                warnings.warn(Warning('Ignore Arguments: <options>.'))
+            if run_metadata is not None:
+                warnings.warn(Warning('Ignore Arguments: <run_metadata>.'))
+        except Warning:
+            pass
+
+        if not isinstance(fetches, list): fetches = [fetches]
+        if len(fetches) < 1: return None
+        return self._run(fetches, feed_dict)
+
+    def _run(self, fetches, feed_dict):
+        if self._closed:
+            raise RuntimeError('Attempted to use a closed Session.')
+
+        # unpack opts and tensors
+        opts = []; tensors = []
+        for target in fetches:
+            if isinstance(target, Optimizer): opts.append(target)
+            elif isinstance(target, VariablesInitializer): tensors.extend(target.var_list)
+            elif isinstance(target, Tensor): tensors.append(target)
+
+        # find minimum solving targets
+        targets = set()
+        for t in tensors: targets.add(t)
+        for opt in opts:
+            for t in opt.objs: targets.add(t)
+
+        targets = list(targets)
+
+        # if existing a transaction before?
+        global _TRANSACTIONS
+        t_key = tuple(fetches + feed_dict.keys()) \
+                if feed_dict is not None else tuple(fetches)
+        transaction = None if not t_key in _TRANSACTIONS else _TRANSACTIONS[t_key]
+
+        # cond.1: run by feeding
+        if feed_dict is not None:
+            # checking
+            for key, value in feed_dict.items():
+                if not isinstance(key, Tensor):
+                    raise TypeError('The key of feed_dict key should be a Tensor.')
+                if key.shape is not None:
+                    if len(key.shape) != len(value.shape):
+                        raise RuntimeError('The Tensor({}) was limited to {} dimensions, \
+                                            while feed a value with {} dimensions.'.
+                                                format(key.name, len(key.shape), len(value.shape)))
+                    for i in xrange(len(key.shape)):
+                        if key.shape[i] is None: continue
+                        if key.shape[i] != value.shape[i]:
+                            raise RuntimeError('The shape of Tensor({}) was limited as ('.format(key.name) +
+                                               ','.join([str(dim) for dim in key.shape]) + '), ' +
+                                               'while feed a value with (' + ','.join([str(dim) for dim in value.shape]) + ').')
+            # create a new transaction
+            if transaction is None:
+                functions = []
+                functions.append(theano.function(inputs=feed_dict.keys(), outputs=targets))
+                for opt in opts:
+                    functions.append(theano.function(updater=opt.updater))
+                _TRANSACTIONS[t_key] = transaction = Transaction(functions)
+            transaction.run(feed_dict.values())
+
+        # cond.2: run without feeding
+        else:
+            # create a new transaction
+            if transaction is None:
+                functions = []
+                functions.append(theano.function(outputs=targets))
+                for opt in opts:
+                    functions.append(theano.function(updater=opt.updater))
+                _TRANSACTIONS[t_key] = transaction = Transaction(functions)
+            transaction.run(None)
+
+        # fetch after running
+        returns = []
+        for target in fetches:
+            if isinstance(target, Optimizer): returns.append(None)
+            elif isinstance(target, VariablesInitializer): returns.append(None)
+            else:
+                np_target = target.get_value()
+                # unpack the scalar if necessary
+                if np_target.size == 1:
+                    returns.append(np_target.flatten()[0])
+                else:
+                    returns.append(np_target)
+
+        # unpack the returns if necessary
+        if len(returns) == 1: return returns[0]
+        else: return returns
+
+
+class Session(BaseSession):
+    """
+    Construct a Session.
+    """
+    def __init__(self, target='', graph=None, config=None):
+        super(Session, self).__init__(target, graph, config=config)
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exec_type, exec_value, exec_tb):
+        pass
+
+    @staticmethod
+    def reset(target, containers=None, config=None):
+        pass
+
+
+class InteractiveSession(BaseSession):
+    """
+    Construct a InteractiveSession.
+    """
+    def __init__(self, target='', graph=None, config=None):
+        super(InteractiveSession, self).__init__(target, graph, config=config)
+
+    def __enter__(self):
+        pass
+
+    def __exit__(self, exec_type, exec_value, exec_tb):
+        pass
+
+    @staticmethod
+    def reset(target, containers=None, config=None):
+        pass
+
+
+def get_default_session():
+    global _default_session
+    if _default_session is None:
+        _default_session = Session()
+    return _default_session
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/contrib/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from .layers import *
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/contrib/framework/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/contrib/framework/ops/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/contrib/framework/ops/variables.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.framework import ops
+from dragon.vm.tensorflow.ops import var_scope as variable_scope
+
+
+def get_variables(scope=None, suffix=None,
+                  collection=ops.GraphKeys.GLOBAL_VARIABLES):
+    if isinstance(scope, variable_scope.VariableScope):
+        scope = scope.name
+    if suffix is not None:
+        scope = (scope or '') + '.*' + suffix
+    return ops.get_collection(collection, scope)
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/contrib/layers/__init__.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from .layers import *
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/contrib/layers/initializers.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import math
+
+from dragon.vm.tensorflow.framework import dtypes
+from dragon.vm.tensorflow.ops import random_ops
+
+__all__ = ['xavier_initializer',
+           'xavier_initializer_conv2d',
+           'variance_scaling_initializer']
+
+
+def xavier_initializer(uniform=True, seed=None, dtype=dtypes.float32):
+    return variance_scaling_initializer(factor=1.0, mode='FAN_AVG',
+                                        uniform=uniform, seed=seed, dtype=dtype)
+
+
+xavier_initializer_conv2d = xavier_initializer
+
+
+def variance_scaling_initializer(factor=2.0, mode='FAN_IN', uniform=False,
+                                 seed=None, dtype=dtypes.float32):
+    if not dtype.is_floating:
+        raise TypeError('Cannot create initializer for non-floating point type.')
+    if mode not in ['FAN_IN', 'FAN_OUT', 'FAN_AVG']:
+        raise TypeError('Unknow mode %s [FAN_IN, FAN_OUT, FAN_AVG]', mode)
+
+    def _initializer(shape, dtype=dtype, partition_info=None):
+        """Initializer function."""
+        if not dtype.is_floating:
+            raise TypeError('Cannot create initializer for non-floating point type.')
+        # Estimating fan_in and fan_out is not possible to do perfectly, but we try.
+        # This is the right thing for matrix multiply and convolutions.
+        if shape:
+            fan_in = float(shape[-2]) if len(shape) > 1 else float(shape[-1])
+            fan_out = float(shape[-1])
+        else:
+            fan_in = 1.0
+            fan_out = 1.0
+        for dim in shape[:-2]:
+            fan_in *= float(dim)
+            fan_out *= float(dim)
+        if mode == 'FAN_IN':
+            # Count only number of input connections.
+            n = fan_in
+        elif mode == 'FAN_OUT':
+            # Count only number of output connections.
+            n = fan_out
+        elif mode == 'FAN_AVG':
+            # Average number of inputs and output connections.
+            n = (fan_in + fan_out) / 2.0
+        if uniform:
+            # To get stddev = math.sqrt(factor / n) need to adjust for uniform.
+            limit = math.sqrt(3.0 * factor / n)
+            return random_ops.random_uniform(shape, -limit, limit,
+                                             dtype, seed=seed)
+        else:
+            # To get stddev = math.sqrt(factor / n) need to adjust for truncated.
+            trunc_stddev = math.sqrt(1.3 * factor / n)
+            return random_ops.truncated_normal(shape, 0.0, trunc_stddev, dtype,
+                                               seed=seed)
+
+    return _initializer

Dragon/python/dragon/vm/tensorflow/contrib/layers/layers.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from collections import defaultdict
+
+import dragon.ops as op_lib
+
+import dragon.vm.tensorflow.framework.ops as ops
+from dragon.vm.tensorflow.contrib.layers import initializers
+from dragon.vm.tensorflow.ops import init_ops
+from dragon.vm.tensorflow.ops import nn
+from dragon.vm.tensorflow.ops import var_scope as vs
+from dragon.vm.tensorflow.layers import layers
+from dragon.vm.tensorflow.layers import normalization as normalization_layers
+
+__all__ = ['flatten']
+
+_LAYERS_UID_DICT = defaultdict(int)
+
+DATA_FORMAT_NCHW = 'NCHW'
+DATA_FORMAT_NHWC = 'NHWC'
+DATA_FORMAT_NCDHW = 'NCDHW'
+DATA_FORMAT_NDHWC = 'NDHWC'
+
+
+def _default_scope(scope, key, indicator):
+    if scope is None:
+        return indicator
+        # global _LAYERS_UID_DICT
+        # _LAYERS_UID_DICT[key] += 1
+        # return '{}{}'.format(indicator, _LAYERS_UID_DICT[key])
+    else:
+        return scope
+
+
+def avg_pool2d(inputs,
+               kernel_size,
+               stride=2,
+               padding='VALID',
+               data_format=DATA_FORMAT_NHWC,
+               outputs_collections=None,
+               scope=None):
+    if data_format not in (DATA_FORMAT_NCHW, DATA_FORMAT_NHWC):
+        raise ValueError('data_format has to be either NCHW or NHWC.')
+    df = ('channels_first' if data_format and data_format.startswith('NC')
+          else 'channels_last')
+    return layers.average_pooling2d(inputs=inputs,
+                                    pool_size=kernel_size,
+                                    strides=stride,
+                                    padding=padding,
+                                    data_format=df)
+
+
+def max_pool2d(inputs,
+               kernel_size,
+               stride=2,
+               padding='VALID',
+               data_format=DATA_FORMAT_NHWC,
+               outputs_collections=None,
+               scope=None):
+    if data_format not in (DATA_FORMAT_NCHW, DATA_FORMAT_NHWC):
+        raise ValueError('data_format has to be either NCHW or NHWC.')
+    df = ('channels_first' if data_format and data_format.startswith('NC')
+          else 'channels_last')
+    return layers.max_pooling2d(inputs=inputs,
+                                pool_size=kernel_size,
+                                strides=stride,
+                                padding=padding,
+                                data_format=df)
+
+
+def convolution(inputs,
+                num_outputs,
+                kernel_size,
+                stride=1,
+                padding='SAME',
+                data_format=None,
+                rate=1,
+                activation_fn=nn.relu,
+                normalizer_fn=None,
+                normalizer_params=None,
+                weights_initializer=initializers.xavier_initializer(),
+                weights_regularizer=None,
+                biases_initializer=init_ops.zeros_initializer(),
+                biases_regularizer=None,
+                reuse=None,
+                variables_collections=None,
+                outputs_collections=None,
+                trainable=True,
+                scope=None):
+    scope = _default_scope(scope, 'CONVOLUTION', 'Conv')
+    if data_format not in [None, 'NHWC', 'NCHW']:
+        raise ValueError('Invalid data_format: %r' % (data_format,))
+    data_format = 'channels_first' if data_format == 'NCHW' else 'channels_last'
+    input_rank = inputs.get_shape().ndims
+
+    with vs.variable_scope(scope, reuse=reuse) as sc:
+        if input_rank == 4:
+            return layers.conv2d(
+                inputs=inputs,
+                filters=num_outputs,
+                kernel_size=kernel_size,
+                strides=stride,
+                padding=padding,
+                data_format=data_format,
+                dilation_rate=rate,
+                activation=activation_fn,
+                use_bias=True if biases_initializer is not None else False,
+                kernel_initializer=weights_initializer,
+                bias_initializer=biases_initializer,
+                bias_regularizer=biases_regularizer,
+                activity_regularizer=None,
+                trainable=trainable,
+                reuse=reuse)
+
+
+# Simple alias.
+convolution2d = convolution
+conv2d = convolution2d
+
+
+def fully_connected(inputs,
+                    num_outputs,
+                    activation_fn=nn.relu,
+                    normalizer_fn=None,
+                    normalizer_params=None,
+                    weights_initializer=initializers.xavier_initializer(),
+                    weights_regularizer=None,
+                    biases_initializer=init_ops.zeros_initializer(),
+                    biases_regularizer=None,
+                    reuse=None,
+                    variables_collections=None,
+                    outputs_collections=None,
+                    trainable=True,
+                    scope=None):
+    scope = _default_scope(scope, 'FULLY_CONNECTED', 'fully_connected')
+    with vs.variable_scope(scope, reuse=reuse) as sc:
+        return layers.dense(
+            inputs=inputs,
+            units=num_outputs,
+            activation=activation_fn,
+            use_bias=True if biases_initializer is not None else False,
+            kernel_initializer=weights_initializer,
+            bias_initializer=biases_initializer,
+            bias_regularizer=biases_regularizer,
+            activity_regularizer=None,
+            trainable=trainable,
+            reuse=reuse)
+
+
+def batch_norm(inputs,
+               decay=0.999,
+               center=True,
+               scale=False,
+               epsilon=0.001,
+               activation_fn=None,
+               param_initializers=None,
+               param_regularizers=None,
+               updates_collections=ops.GraphKeys.UPDATE_OPS,
+               is_training=True,
+               reuse=None,
+               variables_collections=None,
+               outputs_collections=None,
+               trainable=True,
+               batch_weights=None,
+               fused=False,
+               data_format=DATA_FORMAT_NHWC,
+               zero_debias_moving_mean=False,
+               scope=None,
+               renorm=False,
+               renorm_clipping=None,
+               renorm_decay=0.99):
+    scope = _default_scope(scope, 'BATCH_NORM', 'BatchNorm')
+    if data_format not in (DATA_FORMAT_NCHW, DATA_FORMAT_NHWC):
+        raise ValueError('data_format has to be either NCHW or NHWC.')
+    axis = 1 if data_format == DATA_FORMAT_NCHW else -1
+
+    with vs.variable_scope(scope, reuse=reuse) as sc:
+        if not param_initializers:
+            param_initializers = {}
+        beta_initializer = param_initializers.get('beta', init_ops.zeros_initializer())
+        gamma_initializer = param_initializers.get('gamma', init_ops.ones_initializer())
+        moving_mean_initializer = param_initializers.get('moving_mean', init_ops.zeros_initializer())
+        moving_variance_initializer = param_initializers.get('moving_variance', init_ops.ones_initializer())
+
+        if not param_regularizers:
+            param_regularizers = {}
+
+        beta_regularizer = param_regularizers.get('beta')
+        gamma_regularizer = param_regularizers.get('gamma')
+
+        return layers.batch_norm(
+            inputs=inputs,
+            axis=axis,
+            momentum=decay,
+            epsilon=epsilon,
+            center=center,
+            scale=scale,
+            beta_initializer=beta_initializer,
+            gamma_initializer=gamma_initializer,
+            moving_mean_initializer=moving_mean_initializer,
+            moving_variance_initializer=moving_variance_initializer,
+            beta_regularizer=beta_regularizer,
+            gamma_regularizer=gamma_regularizer,
+            trainable=trainable,
+            renorm=renorm,
+            renorm_clipping=renorm_clipping,
+            renorm_momentum=renorm_decay,
+            fused=fused,
+            training=is_training)
+
+
+def flatten(inputs,
+            outputs_collections=None,
+            scope=None):
+    return op_lib.Flatten(inputs, axis=0, keep_axes=2)
+

Dragon/python/dragon/vm/tensorflow/contrib/learn/datasets/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/contrib/learn/datasets/base.py

+# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Base utilities for loading datasets."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import os
+import random
+import time
+
+import shutil
+from six.moves import urllib
+
+Dataset = collections.namedtuple('Dataset', ['data', 'target'])
+Datasets = collections.namedtuple('Datasets', ['train', 'validation', 'test'])
+
+
+def retry(initial_delay,
+          max_delay,
+          factor=2.0,
+          jitter=0.25,
+          is_retriable=None):
+    """Simple decorator for wrapping retriable functions.
+
+    Args:
+      initial_delay: the initial delay.
+      factor: each subsequent retry, the delay is multiplied by this value.
+          (must be >= 1).
+      jitter: to avoid lockstep, the returned delay is multiplied by a random
+          number between (1-jitter) and (1+jitter). To add a 20% jitter, set
+          jitter = 0.2. Must be < 1.
+      max_delay: the maximum delay allowed (actual max is
+          max_delay * (1 + jitter).
+      is_retriable: (optional) a function that takes an Exception as an argument
+          and returns true if retry should be applied.
+    """
+    if factor < 1:
+        raise ValueError('factor must be >= 1; was %f' % (factor,))
+
+    if jitter >= 1:
+        raise ValueError('jitter must be < 1; was %f' % (jitter,))
+
+    # Generator to compute the individual delays
+    def delays():
+        delay = initial_delay
+        while delay <= max_delay:
+            yield delay * random.uniform(1 - jitter, 1 + jitter)
+            delay *= factor
+
+    def wrap(fn):
+        """Wrapper function factory invoked by decorator magic."""
+
+        def wrapped_fn(*args, **kwargs):
+            """The actual wrapper function that applies the retry logic."""
+            for delay in delays():
+                try:
+                    return fn(*args, **kwargs)
+                except Exception as e:  # pylint: disable=broad-except)
+                    if is_retriable is None:
+                        continue
+
+                    if is_retriable(e):
+                        time.sleep(delay)
+                    else:
+                        raise
+            return fn(*args, **kwargs)
+
+        return wrapped_fn
+
+    return wrap
+
+
+_RETRIABLE_ERRNOS = {
+    110,  # Connection timed out [socket.py]
+}
+
+
+def _is_retriable(e):
+    return isinstance(e, IOError) and e.errno in _RETRIABLE_ERRNOS
+
+
+@retry(initial_delay=1.0, max_delay=16.0, is_retriable=_is_retriable)
+def urlretrieve_with_retry(url, filename=None):
+    return urllib.request.urlretrieve(url, filename)
+
+
+def maybe_download(filename, work_directory, source_url):
+    """Download the data from source url, unless it's already here.
+
+    Args:
+        filename: string, name of the file in the directory.
+        work_directory: string, path to working directory.
+        source_url: url to download from if file doesn't exist.
+
+    Returns:
+        Path to resulting file.
+    """
+    if not os.path.exists(work_directory):
+        os.makedirs(work_directory)
+    filepath = os.path.join(work_directory, filename)
+    if not os.path.exists(filepath):
+        temp_file_name, _ = urlretrieve_with_retry(source_url)
+        shutil.copy(temp_file_name, filepath)
+        size = os.path.getsize(filepath)
+        print('Successfully downloaded', filename, size, 'bytes.')
+    return filepath

Dragon/python/dragon/vm/tensorflow/contrib/learn/datasets/mnist.py

+# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Functions for downloading and reading MNIST data."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import gzip
+
+import numpy
+from six.moves import xrange  # pylint: disable=redefined-builtin
+
+from dragon.vm.tensorflow.contrib.learn.datasets import base
+from dragon.vm.tensorflow.framework import dtypes
+from dragon.vm.tensorflow.framework import random_seed
+
+# CVDF mirror of http://yann.lecun.com/exdb/mnist/
+# SOURCE_URL = 'https://storage.googleapis.com/cvdf-datasets/mnist/'
+SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/'
+
+
+def _read32(bytestream):
+    dt = numpy.dtype(numpy.uint32).newbyteorder('>')
+    return numpy.frombuffer(bytestream.read(4), dtype=dt)[0]
+
+
+def extract_images(f):
+    """Extract the images into a 4D uint8 numpy array [index, y, x, depth].
+
+    Args:
+      f: A file object that can be passed into a gzip reader.
+
+    Returns:
+      data: A 4D uint8 numpy array [index, y, x, depth].
+
+    Raises:
+      ValueError: If the bytestream does not start with 2051.
+
+    """
+    print('Extracting', f.name)
+    with gzip.GzipFile(fileobj=f) as bytestream:
+        magic = _read32(bytestream)
+        if magic != 2051:
+            raise ValueError('Invalid magic number %d in MNIST image file: %s' %
+                             (magic, f.name))
+        num_images = _read32(bytestream)
+        rows = _read32(bytestream)
+        cols = _read32(bytestream)
+        buf = bytestream.read(rows * cols * num_images)
+        data = numpy.frombuffer(buf, dtype=numpy.uint8)
+        data = data.reshape(num_images, rows, cols, 1)
+        return data
+
+
+def dense_to_one_hot(labels_dense, num_classes):
+    """Convert class labels from scalars to one-hot vectors."""
+    num_labels = labels_dense.shape[0]
+    index_offset = numpy.arange(num_labels) * num_classes
+    labels_one_hot = numpy.zeros((num_labels, num_classes))
+    labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1
+    return labels_one_hot
+
+
+def extract_labels(f, one_hot=False, num_classes=10):
+    """Extract the labels into a 1D uint8 numpy array [index].
+
+    Args:
+      f: A file object that can be passed into a gzip reader.
+      one_hot: Does one hot encoding for the result.
+      num_classes: Number of classes for the one hot encoding.
+
+    Returns:
+      labels: a 1D uint8 numpy array.
+
+    Raises:
+      ValueError: If the bystream doesn't start with 2049.
+    """
+    print('Extracting', f.name)
+    with gzip.GzipFile(fileobj=f) as bytestream:
+        magic = _read32(bytestream)
+        if magic != 2049:
+            raise ValueError('Invalid magic number %d in MNIST label file: %s' %
+                             (magic, f.name))
+        num_items = _read32(bytestream)
+        buf = bytestream.read(num_items)
+        labels = numpy.frombuffer(buf, dtype=numpy.uint8)
+        if one_hot:
+            return dense_to_one_hot(labels, num_classes)
+        return labels
+
+
+class DataSet(object):
+    def __init__(self,
+                 images,
+                 labels,
+                 fake_data=False,
+                 one_hot=False,
+                 dtype=dtypes.float32,
+                 reshape=True,
+                 seed=None):
+        """Construct a DataSet.
+        one_hot arg is used only if fake_data is true.  `dtype` can be either
+        `uint8` to leave the input as `[0, 255]`, or `float32` to rescale into
+        `[0, 1]`.  Seed arg provides for convenient deterministic testing.
+        """
+        seed1, seed2 = random_seed.get_seed(seed)
+        # If op level seed is not set, use whatever graph level seed is returned
+        numpy.random.seed(seed1 if seed is None else seed2)
+        dtype = dtypes.as_dtype(dtype).base_dtype
+        if dtype not in (dtypes.uint8, dtypes.float32):
+            raise TypeError('Invalid image dtype %r, expected uint8 or float32' %
+                            dtype)
+        if fake_data:
+            self._num_examples = 10000
+            self.one_hot = one_hot
+        else:
+            assert images.shape[0] == labels.shape[0], (
+                'images.shape: %s labels.shape: %s' % (images.shape, labels.shape))
+            self._num_examples = images.shape[0]
+
+            # Convert shape from [num examples, rows, columns, depth]
+            # to [num examples, rows*columns] (assuming depth == 1)
+            if reshape:
+                assert images.shape[3] == 1
+                images = images.reshape(images.shape[0],
+                                        images.shape[1] * images.shape[2])
+            if dtype == dtypes.float32:
+                # Convert from [0, 255] -> [0.0, 1.0].
+                images = images.astype(numpy.float32)
+                images = numpy.multiply(images, 1.0 / 255.0)
+        self._images = images
+        self._labels = labels
+        self._epochs_completed = 0
+        self._index_in_epoch = 0
+
+    @property
+    def images(self):
+        return self._images
+
+    @property
+    def labels(self):
+        return self._labels
+
+    @property
+    def num_examples(self):
+        return self._num_examples
+
+    @property
+    def epochs_completed(self):
+        return self._epochs_completed
+
+    def next_batch(self, batch_size, fake_data=False, shuffle=True):
+        """Return the next `batch_size` examples from this data set."""
+        if fake_data:
+            fake_image = [1] * 784
+            if self.one_hot:
+                fake_label = [1] + [0] * 9
+            else:
+                fake_label = 0
+            return [fake_image for _ in xrange(batch_size)], [
+                fake_label for _ in xrange(batch_size)
+                ]
+        start = self._index_in_epoch
+        # Shuffle for the first epoch
+        if self._epochs_completed == 0 and start == 0 and shuffle:
+            perm0 = numpy.arange(self._num_examples)
+            numpy.random.shuffle(perm0)
+            self._images = self.images[perm0]
+            self._labels = self.labels[perm0]
+        # Go to the next epoch
+        if start + batch_size > self._num_examples:
+            # Finished epoch
+            self._epochs_completed += 1
+            # Get the rest examples in this epoch
+            rest_num_examples = self._num_examples - start
+            images_rest_part = self._images[start:self._num_examples]
+            labels_rest_part = self._labels[start:self._num_examples]
+            # Shuffle the data
+            if shuffle:
+                perm = numpy.arange(self._num_examples)
+                numpy.random.shuffle(perm)
+                self._images = self.images[perm]
+                self._labels = self.labels[perm]
+            # Start next epoch
+            start = 0
+            self._index_in_epoch = batch_size - rest_num_examples
+            end = self._index_in_epoch
+            images_new_part = self._images[start:end]
+            labels_new_part = self._labels[start:end]
+            return numpy.concatenate((images_rest_part, images_new_part), axis=0), numpy.concatenate((labels_rest_part, labels_new_part), axis=0)
+        else:
+            self._index_in_epoch += batch_size
+            end = self._index_in_epoch
+            return self._images[start:end], self._labels[start:end]
+
+
+def read_data_sets(train_dir,
+                   fake_data=False,
+                   one_hot=False,
+                   dtype=dtypes.float32,
+                   reshape=True,
+                   validation_size=5000,
+                   seed=None):
+    if fake_data:
+        def fake():
+            return DataSet(
+                [], [], fake_data=True, one_hot=one_hot, dtype=dtype, seed=seed)
+
+        train = fake()
+        validation = fake()
+        test = fake()
+        return base.Datasets(train=train, validation=validation, test=test)
+
+    TRAIN_IMAGES = 'train-images-idx3-ubyte.gz'
+    TRAIN_LABELS = 'train-labels-idx1-ubyte.gz'
+    TEST_IMAGES = 't10k-images-idx3-ubyte.gz'
+    TEST_LABELS = 't10k-labels-idx1-ubyte.gz'
+
+    local_file = base.maybe_download(TRAIN_IMAGES, train_dir,
+                                     SOURCE_URL + TRAIN_IMAGES)
+    with open(local_file, 'rb') as f:
+        train_images = extract_images(f)
+
+    local_file = base.maybe_download(TRAIN_LABELS, train_dir,
+                                     SOURCE_URL + TRAIN_LABELS)
+    with open(local_file, 'rb') as f:
+        train_labels = extract_labels(f, one_hot=one_hot)
+
+    local_file = base.maybe_download(TEST_IMAGES, train_dir,
+                                     SOURCE_URL + TEST_IMAGES)
+    with open(local_file, 'rb') as f:
+        test_images = extract_images(f)
+
+    local_file = base.maybe_download(TEST_LABELS, train_dir,
+                                     SOURCE_URL + TEST_LABELS)
+    with open(local_file, 'rb') as f:
+        test_labels = extract_labels(f, one_hot=one_hot)
+
+    if not 0 <= validation_size <= len(train_images):
+        raise ValueError(
+            'Validation size should be between 0 and {}. Received: {}.'
+                .format(len(train_images), validation_size))
+
+    validation_images = train_images[:validation_size]
+    validation_labels = train_labels[:validation_size]
+    train_images = train_images[validation_size:]
+    train_labels = train_labels[validation_size:]
+
+    options = dict(dtype=dtype, reshape=reshape, seed=seed)
+
+    train = DataSet(train_images, train_labels, **options)
+    validation = DataSet(validation_images, validation_labels, **options)
+    test = DataSet(test_images, test_labels, **options)
+
+    return base.Datasets(train=train, validation=validation, test=test)
+
+
+def load_mnist(train_dir='MNIST-data'):
+    return read_data_sets(train_dir)

Dragon/python/dragon/vm/tensorflow/contrib/slim/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.contrib.framework.ops.variables import *
+
+from dragon.vm.tensorflow.contrib.layers.layers import \
+    convolution, convolution2d, conv2d
+
+from dragon.vm.tensorflow.contrib.layers.layers import \
+    fully_connected
+
+from dragon.vm.tensorflow.contrib.layers.layers import \
+    batch_norm
+
+from dragon.vm.tensorflow.contrib.layers.layers import \
+    avg_pool2d, max_pool2d
+
+from dragon.vm.tensorflow.contrib.layers.layers import \
+    flatten
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/examples/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+

Dragon/python/dragon/vm/tensorflow/examples/tutorials/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+

Dragon/python/dragon/vm/tensorflow/examples/tutorials/mnist/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+

Dragon/python/dragon/vm/tensorflow/examples/tutorials/mnist/input_data.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.contrib.learn.datasets.mnist import read_data_sets
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/framework/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/framework/constant_op.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+__all__ = ['constant']
+
+import numpy as np
+
+import dragon.core.workspace as ws
+from dragon.core.tensor import Tensor
+
+from dragon.vm.tensorflow.framework import dtypes
+
+
+def constant(value, dtype=None, shape=None, name=None, verify_shape=False):
+    # determine the data type
+    if dtype == None: dtype = dtypes.float32
+    if isinstance(value, np.ndarray):
+        feed = value.astype(dtype.as_numpy_dtype)
+    elif isinstance(value, list):
+        feed = np.array(value, dtype.as_numpy_dtype)
+    else:
+        feed = np.array([value], dtype.as_numpy_dtype)
+
+    # determine the shape
+    if shape is not None:
+        if feed.size == 1:
+            # case 1: broadcast with scalar value
+            c = feed[0]
+            feed = np.zeros(shape, dtype.as_numpy_dtype)
+            feed.fill(c)
+        else:
+            # case 2: reshape directly
+            if verify_shape:
+                if shape is not None:
+                    if len(shape) != len(value.shape):
+                        raise RuntimeError('The constant was limited to {} dimensions, \
+                                                while feed a value with {} dimensions.'.
+                                           format(len(shape), len(value.shape)))
+                    for i in xrange(len(shape)):
+                        if shape[i] is None: continue
+                        if shape[i] != value.shape[i]:
+                            raise RuntimeError('The shape of constant was limited as (' +
+                                               ','.join([str(dim) for dim in shape]) + '), ' +
+                                               'while feed a value with (' + ','.join([str(dim) for dim in value.shape]) + ').')
+            feed = feed.reshape(shape)
+
+    # feed to VM
+    tensor = Tensor(name)
+    tensor.shape = list(feed.shape)
+    ws.FeedTensor(tensor, feed)
+    return tensor

Dragon/python/dragon/vm/tensorflow/framework/dtypes.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import numpy as np
+
+from dragon.vm.tensorflow.protobuf import types_pb2
+
+class DType(object):
+    """
+    The basic data type.
+    """
+    def __init__(self, type_enum):
+        type_enum = int(type_enum)
+        if (type_enum not in types_pb2.DataType.values()
+            or type_enum == types_pb2.DT_INVALID):
+            raise TypeError(
+                "type_enum is not a valid types_pb2.DataType: %s" % type_enum)
+        self._type_enum = type_enum
+
+    @property
+    def base_dtype(self):
+        return self
+
+    @property
+    def real_dtype(self):
+        base = self.base_dtype
+        if base == complex64:
+            return float32
+        elif base == complex128:
+            return float64
+        else:
+            return self
+
+    @property
+    def is_numpy_compatible(self):
+        return (self._type_enum != types_pb2.DT_RESOURCE and
+                self._type_enum != types_pb2.DT_RESOURCE_REF)
+
+    @property
+    def as_numpy_dtype(self):
+        return _TF_TO_NP[self._type_enum]
+
+    @property
+    def as_datatype_enum(self):
+        return self._type_enum
+
+    @property
+    def is_bool(self):
+        return self.base_dtype == bool
+
+    @property
+    def is_integer(self):
+        return (self.is_numpy_compatible and not self.is_quantized and
+                issubclass(self.as_numpy_dtype, np.integer))
+
+    @property
+    def is_floating(self):
+        return self.is_numpy_compatible and issubclass(self.as_numpy_dtype,
+                                                       np.floating)
+
+    @property
+    def is_complex(self):
+        return self.base_dtype in (complex64, complex128)
+
+    @property
+    def is_quantized(self):
+        return self.base_dtype in [qint8, quint8, qint16, quint16, qint32, bfloat16]
+
+    @property
+    def is_unsigned(self):
+        try:
+            return self.min == 0
+        except TypeError:
+            return False
+
+    @property
+    def min(self):
+        if (self.is_quantized or self.base_dtype in
+            (bool, string, complex64, complex128)):
+            raise TypeError("Cannot find minimum value of %s." % self)
+        try:
+            return np.finfo(self.as_numpy_dtype()).min
+        except:
+            try:
+                return np.iinfo(self.as_numpy_dtype()).min
+            except:
+                raise TypeError("Cannot find minimum value of %s." % self)
+
+    @property
+    def max(self):
+        if (self.is_quantized or self.base_dtype in
+            (bool, string, complex64, complex128)):
+            raise TypeError("Cannot find maximum value of %s." % self)
+        try:
+            return np.finfo(self.as_numpy_dtype()).max
+        except:
+            try:
+                return np.iinfo(self.as_numpy_dtype()).max
+            except:
+                raise TypeError("Cannot find maximum value of %s." % self)
+
+    @property
+    def limits(self, clip_negative=True):
+        min, max = dtype_range[self.as_numpy_dtype]
+        if clip_negative:
+            min = 0
+        return min, max
+
+    def is_compatible_with(self, other):
+        other = as_dtype(other)
+        return self._type_enum in (
+            other.as_datatype_enum, other.base_dtype.as_datatype_enum)
+
+    def __eq__(self, other):
+        if other is None:
+            return False
+        try:
+            dtype = as_dtype(other).as_datatype_enum
+            return self._type_enum == dtype
+        except TypeError:
+            return False
+
+    def __ne__(self, other):
+        return not self.__eq__(other)
+
+    @property
+    def name(self):
+        """Returns the string name for this `DType`."""
+        return _TYPE_TO_STRING[self._type_enum]
+
+    def __int__(self):
+        return self._type_enum
+
+    def __str__(self):
+        return "<dtype: %r>" % self.name
+
+    def __hash__(self):
+        return self._type_enum
+
+
+# Define data type range of numpy dtype
+dtype_range = {np.bool_: (False, True),
+               np.bool8: (False, True),
+               np.uint8: (0, 255),
+               np.uint16: (0, 65535),
+               np.int8: (-128, 127),
+               np.int16: (-32768, 32767),
+               np.int64: (-2 ** 63, 2 ** 63 - 1),
+               np.uint64: (0, 2 ** 64 - 1),
+               np.int32: (-2 ** 31, 2 ** 31 - 1),
+               np.uint32: (0, 2 ** 32 - 1),
+               np.float32: (-1, 1),
+               np.float64: (-1, 1)}
+
+
+# Define standard wrappers for the types_pb2.DataType enum.
+resource = DType(types_pb2.DT_RESOURCE)
+float16 = DType(types_pb2.DT_HALF)
+half = float16
+float32 = DType(types_pb2.DT_FLOAT)
+float64 = DType(types_pb2.DT_DOUBLE)
+double = float64
+int32 = DType(types_pb2.DT_INT32)
+uint8 = DType(types_pb2.DT_UINT8)
+uint16 = DType(types_pb2.DT_UINT16)
+int16 = DType(types_pb2.DT_INT16)
+int8 = DType(types_pb2.DT_INT8)
+string = DType(types_pb2.DT_STRING)
+complex64 = DType(types_pb2.DT_COMPLEX64)
+complex128 = DType(types_pb2.DT_COMPLEX128)
+int64 = DType(types_pb2.DT_INT64)
+bool = DType(types_pb2.DT_BOOL)
+qint8 = DType(types_pb2.DT_QINT8)
+quint8 = DType(types_pb2.DT_QUINT8)
+qint16 = DType(types_pb2.DT_QINT16)
+quint16 = DType(types_pb2.DT_QUINT16)
+qint32 = DType(types_pb2.DT_QINT32)
+bfloat16 = DType(types_pb2.DT_BFLOAT16)
+
+
+# Standard mappings between types_pb2.DataType values and string names.
+_TYPE_TO_STRING = {
+    types_pb2.DT_HALF: "float16",
+    types_pb2.DT_FLOAT: "float32",
+    types_pb2.DT_DOUBLE: "float64",
+    types_pb2.DT_INT32: "int32",
+    types_pb2.DT_UINT8: "uint8",
+    types_pb2.DT_UINT16: "uint16",
+    types_pb2.DT_INT16: "int16",
+    types_pb2.DT_INT8: "int8",
+    types_pb2.DT_STRING: "string",
+    types_pb2.DT_COMPLEX64: "complex64",
+    types_pb2.DT_COMPLEX128: "complex128",
+    types_pb2.DT_INT64: "int64",
+    types_pb2.DT_BOOL: "bool",
+    types_pb2.DT_QINT8: "qint8",
+    types_pb2.DT_QUINT8: "quint8",
+    types_pb2.DT_QINT16: "qint16",
+    types_pb2.DT_QUINT16: "quint16",
+    types_pb2.DT_QINT32: "qint32",
+    types_pb2.DT_BFLOAT16: "bfloat16",
+    types_pb2.DT_RESOURCE: "resource",
+    types_pb2.DT_HALF_REF: "float16_ref",
+    types_pb2.DT_FLOAT_REF: "float32_ref",
+    types_pb2.DT_DOUBLE_REF: "float64_ref",
+    types_pb2.DT_INT32_REF: "int32_ref",
+    types_pb2.DT_UINT8_REF: "uint8_ref",
+    types_pb2.DT_UINT16_REF: "uint16_ref",
+    types_pb2.DT_INT16_REF: "int16_ref",
+    types_pb2.DT_INT8_REF: "int8_ref",
+    types_pb2.DT_STRING_REF: "string_ref",
+    types_pb2.DT_COMPLEX64_REF: "complex64_ref",
+    types_pb2.DT_COMPLEX128_REF: "complex128_ref",
+    types_pb2.DT_INT64_REF: "int64_ref",
+    types_pb2.DT_BOOL_REF: "bool_ref",
+    types_pb2.DT_QINT8_REF: "qint8_ref",
+    types_pb2.DT_QUINT8_REF: "quint8_ref",
+    types_pb2.DT_QINT16_REF: "qint16_ref",
+    types_pb2.DT_QUINT16_REF: "quint16_ref",
+    types_pb2.DT_QINT32_REF: "qint32_ref",
+    types_pb2.DT_BFLOAT16_REF: "bfloat16_ref",
+    types_pb2.DT_RESOURCE_REF: "resource_ref",
+}
+
+# Standard mappings between types_pb2.DataType values and numpy.dtypes.
+
+# Numpy representation for quantized dtypes.
+_np_qint8 = np.dtype([("qint8", np.int8, 1)])
+_np_quint8 = np.dtype([("quint8", np.uint8, 1)])
+_np_qint16 = np.dtype([("qint16", np.int16, 1)])
+_np_quint16 = np.dtype([("quint16", np.uint16, 1)])
+_np_qint32 = np.dtype([("qint32", np.int32, 1)])
+
+_NP_TO_TF = frozenset([
+    (np.float16, float16),
+    (np.float32, float32),
+    (np.float64, float64),
+    (np.int32, int32),
+    (np.int64, int64),
+    (np.uint8, uint8),
+    (np.uint16, uint16),
+    (np.int16, int16),
+    (np.int8, int8),
+    (np.complex64, complex64),
+    (np.complex128, complex128),
+    (np.object, string),
+    (np.bool, bool),
+    (_np_qint8, qint8),
+    (_np_quint8, quint8),
+    (_np_qint16, qint16),
+    (_np_quint16, quint16),
+    (_np_qint32, qint32),
+    # NOTE(touts): Intentionally no way to feed a DT_BFLOAT16.
+])
+
+_TF_TO_NP = {
+    types_pb2.DT_HALF: np.float16,
+    types_pb2.DT_FLOAT: np.float32,
+    types_pb2.DT_DOUBLE: np.float64,
+    types_pb2.DT_INT32: np.int32,
+    types_pb2.DT_UINT8: np.uint8,
+    types_pb2.DT_UINT16: np.uint16,
+    types_pb2.DT_INT16: np.int16,
+    types_pb2.DT_INT8: np.int8,
+    # NOTE(touts): For strings we use np.object as it supports variable length
+    # strings.
+    types_pb2.DT_STRING: np.object,
+    types_pb2.DT_COMPLEX64: np.complex64,
+    types_pb2.DT_COMPLEX128: np.complex128,
+    types_pb2.DT_INT64: np.int64,
+    types_pb2.DT_BOOL: np.bool,
+    types_pb2.DT_QINT8: _np_qint8,
+    types_pb2.DT_QUINT8: _np_quint8,
+    types_pb2.DT_QINT16: _np_qint16,
+    types_pb2.DT_QUINT16: _np_quint16,
+    types_pb2.DT_QINT32: _np_qint32,
+    types_pb2.DT_BFLOAT16: np.uint16,
+}
+
+
+def as_dtype(type_value):
+    if isinstance(type_value, DType):
+        return type_value
+
+    if isinstance(type_value, np.dtype):
+        if type_value.type == np.string_ or type_value.type == np.unicode_:
+            return string
+
+    for key, val in _NP_TO_TF:
+        try:
+            if key == type_value:
+                return val
+        except TypeError as e:
+            raise TypeError("Cannot convert {} to a dtype. {}".format(type_value, e))
+
+    raise TypeError("Cannot convert value %r to a TensorFlow DType." % type_value)
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/framework/framework_lib.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from .ops import Graph
+
+# Utilities used when building a Graph.
+from dragon.vm.tensorflow.framework.ops import device
+from dragon.vm.tensorflow.framework.ops import name_scope
+from dragon.vm.tensorflow.framework.ops import get_default_graph
+from dragon.vm.tensorflow.framework.ops import add_to_collection
+from dragon.vm.tensorflow.framework.ops import get_collection
+from dragon.vm.tensorflow.framework.ops import convert_to_tensor
+from dragon.vm.tensorflow.framework.ops import GraphKeys
+from dragon.vm.tensorflow.framework.constant_op import *
+
+from dragon.vm.tensorflow.framework.dtypes import *
+
+
+from dragon.vm.tensorflow.framework.tensor_shape import Dimension
+from dragon.vm.tensorflow.framework.tensor_shape import TensorShape
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/framework/ops.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import re
+
+from dragon.core.tensor import Tensor
+from dragon.core.scope import TensorScope, DeviceScope
+
+from dragon.vm.tensorflow.framework import dtypes
+
+_default_graph = None
+
+
+def convert_to_tensor(value, dtype=None, name=None, **kwargs):
+    """Converts the given value to a Tensor.
+
+    Parameters
+    ----------
+    value : basic type, list or numpy.ndarray
+        The value to convert.
+    dtype : Dtype or None
+        The data type. If ``None``, inferred from the type of `value`.
+    name : str or None
+        The Optional name.
+
+    Returns
+    -------
+    Tensor
+        The output tensor.
+
+    """
+    if dtype is not None:
+        if not isinstance(dtype, str):
+            if isinstance(dtype, dtypes.DType):
+                dtype = dtype.name
+            else:
+                raise ValueError('The dtype should be a str of a tf.Dtype.')
+    tensor = Tensor(name=name, dtype=dtype)
+    tensor.set_value(value)
+    return tensor
+
+
+class Graph(object):
+    """
+    A virtual graph.
+    """
+    def __init__(self):
+        self._collections = {}
+
+    def get_collection_ref(self, name):
+        coll_list = self._collections.get(name, None)
+        if coll_list is None:
+            coll_list = []
+            self._collections[name] = coll_list
+        return coll_list
+
+    def get_collection(self, name, scope=None):
+        coll_list = self._collections.get(name, None)
+        if coll_list is None:
+            return []
+        if scope is None:
+            return list(coll_list)
+        else:
+            filter_coll_list = []
+            regex = re.compile(scope)
+            for item in coll_list:
+                if hasattr(item, "name") and regex.match(item.name):
+                    filter_coll_list.append(item)
+            return filter_coll_list
+
+    def add_to_collection(self, name, value):
+        if name not in self._collections:
+            self._collections[name] = [value]
+        else:
+            self._collections[name].append(value)
+
+    def add_to_collections(self, names, value):
+        for name in names:
+            self.add_to_collection(name, value)
+
+    def device(self, device_name_or_function):
+        if not isinstance(device_name_or_function, str):
+            raise TypeError('The device function should be a str.')
+        device_and_id = device_name_or_function.split('/')[1]
+        device, id = device_and_id.split(':')
+        if device not in ['cpu', 'gpu']:
+            raise ValueError('The device should either be cpu or gpu.')
+        try:
+            id = int(id)
+        except Exception as e:
+            raise ValueError('The device id should be a integer.')
+        return DeviceScope(device, id=id, use_cudnn=True)
+
+
+def device(device_name_or_function):
+  return get_default_graph().device(device_name_or_function)
+
+
+def get_default_graph():
+    global _default_graph
+    if _default_graph is None:
+        _default_graph = Graph()
+    return _default_graph
+
+
+class GraphKeys(object):
+  GLOBAL_VARIABLES = "variables"
+  # Key to collect local variables that are local to the machine and are not
+  # saved/restored.
+  LOCAL_VARIABLES = "local_variables"
+  # Key to collect model variables defined by layers.
+  MODEL_VARIABLES = "model_variables"
+  # Key to collect Variable objects that will be trained by the
+  # optimizers.
+  TRAINABLE_VARIABLES = "trainable_variables"
+  # Key to collect summaries.
+  SUMMARIES = "summaries"
+  # Key to collect QueueRunners.
+  QUEUE_RUNNERS = "queue_runners"
+  # Key to collect table initializers.
+  TABLE_INITIALIZERS = "table_initializer"
+  # Key to collect asset filepaths. An asset represents an external resource
+  # like a vocabulary file.
+  ASSET_FILEPATHS = "asset_filepaths"
+  # Key to collect Variable objects that keep moving averages.
+  MOVING_AVERAGE_VARIABLES = "moving_average_variables"
+  # Key to collect regularization losses at graph construction.
+  REGULARIZATION_LOSSES = "regularization_losses"
+  # Key to collect concatenated sharded variables.
+  CONCATENATED_VARIABLES = "concatenated_variables"
+  # Key to collect savers.
+  SAVERS = "savers"
+  # Key to collect weights
+  WEIGHTS = "weights"
+  # Key to collect biases
+  BIASES = "biases"
+  # Key to collect activations
+  ACTIVATIONS = "activations"
+  # Key to collect update_ops
+  UPDATE_OPS = "update_ops"
+  # Key to collect losses
+  LOSSES = "losses"
+  # Key to collect BaseSaverBuilder.SaveableObject instances for checkpointing.
+  SAVEABLE_OBJECTS = "saveable_objects"
+  # Key to collect all shared resources used by the graph which need to be
+  # initialized once per cluster.
+  RESOURCES = "resources"
+  # Key to collect all shared resources used in this graph which need to be
+  # initialized once per session.
+  LOCAL_RESOURCES = "local_resources"
+  # Trainable resource-style variables.
+  TRAINABLE_RESOURCE_VARIABLES = "trainable_resource_variables"
+
+  # Key to indicate various ops.
+  INIT_OP = "init_op"
+  LOCAL_INIT_OP = "local_init_op"
+  READY_OP = "ready_op"
+  READY_FOR_LOCAL_INIT_OP = "ready_for_local_init_op"
+  SUMMARY_OP = "summary_op"
+  GLOBAL_STEP = "global_step"
+
+  # Used to count the number of evaluations performed during a single evaluation
+  # run.
+  EVAL_STEP = "eval_step"
+  TRAIN_OP = "train_op"
+
+  # Key for control flow context.
+  COND_CONTEXT = "cond_context"
+  WHILE_CONTEXT = "while_context"
+
+  # Key for streaming model ports.
+  # NOTE(yuanbyu): internal and experimental.
+  _STREAMING_MODEL_PORTS = "streaming_model_ports"
+
+
+def get_collection_ref(key):
+  return get_default_graph().get_collection_ref(key)
+
+
+def get_collection(key, scope=None):
+    return get_default_graph().get_collection(key, scope)
+
+
+def add_to_collection(name, value):
+    get_default_graph().add_to_collection(name, value)
+
+
+def add_to_collections(names, value):
+    get_default_graph().add_to_collections(names, value)
+
+
+def name_scope(name, default_name=None, values=None):
+    n = default_name if name is None else name
+    n = '' if n is None else n
+    return TensorScope(prefix=n)

Dragon/python/dragon/vm/tensorflow/framework/random_seed.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import dragon.config as config
+
+
+DEFAULT_GRAPH_SEED = 87654321
+_MAXINT32 = 2**31 - 1
+
+
+def _truncate_seed(seed):
+  return seed % _MAXINT32  # Truncate to fit into 32-bit integer
+
+
+def get_seed(op_seed):
+    """Return the global random seed.
+
+    Parameters
+    ----------
+    op_seed : int
+        The optional seed to use.
+
+    Return
+    ------
+    tuple
+        A tuple of two ints for using.
+
+    """
+    graph_seed = config.GetRandomSeed()
+    if graph_seed is not None:
+        if op_seed is None:
+            # pylint: disable=protected-access
+            op_seed = graph_seed
+        seeds = _truncate_seed(graph_seed), _truncate_seed(op_seed)
+    else:
+        if op_seed is not None:
+            seeds = DEFAULT_GRAPH_SEED, _truncate_seed(op_seed)
+        else:
+            seeds = None, None
+    return seeds
+
+
+def set_random_seed(seed):
+    """Set the global random seed.
+
+    Parameters
+    ----------
+    seed : int
+        The seed to use.
+
+    """
+    config.SetRandomSeed(seed)
+

Dragon/python/dragon/vm/tensorflow/framework/tensor_shape.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.core.tensor import Tensor
+
+class Dimension(object):
+    def __init__(self, value):
+        if value is None:
+            self._value = None
+        else:
+            self._value = int(value)
+            if self._value < 0:
+                raise ValueError("Dimension %d must be >= 0" % self._value)
+
+    @property
+    def value(self):
+        return self._value
+
+    def __repr__(self):
+        return "Dimension(%s)" % repr(self._value)
+
+    def __str__(self):
+        value = self._value
+        return "?" if value is None else str(value)
+
+    def __eq__(self, other):
+        try:
+            other = as_dimension(other)
+        except (TypeError, ValueError):
+            return NotImplemented
+        if self._value is None or other.value is None:
+            return None
+        return self._value == other.value
+
+
+def as_dimension(value):
+    if isinstance(value, Dimension): return value
+    else: return Dimension(value)
+
+
+class TensorShape(object):
+    def __init__(self, dims):
+        if dims is None:
+            self._dims = None
+        elif isinstance(dims, TensorShape):
+            self._dims = dims.dims
+        else:
+            try:
+                dims_iter = iter(dims)
+            except TypeError:
+                self._dims = [as_dimension(dims)]
+            else:
+                self._dims = [as_dimension(d) for d in dims_iter]
+
+    @property
+    def dims(self):
+        return self._dims
+
+    @property
+    def ndims(self):
+        if self._dims is None:
+            return None
+        else:
+            return len(self._dims)
+
+    def as_list(self):
+        if self._dims is None:
+            raise ValueError("as_list() is not defined on an unknown TensorShape.")
+        return [dim.value for dim in self._dims]
+
+    def __repr__(self):
+        return "TensorShape(%r)" % self._dims
+
+    def __str__(self):
+        if self.ndims is None:
+            return "<unknown>"
+        elif self.ndims == 1:
+            return "(%s,)" % self._dims[0]
+        else:
+            return "(%s)" % ", ".join(str(d) for d in self._dims)
+
+    def __getitem__(self, key):
+        if self._dims is not None:
+            if isinstance(key, slice):
+                return TensorShape(self._dims[key])
+            else:
+                return self._dims[key]
+        else:
+            return Dimension(None)
+
+
+def get_shape(self):
+    """
+    Construct the shape descriptor.
+
+    Returns
+    -------
+    TensorShape
+        The shape description.
+
+    """
+    return TensorShape(self.shape)
+
+Tensor.get_shape = get_shape
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/layers/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+

Dragon/python/dragon/vm/tensorflow/layers/base.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import re
+import collections
+import weakref
+
+from dragon.core.scope import get_tensor_scope
+
+from dragon.vm.tensorflow.framework import ops
+from dragon.vm.tensorflow.framework import dtypes
+from dragon.vm.tensorflow.ops import var_scope as vs
+from dragon.vm.tensorflow.ops import variables as tf_variables
+from dragon.vm.tensorflow.util import nest
+
+
+class Layer(object):
+    def __init__(self, trainable=True, name=None, dtype=dtypes.float32, **kwargs):
+        allowed_kwargs = {'_scope',
+                          '_reuse'}
+        for kwarg in kwargs:
+            if kwarg not in allowed_kwargs:
+                raise TypeError('Keyword argument not understood:', kwarg)
+
+        self.trainable = trainable
+        self.built = False
+        self._trainable_weights = []
+        self._non_trainable_weights = []
+        self._updates = []
+        self._losses = []
+        self._reuse = kwargs.get('_reuse')
+        self._graph = ops.get_default_graph()
+        self._per_input_losses = {}
+        self._per_input_updates = {}
+        self.dtype = dtypes.as_dtype(dtype)
+        self.input_spec = None
+
+        # Determine layer name
+        if name is None:
+            base_name = _to_snake_case(self.__class__.__name__)
+            self.name = _unique_layer_name(base_name)
+        else:
+            base_name = name
+            self.name = name
+
+        self._base_name = base_name
+
+        # Determine variable scope.
+        self._scope = None
+
+    def build(self, _):
+        self.built = True
+
+    def call(self, inputs, *args, **kwargs):
+        raise NotImplementedError
+
+    @property
+    def updates(self):
+        return self._updates
+
+    def __call__(self, inputs, *args, **kwargs):
+        with vs.variable_scope(self._scope,
+                               reuse=self.built or self._reuse) as scope:
+            with ops.name_scope(scope.original_name_scope):
+                if not self.built:
+                    input_shapes = [x.get_shape()
+                                    for x in nest.flatten(inputs)]
+                    if len(input_shapes) == 1:
+                        self.build(input_shapes[0])
+                    else:
+                        self.build(input_shapes)
+                outputs = self.call(inputs, *args, **kwargs)
+                # # Apply activity regularization.
+                # # Note that it should be applied every time the layer creates a new
+                # # output, since it is output-specific.
+                # if hasattr(self, 'activity_regularizer') and self.activity_regularizer:
+                #     output_list = _to_list(outputs)
+                #     for output in output_list:
+                #         with ops.name_scope('ActivityRegularizer'):
+                #             activity_regularization = self.activity_regularizer(output)
+                #         self.add_loss(activity_regularization)
+                #         _add_elements_to_collection(
+                #             activity_regularization, ops.GraphKeys.REGULARIZATION_LOSSES)
+        # Update global default collections.
+        _add_elements_to_collection(self.updates, ops.GraphKeys.UPDATE_OPS)
+        self.built = True
+        return outputs
+
+    def add_variable(self, name, shape, dtype=None, trainable=True,
+                     initializer=None, regularizer=None):
+        if dtype is None:
+            dtype = self.dtype
+        existing_variables = set(tf_variables.global_variables())
+
+        with vs.variable_scope(self._scope,
+                               reuse=self.built or self._reuse) as scope:
+            with ops.name_scope(scope.original_name_scope):
+
+                full_name = get_tensor_scope() + name
+                variable = vs.get_variable(name,
+                                           shape=shape,
+                                           initializer=initializer,
+                                           dtype=dtypes.as_dtype(dtype),
+                                           trainable=trainable and self.trainable)
+                if variable in existing_variables:
+                    # Work only if the layer is built
+                    return variable
+                if regularizer:
+                    raise NotImplementedError()
+
+        if trainable:
+            self._trainable_weights.append(variable)
+        else:
+            self._non_trainable_weights.append(variable)
+        return variable
+
+    def apply(self, inputs, *args, **kwargs):
+        return self.__call__(inputs, *args, **kwargs)
+
+
+class InputSpec(object):
+    def __init__(self, dtype=None, shape=None,
+                 ndim=None, max_ndim=None, min_ndim=None, axes=None):
+        self.dtype = dtype
+        self.shape = shape
+        if shape is not None:
+            self.ndim = len(shape)
+        else:
+            self.ndim = ndim
+        self.max_ndim = max_ndim
+        self.min_ndim = min_ndim
+        self.axes = axes or {}
+
+
+def _to_snake_case(name):
+    intermediate = re.sub('(.)([A-Z][a-z0-9]+)', r'\1_\2', name)
+    insecure = re.sub('([a-z])([A-Z])', r'\1_\2', intermediate).lower()
+    if insecure[0] != '_': return insecure
+    return 'private' + insecure
+
+
+PER_GRAPH_LAYER_NAME_UIDS = weakref.WeakKeyDictionary()
+
+
+def _unique_layer_name(name):
+    graph = ops.get_default_graph()
+    if graph not in PER_GRAPH_LAYER_NAME_UIDS:
+        PER_GRAPH_LAYER_NAME_UIDS[graph] = collections.defaultdict(int)
+    layer_name_uids = PER_GRAPH_LAYER_NAME_UIDS[graph]
+    layer_name_uids[name] += 1
+
+
+def _to_list(x):
+    if isinstance(x, (list, tuple)):
+        return list(x)
+    return [x]
+
+
+def _add_elements_to_collection(elements, collection_list):
+    elements = _to_list(elements)
+    collection_list = _to_list(collection_list)
+    for name in collection_list:
+        collection = ops.get_collection_ref(name)
+        collection_set = set(collection)
+        for element in elements:
+            if element not in collection_set:
+                collection.append(element)

Dragon/python/dragon/vm/tensorflow/layers/convolutional.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.framework import tensor_shape
+from dragon.vm.tensorflow.layers import base, utils
+from dragon.vm.tensorflow.ops import init_ops
+from dragon.vm.tensorflow.ops import nn
+
+
+class _Conv(base.Layer):
+    def __init__(self, rank,
+                 filters,
+                 kernel_size,
+                 strides=1,
+                 padding='valid',
+                 data_format='channels_last',
+                 dilation_rate=1,
+                 activation=None,
+                 use_bias=True,
+                 kernel_initializer=None,
+                 bias_initializer=init_ops.zeros_initializer(),
+                 kernel_regularizer=None,
+                 bias_regularizer=None,
+                 activity_regularizer=None,
+                 trainable=True,
+                 name=None,
+                 **kwargs):
+        super(_Conv, self).__init__(trainable=trainable, name=name, **kwargs)
+        self.rank = rank
+        self.filters = filters
+        self.kernel_size = utils.normalize_tuple(kernel_size, rank, 'kernel_size')
+        self.strides = utils.normalize_tuple(strides, rank, 'strides')
+        self.padding = utils.normalize_padding(padding)
+        self.data_format = utils.normalize_data_format(data_format)
+        self.dilation_rate = utils.normalize_tuple(dilation_rate, rank, 'dilation_rate')
+        self.activation = activation
+        self.use_bias = use_bias
+        self.kernel_initializer = kernel_initializer
+        self.bias_initializer = bias_initializer
+        self.kernel_regularizer = kernel_regularizer
+        self.bias_regularizer = bias_regularizer
+        self.activity_regularizer = activity_regularizer
+        self.input_spec = base.InputSpec(ndim=self.rank + 2)
+
+    def build(self, input_shape):
+        input_shape = tensor_shape.TensorShape(input_shape)
+        if self.data_format == 'channels_first':
+            channel_axis = 1
+        else:
+            channel_axis = -1
+        if input_shape[channel_axis].value is None:
+            raise ValueError('The channel dimension of the inputs '
+                             'should be defined. Found `None`.')
+        input_dim = input_shape[channel_axis].value
+
+        if self.data_format == 'channels_first':
+            # For channels first: (n_out, n_in, k_h, k_w)
+            kernel_shape = (self.filters, input_dim) + self.kernel_size
+        else:
+            # For channels last: (k_h, k_w, n_in, n_out)
+            kernel_shape = self.kernel_size + (input_dim, self.filters)
+
+        self.kernel = self.add_variable(name='kernel',
+                                        shape=kernel_shape,
+                                        initializer=self.kernel_initializer,
+                                        regularizer=self.kernel_regularizer,
+                                        trainable=True,
+                                        dtype=self.dtype)
+        if self.use_bias:
+            self.bias = self.add_variable(name='bias',
+                                          shape=(self.filters,),
+                                          initializer=self.bias_initializer,
+                                          regularizer=self.bias_regularizer,
+                                          trainable=True,
+                                          dtype=self.dtype)
+        else:
+            self.bias = None
+        self.input_spec = base.InputSpec(ndim=self.rank + 2,
+                                         axes={channel_axis: input_dim})
+        self.built = True
+
+    def call(self, inputs):
+        tf_data_format = \
+            utils.convert_data_format(self.data_format, self.rank + 2)
+        outputs = nn.convolution(
+            input=inputs,
+            filter=self.kernel,
+            dilation_rate=self.dilation_rate,
+            strides=self.strides,
+            padding=self.padding.upper(),
+            data_format=tf_data_format)
+
+        if self.bias is not None:
+            outputs = nn.bias_add(outputs, self.bias, data_format=tf_data_format)
+
+        if self.activation is not None:
+            return self.activation(outputs)
+        return outputs
+
+
+class Conv2D(_Conv):
+    def __init__(self, filters,
+                 kernel_size,
+                 strides=(1, 1),
+                 padding='valid',
+                 data_format='channels_last',
+                 dilation_rate=(1, 1),
+                 activation=None,
+                 use_bias=True,
+                 kernel_initializer=None,
+                 bias_initializer=init_ops.zeros_initializer(),
+                 kernel_regularizer=None,
+                 bias_regularizer=None,
+                 activity_regularizer=None,
+                 trainable=True,
+                 name=None,
+                 **kwargs):
+        super(Conv2D, self).__init__(
+            rank=2,
+            filters=filters,
+            kernel_size=kernel_size,
+            strides=strides,
+            padding=padding,
+            data_format=data_format,
+            dilation_rate=dilation_rate,
+            activation=activation,
+            use_bias=use_bias,
+            kernel_initializer=kernel_initializer,
+            bias_initializer=bias_initializer,
+            kernel_regularizer=kernel_regularizer,
+            bias_regularizer=bias_regularizer,
+            activity_regularizer=activity_regularizer,
+            trainable=trainable,
+            name=name, **kwargs)
+
+
+def conv2d(inputs,
+           filters,
+           kernel_size,
+           strides=(1, 1),
+           padding='valid',
+           data_format='channels_last',
+           dilation_rate=(1, 1),
+           activation=None,
+           use_bias=True,
+           kernel_initializer=None,
+           bias_initializer=init_ops.zeros_initializer(),
+           kernel_regularizer=None,
+           bias_regularizer=None,
+           activity_regularizer=None,
+           trainable=True,
+           name=None,
+           reuse=None):
+    layer = Conv2D(
+        filters=filters,
+        kernel_size=kernel_size,
+        strides=strides,
+        padding=padding,
+        data_format=data_format,
+        dilation_rate=dilation_rate,
+        activation=activation,
+        use_bias=use_bias,
+        kernel_initializer=kernel_initializer,
+        bias_initializer=bias_initializer,
+        kernel_regularizer=kernel_regularizer,
+        bias_regularizer=bias_regularizer,
+        activity_regularizer=activity_regularizer,
+        trainable=trainable,
+        name=name,
+        _reuse=reuse,
+        _scope=name)
+    return layer.apply(inputs)

Dragon/python/dragon/vm/tensorflow/layers/core.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.framework import tensor_shape
+from dragon.vm.tensorflow.layers import base, utils
+from dragon.vm.tensorflow.ops import init_ops
+from dragon.vm.tensorflow.ops import nn, standard_ops
+
+
+class Dense(base.Layer):
+    def __init__(self, units,
+                 activation=None,
+                 use_bias=True,
+                 kernel_initializer=None,
+                 bias_initializer=init_ops.zeros_initializer(),
+                 kernel_regularizer=None,
+                 bias_regularizer=None,
+                 activity_regularizer=None,
+                 trainable=True,
+                 name=None,
+                 **kwargs):
+        super(Dense, self).__init__(trainable=trainable, name=name, **kwargs)
+        self.units = units
+        self.activation = activation
+        self.use_bias = use_bias
+        self.kernel_initializer = kernel_initializer
+        self.bias_initializer = bias_initializer
+        self.kernel_regularizer = kernel_regularizer
+        self.bias_regularizer = bias_regularizer
+        self.activity_regularizer = activity_regularizer
+        self.input_spec = base.InputSpec(min_ndim=2)
+
+    def build(self, input_shape):
+        input_shape = tensor_shape.TensorShape(input_shape)
+        if input_shape[-1].value is None:
+            raise ValueError('The last dimension of the inputs to `Dense` '
+                             'should be defined. Found `None`.')
+        self.input_spec = base.InputSpec(min_ndim=2,
+                                         axes={-1: input_shape[-1].value})
+        self.kernel = self.add_variable('kernel',
+                                        shape=[input_shape[-1].value, self.units],
+                                        initializer=self.kernel_initializer,
+                                        regularizer=self.kernel_regularizer,
+                                        dtype=self.dtype,
+                                        trainable=True)
+        if self.use_bias:
+            self.bias = self.add_variable('bias',
+                                          shape=[self.units, ],
+                                          initializer=self.bias_initializer,
+                                          regularizer=self.bias_regularizer,
+                                          dtype=self.dtype,
+                                          trainable=True)
+        else:
+            self.bias = None
+        self.built = True
+
+    def call(self, inputs):
+        shape = inputs.get_shape().as_list()
+        output_shape = shape[:-1] + [self.units]
+        if len(output_shape) > 2:
+            raise NotImplementedError()
+        else:
+            outputs = standard_ops.matmul(inputs, self.kernel)
+        if self.use_bias:
+            outputs = outputs + self.bias
+        if self.activation is not None:
+            return self.activation(outputs)
+        return outputs
+
+
+def dense(inputs, units,
+          activation=None,
+          use_bias=True,
+          kernel_initializer=None,
+          bias_initializer=init_ops.zeros_initializer(),
+          kernel_regularizer=None,
+          bias_regularizer=None,
+          activity_regularizer=None,
+          trainable=True,
+          name=None,
+          reuse=None):
+    layer = Dense(units,
+                  activation=activation,
+                  use_bias=use_bias,
+                  kernel_initializer=kernel_initializer,
+                  bias_initializer=bias_initializer,
+                  kernel_regularizer=kernel_regularizer,
+                  bias_regularizer=bias_regularizer,
+                  activity_regularizer=activity_regularizer,
+                  trainable=trainable,
+                  name=name,
+                  _scope=name,
+                  _reuse=reuse)
+    return layer.apply(inputs)

Dragon/python/dragon/vm/tensorflow/layers/layers.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.layers.convolutional import conv2d
+
+from dragon.vm.tensorflow.layers.core import dense
+
+from dragon.vm.tensorflow.layers.normalization import \
+    batch_normalization, batch_norm, BatchNorm
+
+from dragon.vm.tensorflow.layers.pooling import \
+    average_pooling2d, max_pooling2d
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/layers/normalization.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import dragon.ops as ops
+
+from dragon.vm.tensorflow.framework import tensor_shape
+from dragon.vm.tensorflow.layers import base
+from dragon.vm.tensorflow.ops import init_ops
+
+
+class BatchNormalization(base.Layer):
+    def __init__(self,
+                 axis=-1,
+                 momentum=0.99,
+                 epsilon=1e-3,
+                 center=True,
+                 scale=True,
+                 beta_initializer=init_ops.zeros_initializer(),
+                 gamma_initializer=init_ops.ones_initializer(),
+                 moving_mean_initializer=init_ops.zeros_initializer(),
+                 moving_variance_initializer=init_ops.ones_initializer(),
+                 beta_regularizer=None,
+                 gamma_regularizer=None,
+                 renorm=False,
+                 renorm_clipping=None,
+                 renorm_momentum=0.99,
+                 fused=False,
+                 trainable=True,
+                 name=None,
+                 **kwargs):
+        super(BatchNormalization, self).__init__(trainable=trainable, name=name, **kwargs)
+        self.axis = axis
+        self.momentum = momentum
+        self.epsilon = epsilon
+        self.center = center
+        self.scale = scale
+        self.beta_initializer = beta_initializer
+        self.gamma_initializer = gamma_initializer
+        self.moving_mean_initializer = moving_mean_initializer
+        self.moving_variance_initializer = moving_variance_initializer
+        self.beta_regularizer = beta_regularizer
+        self.gamma_regularizer = gamma_regularizer
+        self.renorm = renorm
+        self.fused = fused
+        self.trainable = trainable
+        if fused:
+            if not center or not scale:
+                raise ValueError('fused norm requires both center and scale set to be True.')
+        if renorm:
+            raise ValueError('renorm is currently not supported.')
+
+    def build(self, input_shape):
+        input_shape = tensor_shape.TensorShape(input_shape)
+        if not input_shape.ndims:
+            raise ValueError('Input has undefined rank:', input_shape)
+        ndim = input_shape.ndims
+        if self.fused and ndim != 4:
+            raise ValueError(
+                'Only 4D inputs are currently supported with fused batch norm. '
+                'Consider reshaping the input to 4D and reshape the output back '
+                'to its original shape. Got input rank: ', ndim)
+        if self.axis < 0:
+            axis = ndim + self.axis
+        else:
+            axis = self.axis
+        if axis < 0 or axis >= ndim:
+            raise ValueError('Value of `axis` argument ' + str(self.axis) +
+                             ' is out of range for input with rank ' + str(ndim))
+        if axis + 1 == ndim:
+            self._data_format = 'NHWC'
+        elif axis == 1:
+            self._data_format = 'NCHW'
+        else:
+            raise ValueError(
+                'Only axis 1 or last axis are currently supported dimensions for '
+                    'batch norm. Got `axis` dimension: ', axis)
+
+        param_dim = input_shape[axis]
+        if not param_dim.value:
+            raise ValueError('Input has undefined `axis` dimension. Input shape: ', input_shape)
+        self.input_spec = base.InputSpec(ndim=ndim, axes={self.axis: param_dim.value})
+
+        if self.center:
+            self.beta = self.add_variable(name='beta',
+                                          shape=(param_dim.value,),
+                                          initializer=self.beta_initializer,
+                                          regularizer=self.beta_regularizer,
+                                          trainable=self.trainable)
+        else:
+            self.beta = None
+        if self.scale:
+            self.gamma = self.add_variable(name='gamma',
+                                           shape=(param_dim.value,),
+                                           initializer=self.gamma_initializer,
+                                           regularizer=self.gamma_regularizer,
+                                           trainable=True)
+        else:
+            self.gamma = None
+
+        self.moving_mean = self.add_variable(name='moving_mean',
+                                             shape=(param_dim.value,),
+                                             initializer=self.moving_mean_initializer,
+                                             trainable=False)
+        self.moving_variance = self.add_variable(name='moving_variance',
+                                                 shape=(param_dim.value,),
+                                                 initializer=self.moving_variance_initializer,
+                                                 trainable=False)
+        if self.renorm: pass
+        self.built = True
+
+    def call(self, inputs, training=False):
+        use_stats = 0 if training else 1
+        if self.fused:
+            return ops.FusedBatchNorm([inputs, self.moving_mean,
+                                               self.moving_variance,
+                                               self.gamma,
+                                               self.beta],
+                                       axis=self.axis,
+                                       momentum=self.momentum,
+                                       eps=self.epsilon,
+                                       use_stats=use_stats,
+                                       mode='DEFAULT')
+
+        x_norm = ops.BatchNorm([inputs, self.moving_mean,
+                                        self.moving_variance],
+                               axis=self.axis,
+                               momentum=self.momentum,
+                               eps=self.epsilon,
+                               use_stats=use_stats,
+                               mode='DEFAULT')
+        if self.gamma is not None:
+            # use scale
+            if self.beta is not None:
+                return ops.Scale([x_norm, self.gamma, self.beta], axis=self.axis, num_axes=1)
+            else:
+                return ops.Scale([x_norm, self.gamma], axis=self.axis, num_axes=1)
+        else:
+            # do not use scale
+            if self.beta is not None:
+                return ops.BiasAdd([x_norm, self.beta], data_format=self._data_format)
+            else:
+                return x_norm
+
+
+def batch_normalization(inputs,
+                        axis=-1,
+                        momentum=0.99,
+                        epsilon=1e-3,
+                        center=True,
+                        scale=True,
+                        beta_initializer=init_ops.zeros_initializer(),
+                        gamma_initializer=init_ops.ones_initializer(),
+                        moving_mean_initializer=init_ops.zeros_initializer(),
+                        moving_variance_initializer=init_ops.ones_initializer(),
+                        beta_regularizer=None,
+                        gamma_regularizer=None,
+                        training=False,
+                        trainable=True,
+                        name=None,
+                        reuse=None,
+                        renorm=False,
+                        renorm_clipping=None,
+                        renorm_momentum=0.99,
+                        fused=False):
+  layer = BatchNormalization(
+      axis=axis,
+      momentum=momentum,
+      epsilon=epsilon,
+      center=center,
+      scale=scale,
+      beta_initializer=beta_initializer,
+      gamma_initializer=gamma_initializer,
+      moving_mean_initializer=moving_mean_initializer,
+      moving_variance_initializer=moving_variance_initializer,
+      beta_regularizer=beta_regularizer,
+      gamma_regularizer=gamma_regularizer,
+      renorm=renorm,
+      renorm_clipping=renorm_clipping,
+      renorm_momentum=renorm_momentum,
+      fused=fused,
+      trainable=trainable,
+      name=name,
+      _reuse=reuse,
+      _scope=name)
+  return layer.apply(inputs, training=training)
+
+
+# Aliases
+BatchNorm = BatchNormalization
+batch_norm = batch_normalization
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/layers/pooling.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.ops import nn
+
+from dragon.vm.tensorflow.layers import base, utils
+
+
+class _Pooling2D(base.Layer):
+    def __init__(self, pool_function, pool_size, strides,
+                 padding='valid', data_format='channels_last',
+                 name=None, **kwargs):
+        super(_Pooling2D, self).__init__(name=name, **kwargs)
+        self.pool_function = pool_function
+        self.pool_size = utils.normalize_tuple(pool_size, 2, 'pool_size')
+        self.strides = utils.normalize_tuple(strides, 2, 'strides')
+        self.padding = utils.normalize_padding(padding)
+        self.data_format = utils.normalize_data_format(data_format)
+        self.input_spec = base.InputSpec(ndim=4)
+
+    def call(self, inputs):
+        if self.data_format == 'channels_last':
+            pool_shape = (1,) + self.pool_size + (1,)
+            strides = (1,) + self.strides + (1,)
+        else:
+            pool_shape = (1, 1) + self.pool_size
+            strides = (1, 1) + self.strides
+        outputs = self.pool_function(
+            inputs,
+            ksize=pool_shape,
+            strides=strides,
+            padding=self.padding.upper(),
+            data_format=utils.convert_data_format(self.data_format, 4))
+        return outputs
+
+
+class AveragePooling2D(_Pooling2D):
+    def __init__(self, pool_size, strides,
+                 padding='valid', data_format='channels_last',
+                 name=None, **kwargs):
+        super(AveragePooling2D, self).__init__(
+            nn.avg_pool,
+            pool_size=pool_size, strides=strides, padding=padding,
+            data_format=data_format, name=name, **kwargs)
+
+
+def average_pooling2d(inputs,
+                      pool_size, strides,
+                      padding='valid', data_format='channels_last',
+                      name=None):
+    layer = AveragePooling2D(pool_size=pool_size, strides=strides,
+                             padding=padding, data_format=data_format,
+                             name=name)
+    return layer.apply(inputs)
+
+
+class MaxPooling2D(_Pooling2D):
+    def __init__(self, pool_size, strides,
+                 padding='valid', data_format='channels_last',
+                 name=None, **kwargs):
+        super(MaxPooling2D, self).__init__(
+            nn.max_pool,
+            pool_size=pool_size, strides=strides, padding=padding,
+            data_format=data_format, name=name, **kwargs)
+
+
+def max_pooling2d(inputs,
+                      pool_size, strides,
+                      padding='valid', data_format='channels_last',
+                      name=None):
+    layer = MaxPooling2D(pool_size=pool_size, strides=strides,
+                         padding=padding, data_format=data_format,
+                         name=name)
+    return layer.apply(inputs)
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/layers/utils.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+def convert_data_format(data_format, ndim):
+    if data_format == 'channels_last':
+        if ndim in (3, 4, 5):
+            return 'NHWC'
+        else:
+            raise ValueError('Input rank not supported:', ndim)
+    elif data_format == 'channels_first':
+        if ndim in (3, 4, 5):
+            return 'NCHW'
+        else:
+            raise ValueError('Input rank not supported:', ndim)
+    else:
+        raise ValueError('Invalid data_format:', data_format)
+
+
+def normalize_tuple(value, n, name):
+    if isinstance(value, int):
+        return (value,) * n
+    else:
+        try:
+            value_tuple = tuple(value)
+        except TypeError:
+            raise ValueError('The `' + name + '` argument must be a tuple of ' +
+                             str(n) + ' integers. Received: ' + str(value))
+        if len(value_tuple) != n:
+            raise ValueError('The `' + name + '` argument must be a tuple of ' +
+                             str(n) + ' integers. Received: ' + str(value))
+        for single_value in value_tuple:
+            try:
+                int(single_value)
+            except ValueError:
+                raise ValueError('The `' + name + '` argument must be a tuple of ' +
+                                 str(n) + ' integers. Received: ' + str(value) + ' '
+                                                                                 'including element ' + str(single_value) + ' of type' +
+                                 ' ' + str(type(single_value)))
+        return value_tuple
+
+
+def normalize_data_format(value):
+    data_format = value.lower()
+    if data_format not in {'channels_first', 'channels_last'}:
+        raise ValueError('The `data_format` argument must be one of '
+                         '"channels_first", "channels_last". Received: ' +
+                         str(value))
+    return data_format
+
+
+def normalize_padding(value):
+    padding = value.lower()
+    if padding not in {'valid', 'same'}:
+        raise ValueError('The `padding` argument must be one of "valid", "same". '
+                         'Received: ' + str(padding))
+    return padding

Dragon/python/dragon/vm/tensorflow/ops/__init__.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/array_ops.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+__all__ = [
+    'expand_dims',
+    'shape',
+    'zeros',
+    'ones',
+    'placeholder',
+    'concat',
+    'transpose',
+    'tile',
+    'reshape'
+]
+
+
+import dragon.ops as ops
+from dragon.core.tensor import Tensor
+
+from dragon.vm.tensorflow.framework import dtypes
+
+
+def expand_dims(input, axis=None, name=None, dim=None):
+
+    if dim is not None:
+        if axis is not None:
+            raise ValueError("cannot specify both 'axis' and 'dim'.")
+        axis = dim
+
+    return ops.ExpandDims(input, axis=axis, name=name)
+
+
+def shape(input, name=None, out_type=dtypes.float32):
+
+    return ops.Shape(input, name=None)
+
+
+def zeros(shape, dtype=dtypes.float32, name=None):
+
+    return ops.Fill(shape, value=0.0, name=name)
+
+
+def ones(shape, dtype=dtypes.float32, name=None):
+
+    return ops.Fill(shape, value=1.0, name=name)
+
+
+def placeholder(dtype, shape=None, name=None):
+    # check data type
+    if dtype is not None:
+        if not isinstance(dtype, dtypes.DType):
+            raise TypeError('The dtype should be a valid tf data type.')
+        dtype = dtype.name
+
+    return Tensor(name=name, shape=shape, dtype=dtype).Placeholder()
+
+
+def concat(values, axis, name=None):
+
+    return ops.Concat(values, axis=axis, name=name)
+
+
+def transpose(a, perm=None, name=None):
+
+    return ops.Transpose(a, perm=perm, name=name)
+
+
+def tile(input, multiples, name=None):
+
+    return ops.Tile(input, multiples=multiples, name=name)
+
+
+def reshape(tensor, shape, name=None):
+
+    return ops.Reshape(tensor, shape=shape, name=None)

Dragon/python/dragon/vm/tensorflow/ops/control_flow_ops.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+__all__ = ['equal']
+
+import dragon.ops as ops
+
+
+def equal(a, b, name=None):
+
+    return ops.Equal([a, b])
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/dtypes.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import numpy as np
+
+__all__ = ['int32', 'int64', 'float32', 'bool']
+
+int32 = np.int32
+int64 = np.int64
+float32 = np.float32
+bool = np.bool
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/gradients_impl.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.core.tensor import Tensor
+import dragon.vm.theano.tensor as T
+
+def gradients(ys, xs, **kwargs):
+    """Compute the gradients for variables with respect to the cost.
+
+    Parameters
+    ----------
+    ys : Tensor or list of Tensor
+        The tensor(s) to be differentiated.
+    xs : Tensor or list of Tensor
+        The tensor(s to be used for differentiation.
+
+    Returns
+    -------
+    Tensor or list of Tensor
+        The gradients of variables.
+
+    """
+    if not isinstance(ys, list):
+        ys = [ys]
+    for y in ys:
+        dxs = T.grad(y, xs)
+    return dxs
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/init_ops.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+__all__ = [
+    'zeros_initializer',
+    'ones_initializer',
+    'constant_initializer',
+    'random_uniform_initializer',
+    'random_normal_initializer',
+    'truncated_normal_initializer',
+    'variance_scaling_initializer',
+    'glorot_uniform_initializer',
+    'glorot_normal_initializer',
+]
+
+import dragon.ops as ops
+
+from dragon.vm.tensorflow.framework import dtypes
+
+
+class Initializer(object):
+    """
+    The basic Initializer.
+    """
+    def __call__(self, shape, dtype=None, **kwargs):
+        raise NotImplementedError
+
+
+class Zeros(Initializer):
+    """The initializer that sets tensors to 0.
+
+    Parameters
+    ----------
+    shape : list, tuple or Tensor
+        The shape of the initializer.
+    dtype : DType
+        The data type.
+
+    Returns
+    -------
+    Tensor
+        The initializer.
+
+    """
+    def __init__(self, dtype=dtypes.float32):
+        self.dtype = dtype
+
+    def __call__(self, shape, dtype=None, **kwargs):
+        if dtype is None: dtype = self.dtype
+        return ops.Fill(shape, value=0)
+
+
+class Ones(Initializer):
+    """The initializer that sets tensors to 1.
+
+    Parameters
+    ----------
+    shape : list, tuple or Tensor
+        The shape of the initializer.
+    dtype : DType
+        The data type.
+
+    Returns
+    -------
+    Tensor
+        The initializer.
+
+    """
+    def __init__(self, dtype=dtypes.float32):
+        self.dtype = dtype
+
+    def __call__(self, shape, dtype=None, **kwargs):
+        if dtype is None: dtype = self.dtype
+        return ops.Fill(shape, value=1)
+
+
+class Constant(Initializer):
+    def __init__(self, value=0, dtype=dtypes.float32):
+        self.value = value
+        self.dtype = dtype
+
+    def __call__(self, shape, dtype=None, **kwargs):
+        if dtype is None: dtype = self.dtype
+        return ops.Fill(shape, value=self.value)
+
+
+class RandomUniform(Initializer):
+    def __init__(self, minval=0, maxval=1, dtype=dtypes.float32):
+        self.minval = minval
+        self.maxval = maxval
+        self.dtype = dtype
+
+    def __call__(self, shape, dtype=None, **kwargs):
+        if dtype is None: dtype = self.dtype
+        return ops.RandomUniform(shape, self.minval, self.maxval)
+
+
+class RandomNormal(Initializer):
+    def __init__(self, mean=0.0, stddev=1.0, dtype=dtypes.float32):
+        self.mean = mean
+        self.stddev = stddev
+        assert dtype == dtypes.float32
+        self.dtype = dtype
+
+    def __call__(self, shape, dtype=None, **kwargs):
+        if dtype is None: dtype = self.dtype
+        return ops.RandomNormal(shape, self.mean, self.stddev)
+
+
+class TruncatedNormal(Initializer):
+    def __init__(self, mean=0.0, stddev=1.0, dtype=dtypes.float32):
+        self.mean = mean
+        self.stddev = stddev
+        assert dtype == dtypes.float32
+        self.dtype = dtype
+
+    def __call__(self, shape, dtype=None, **kwargs):
+        if dtype is None: dtype = self.dtype
+        return ops.TruncatedNormal(shape, self.mean, self.stddev)
+
+
+class VarianceScaling(Initializer):
+    def __init__(self, scale=1.0,
+                 mode="fan_in",
+                 distribution="normal",
+                 dtype=dtypes.float32):
+        if scale <= 0.:
+            raise ValueError("`scale` must be positive float.")
+        if mode not in {"fan_in", "fan_out", "fan_avg"}:
+            raise ValueError("Invalid `mode` argument:", mode)
+        distribution = distribution.lower()
+        if distribution not in {"normal", "uniform"}:
+            raise ValueError("Invalid `distribution` argument:", distribution)
+        self.scale = scale
+        self.mode = mode
+        self.distribution = distribution
+        self.dtype = dtype
+
+    def __call__(self, shape, dtype=None, **kwargs):
+        if dtype is None: dtype = self.dtype
+        if self.distribution == "normal":
+            return ops.GlorotNormal(shape=shape, scale=self.scale, mode=self.mode)
+        else:
+            return ops.GlorotUniform(shape=shape, scale=self.scale, mode=self.mode)
+
+
+zeros_initializer = Zeros
+ones_initializer = Ones
+constant_initializer = Constant
+random_uniform_initializer = RandomUniform
+random_normal_initializer = RandomNormal
+truncated_normal_initializer = TruncatedNormal
+variance_scaling_initializer = VarianceScaling
+
+
+def glorot_uniform_initializer(dtype=dtypes.float32):
+
+    return variance_scaling_initializer(scale=6.0,
+                                        mode='fan_avg',
+                                        distribution='uniform',
+                                        dtype=dtype)
+
+
+def glorot_normal_initializer(dtype=dtypes.float32):
+
+    return variance_scaling_initializer(scale=2.0,
+                                        mode='fan_avg',
+                                        distribution='normal',
+                                        dtype=dtype)
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/math_ops.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+__all__ = [
+    'argmax',
+    'matmul',
+    'add',
+    'subtract',
+    'multiply',
+    'divide',
+    'sub',
+    'mul',
+    'div',
+    'log',
+    'exp',
+    'square',
+    'sqrt',
+    'reduce_sum',
+    'reduce_mean',
+    'sigmoid',
+    'tanh',
+    'add_n'
+]
+
+from six.moves import range as xrange
+
+import dragon.ops as ops
+
+
+def argmax(input, axis=None, name=None, dimension=None):
+
+    if dimension is not None:
+        if axis is not None:
+            raise ValueError("cannot specify both 'axis' and 'dimension'.")
+        axis = dimension
+    elif axis is None: axis = 0
+    return ops.Argmax(input, axis=axis, name=name)
+
+
+def matmul(a,
+           b,
+           transpose_a=False,
+           transpose_b=False,
+           name=None):
+
+    return ops.Matmul([a, b], TransA=transpose_a, TransB=transpose_b, name=name)
+
+
+def add(x, y, name=None):
+
+    return ops.Add([x, y], name=None)
+
+
+def subtract(x, y, name=None):
+
+    return ops.Sub([x, y], name=name)
+
+
+def multiply(x, y, name=None):
+
+    return ops.Mul([x, y], name=name)
+
+
+def divide(x, y, name=None):
+
+    return ops.Div([x, y], name=name)
+
+
+def mul(x, y, name=None):
+
+    return multiply(x, y, name)
+
+
+def sub(x, y, name=None):
+
+    return subtract(x, y, name)
+
+
+def div(x, y, name=None):
+
+    return divide(x, y, name=name)
+
+
+def log(x, name=None):
+
+    return ops.Log(x, name=name)
+
+
+def exp(x, name=None):
+
+    return ops.Exp(x, name=name)
+
+
+def square(x, name=None):
+
+    return ops.Square(x, name=name)
+
+
+def sqrt(x, name=None):
+
+    return ops.Pow(x, power=0.5, name=name)
+
+
+def pow(x, power, name=None):
+
+    return ops.Pow(x, power=power, name=name)
+
+
+def reduce_sum(input_tensor,
+               axis=None,
+               keep_dims=False,
+               name=None,
+               reduction_indices=None):
+
+    if reduction_indices is not None:
+        if axis is not None:
+            raise ValueError("cannot specify both 'axis' and 'reduction_indices'.")
+        axis = reduction_indices
+    elif axis is None: axis = -1 # reduce all
+    if isinstance(axis, list) or isinstance(axis, tuple): # reduce continuously
+        if len(axis) < 1:
+            raise RuntimeError('reduce axes should at least have one.')
+        if len(axis) == 1:
+            return ops.Sum(input_tensor, axis=axis[0], keep_dims=keep_dims)
+        else:
+            ret = ops.Sum(input_tensor, axis=axis[0], keep_dims=True)
+            for i in xrange(1, len(axis) - 1):
+                ret = ops.Sum(ret, axis=axis[i], keep_dims=True)
+            return ops.Sum(ret, axis=axis[len(axis) - 1], keep_dims=keep_dims)
+    else:
+        return ops.Sum(input_tensor, axis=axis, keep_dims=keep_dims)
+
+
+def reduce_mean(input_tensor,
+                axis=None,
+                keep_dims=False,
+                name=None,
+                reduction_indices=None):
+
+    if reduction_indices is not None:
+        if axis is not None:
+            raise ValueError("cannot specify both 'axis' and 'reduction_indices'.")
+        axis = reduction_indices
+    elif axis is None: axis = -1 # reduce all
+    if isinstance(axis, list) or isinstance(axis, tuple): # reduce continuously
+        if len(axis) < 1:
+            raise RuntimeError('reduce axes should at least have one.')
+        if len(axis) == 1:
+            return ops.Mean(input_tensor, axis=axis[0], keep_dims=keep_dims)
+        else:
+            ret = ops.Mean(input_tensor, axis=axis[0], keep_dims=True)
+            for i in xrange(1, len(axis) - 1):
+                ret = ops.Mean(ret, axis=axis[i], keep_dims=True)
+            return ops.Mean(ret, axis=axis[len(axis) - 1], keep_dims=keep_dims)
+    else:
+        return ops.Mean(input_tensor, axis=axis, keep_dims=keep_dims)
+
+
+def sigmoid(x, name=None):
+
+    return ops.Sigmoid(x, name=name)
+
+
+def tanh(x, name=None):
+
+    return ops.Tanh(x, name=name)
+
+
+def add_n(inputs, name=None):
+
+    return ops.Eltwise(inputs, operation='SUM', name=name)

Dragon/python/dragon/vm/tensorflow/ops/nn.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.ops.nn_ops import *
+from dragon.vm.tensorflow.ops.nn_impl import *

Dragon/python/dragon/vm/tensorflow/ops/nn_impl.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.core.tensor import Tensor
+import dragon.ops as ops
+
+def batch_normalization(x, mean, variance,
+                        offset, scale,
+                        decay=0.9, variance_epsilon=1e-3, name=None):
+    raise NotImplementedError('Deprecated. Use tf.layer.batch_normalization.')
+
+
+
+def batch_norm_with_global_normalization(t, m, v,
+                                         beta, gamma,
+                                         decay=0.9, variance_epsilon=1e-3,
+                                         scale_after_normalization=True, name=None):
+    raise NotImplementedError('Deprecated. Use tf.layer.batch_normalization.')
+
+
+def l2_normalize(x, dim, epsilon=1e-12, name=None):
+    return ops.L2Norm(inputs=x,
+                      axis=dim,
+                      num_axes=1,
+                      eps=epsilon)

Dragon/python/dragon/vm/tensorflow/ops/nn_ops.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.core.tensor import Tensor
+import dragon.ops as ops
+
+__all__ = [
+    'convolution',
+    'relu',
+    'softmax',
+    'conv2d',
+    'conv2d_transpose',
+    'avg_pool',
+    'max_pool',
+    'xw_plus_b',
+    'bias_add',
+    'dropout',
+    'sigmoid_cross_entropy_with_logits',
+    'softmax_cross_entropy_with_logits',
+    'sparse_softmax_cross_entropy_with_logits',
+    'l2_loss'
+]
+
+
+def convolution(input, filter, padding, strides=None,
+                dilation_rate=None, name=None, data_format=None):
+    num_total_dims = filter.get_shape().ndims
+    if num_total_dims is None:
+        num_total_dims = input.get_shape().ndims
+    if num_total_dims is None:
+        raise ValueError("rank of input or filter must be known.")
+    num_spatial_dims = num_total_dims - 2
+
+    # make default parameters
+    if data_format is None:
+        data_format = 'NHWC'
+    if strides is None:
+        strides = [1] * num_total_dims
+    else:
+        if len(strides) != num_total_dims:
+            _strides = [1] * num_total_dims
+            _n_provides = len(strides)
+            if data_format == 'NHWC':
+                _strides[1 : 1 + _n_provides] = strides
+            else:
+                _strides[2 : 2 + _n_provides] = strides
+            strides = _strides
+    if dilation_rate is not None:
+        if len(dilation_rate) != num_total_dims:
+            _dilation_rate = [1] * num_total_dims
+            _n_provides = len(dilation_rate)
+            if data_format == 'NHWC':
+                _dilation_rate[1 : 1 + _n_provides] = dilation_rate
+            else:
+                _dilation_rate[2 : 2 + _n_provides] = dilation_rate
+            dilation_rate = _dilation_rate
+
+    if num_spatial_dims == 2:
+        return conv2d(input, filter,
+                      strides, padding, dilation_rate,
+                      data_format, name)
+    else:
+        raise NotImplementedError('conv{}d is not implemented.'.format(num_spatial_dims))
+
+
+def relu(features, name=None):
+
+   return ops.Relu(features, name=name)
+
+
+def softmax(logits, dim=-1, name=None):
+
+    return ops.Softmax(logits, axis=dim)
+
+
+def conv2d(input, filter, strides, padding, dilation_rate=None,
+           data_format='NHWC', name=None, **kwargs):
+    """Compute 2D convolution according to the given 4D ``input`` and ``filter``.
+
+    For **NHWC** format, filter should be as ``[filter_height, filter_width, in_channels, out_channels]``.
+
+    For **NCHW** format, filter should be as ``[out_channels, in_channels, filter_height, filter_width]``.
+
+    Parameters
+    ----------
+    input : Tensor
+        The input tensor.
+    filter : Tensor
+        The filter tensor.
+    strides : list of int
+        The strides with length 4.
+    padding : str
+        The padding algorithm. ``VALID`` or ``SAME``.
+    dilation_rate : list of int or None
+        The dilation rates with with length 4.
+    data_format : str
+        The data format. ``NHWC`` or ``NCHW``.
+    name : str
+        The optional name for this operator.
+
+    Returns
+    -------
+    Tensor
+        The output tensor.
+
+    """
+    if filter.shape is None:
+        raise ValueError('filter must have a valid shape.')
+    else:
+        if len(filter.shape) != 4:
+            raise ValueError('filter must be a 4D Tensor.')
+    if len(strides) != 4:
+        raise ValueError('strides must be a list with length 4.')
+    if dilation_rate is not None:
+        if len(dilation_rate) != 4:
+            raise ValueError(' dilation_rate must be a list with length 4.')
+
+    if data_format == 'NHWC':
+        output = ops.Conv2d([input, filter],
+                            num_output=filter.shape[3],
+                            kernel_size=filter.shape[0:2],
+                            stride=strides[1:3],
+                            dilation=dilation_rate[1:3] if dilation_rate is not None else 1,
+                            padding=padding,
+                            data_format=data_format)
+        return output
+    elif data_format == 'NCHW':
+        output = ops.Conv2d([input, filter],
+                            num_output=filter.shape[0],
+                            kernel_size=filter.shape[2:4],
+                            stride=strides[2:4],
+                            dilation=dilation_rate[2:4] if dilation_rate is not None else 1,
+                            padding=padding,
+                            data_format=data_format)
+        return output
+    else:
+        raise ValueError('Unknown data format: {}'.format(data_format))
+
+
+def conv2d_transpose(value, filter, output_shape, strides, padding='SAME',
+                     data_format='NHWC', name=None):
+    """Compute 2D deconvolution according to the given 4D ``input`` and ``filter``.
+
+    For **NHWC** format, filter should be as ``[filter_height, filter_width, out_channels, in_channels]``.
+
+    For **NCHW** format, filter should be as ``[in_channels, out_channels, filter_height, filter_width]``.
+
+    ``output_shape`` will be ignored if padding algorithm is **VALID**.
+
+    Parameters
+    ----------
+    input : Tensor
+        The input tensor.
+    filter : Tensor
+        The filter tensor.
+    output_shape : list of int
+        The deterministic output shape for **SAME** padding.
+    strides : list of int
+        The strides with length 4.
+    padding : str
+        The padding algorithm. ``VALID`` or ``SAME``.
+    data_format : str
+        The data format. ``NHWC`` or ``NCHW``.
+    name : str
+        The optional name for this operator.
+
+    Returns
+    -------
+    Tensor
+        The output tensor.
+
+    """
+    if filter.shape is None:
+        raise ValueError('filter must have a valid shape.')
+    else:
+        if len(filter.shape) != 4:
+            raise ValueError('filter must be a 4D Tensor.')
+    if len(strides) != 4:
+        raise ValueError('strides must be a list with length 4.')
+    if not isinstance(output_shape, list):
+        raise TypeError('output_shape should be a list.')
+    if len(output_shape) != 4:
+        raise ValueError('output_shape should be a list with length 4.')
+
+    if data_format == 'NHWC':
+        output = ops.Conv2dTranspose([value, filter],
+                                     num_output=filter.shape[2],
+                                     kernel_size=filter.shape[0:2],
+                                     stride=strides[1:3],
+                                     padding=padding,
+                                     data_format=data_format,
+                                     output_shape=output_shape)
+        return output
+    elif data_format == 'NCHW':
+        output = ops.Conv2dTranspose([value, filter],
+                                     num_output=filter.shape[1],
+                                     kernel_size=filter.shape[2:4],
+                                     stride=strides[2:4],
+                                     padding=padding,
+                                     data_format=data_format,
+                                     output_shape=output_shape)
+        return output
+    else:
+        raise ValueError('Unknown data format: {}'.format(data_format))
+
+
+def avg_pool(value, ksize, strides, padding, data_format='NHWC', name=None):
+    """Perform avg pooling on spatial axes.
+
+    Parameters
+    ----------
+    value : Tensor
+        The input tensor.
+    ksize : list of int
+        The kernel size with length >= 4.
+    strides : list of int
+        The strides with length >= 4.
+    padding : str
+        The padding algorithm. ``VALID`` or ``SAME``.
+    data_format : str
+        The data format. ``NHWC`` or ``NCHW``.
+    name : None or str
+        The optional name of op.
+
+    Returns
+    -------
+    Tensor
+        The output tensor.
+
+    """
+    if len(ksize) < 4:
+        raise ValueError('ksize must be a list with length >=4.')
+    if len(strides) < 4:
+        raise ValueError('strides must be a list with length >=4.')
+    if len(ksize) != len(strides):
+        raise ValueError('ksize and strides should have the same length.')
+    if len(ksize) == 4:
+        if data_format == 'NHWC':
+            if ksize[0] != 1 or ksize[3] != 1 or strides[0] != 1 or strides[3] != 1:
+                raise ValueError('The pooling can only be performed on spatial axes.')
+            return ops.Pool2d(value, [ksize[1], ksize[2]], [strides[1], strides[2]],
+                              padding=padding, data_format=data_format, mode='AVG')
+        if data_format == 'NCHW':
+            if ksize[0] != 1 or ksize[1] != 1 or strides[0] != 1 or strides[1] != 1:
+                raise ValueError('The pooling can only be performed on spatial axes.')
+            return ops.Pool2d(value, [ksize[2], ksize[3]], [strides[2], strides[3]],
+                              padding=padding, data_format=data_format, mode='AVG')
+    else:
+        raise NotImplementedError('Pool{}d has not been implemented yet.'.format(len(ksize) - 2))
+
+
+def max_pool(value, ksize, strides, padding, data_format='NHWC', name=None):
+    """Perform max pooling on spatial axes.
+
+    Parameters
+    ----------
+    value : Tensor
+        The input tensor.
+    ksize : list of int
+        The kernel size with length >= 4.
+    strides : list of int
+        The strides with length >= 4.
+    padding : str
+        The padding algorithm. ``VALID`` or ``SAME``.
+    data_format : str
+        The data format. ``NHWC`` or ``NCHW``.
+    name : None or str
+        The optional name of op.
+
+    Returns
+    -------
+    Tensor
+        The output tensor.
+
+    """
+    if len(ksize) < 4:
+        raise ValueError('ksize must be a list with length >=4.')
+    if len(strides) < 4:
+        raise ValueError('strides must be a list with length >=4.')
+    if len(ksize) != len(strides):
+        raise ValueError('ksize and strides should have the same length.')
+    if len(ksize) == 4:
+        if data_format == 'NHWC':
+            if ksize[0] != 1 or ksize[3] != 1 or strides[0] != 1 or strides[3] != 1:
+                raise ValueError('The pooling can only be performed on spatial axes.')
+            return ops.Pool2d(value, [ksize[1], ksize[2]], [strides[1], strides[2]],
+                              padding=padding, data_format=data_format, mode='MAX')
+        if data_format == 'NCHW':
+            if ksize[0] != 1 or ksize[1] != 1 or strides[0] != 1 or strides[1] != 1:
+                raise ValueError('The pooling can only be performed on spatial axes.')
+            return ops.Pool2d(value, [ksize[2], ksize[3]], [strides[2], strides[3]],
+                              padding=padding, data_format=data_format, mode='MAX')
+    else:
+        raise NotImplementedError('Pool{}d has not been implemented yet.'.format(len(ksize) - 2))
+
+
+def xw_plus_b(x, weights, biases, name=None):
+
+    if weights.shape is None:
+        raise ValueError('weights must have a valid shape.')
+    else:
+        if len(weights.shape) != 2:
+            raise ValueError('weights must be a 2D Tensor')
+
+    if biases.shape is None:
+        raise ValueError('biases must a have a valid shape.')
+    else:
+        if len(biases.shape) != 1:
+            raise ValueError('biases must be a 1D Tensor')
+        if weights.shape[1] != biases.shape[0]:
+            raise ValueError('the shape of weights and biaes are incompatible.')
+
+    return ops.InnerProduct([x, weights, biases], num_output=weights.shape[1], TransW=False)
+
+
+def bias_add(value, bias, data_format='NHWC', name=None):
+    return ops.BiasAdd([value, bias], data_format=data_format)
+
+def sigmoid_cross_entropy_with_logits(logits, targets, name=None):
+
+    return ops.SigmoidCrossEntropy([logits, targets], normalization='UNIT', name=None)
+
+
+def softmax_cross_entropy_with_logits(_sentinel=None,
+                                      labels=None, logits=None,
+                                      dim=-1, name=None):
+
+    if _sentinel is not None:
+        raise ValueError('Only call `softmax_cross_entropy_with_logits` '
+                         'with named arguments (labels=..., logits=..., ...)')
+
+    if dim == -1: dim = 1
+    return ops.SoftmaxCrossEntropy([logits, labels], axis=dim, normalization='UNIT', name=name)
+
+
+def sparse_softmax_cross_entropy_with_logits(logits, labels, dim=-1, name=None):
+
+    if dim == -1: dim = 1
+    return ops.SparseSoftmaxCrossEntropy([logits, labels], axis=dim, normalization='UNIT', name=name)
+
+
+def l2_loss(t, name=None):
+    return (ops.Reduce(ops.Square(t), operation='SUM') * 0.5)
+
+
+def dropout(x, keep_prob, name=None):
+    return ops.Dropout(x, 1 - keep_prob)

Dragon/python/dragon/vm/tensorflow/ops/random_ops.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+__all__ = [
+    'random_normal',
+    'truncated_normal',
+    'random_uniform'
+]
+
+import dragon.ops as ops
+
+from dragon.vm.tensorflow.framework import dtypes
+
+
+def random_normal(shape,
+                  mean=0.0,
+                  stddev=1.0,
+                  dtype=dtypes.float32,
+                  seed=None,
+                  name=None):
+    return ops.RandomNormal(shape, mean, stddev)
+
+
+def truncated_normal(shape,
+                     mean=0.0,
+                     stddev=1.0,
+                     dtype=dtypes.float32,
+                     seed=None,
+                     name=None):
+    return ops.TruncatedNormal(shape, mean, stddev)
+
+
+def random_uniform(shape,
+                   minval=0,
+                   maxval=None,
+                   dtype=dtypes.float32,
+                   seed=None,
+                   name=None):
+    return ops.RandomUniform(shape, minval, maxval)
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/standard_ops.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.ops.variables import *
+from dragon.vm.tensorflow.ops.var_scope import *
+
+# Ops
+from dragon.vm.tensorflow.ops.init_ops import *
+from dragon.vm.tensorflow.ops.random_ops import *
+from dragon.vm.tensorflow.ops.math_ops import *
+from dragon.vm.tensorflow.ops.array_ops import *
+from dragon.vm.tensorflow.ops.control_flow_ops import *
+from dragon.vm.tensorflow.ops.nn_ops import *
+from dragon.vm.tensorflow.ops.gradients_impl import gradients
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/var_scope.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.framework import dtypes
+from dragon.vm.tensorflow.ops.variables import Variable
+from dragon.vm.tensorflow.ops import init_ops
+
+_VARSCOPE = None
+_VARSTORE = {}
+
+class VariableScope(object):
+    """
+    Construct a Variable.
+    """
+    def __init__(self, reuse, name='', name_scope='', **kwargs):
+        self._name = name
+        self._reuse = reuse
+        self._name_scope = name_scope
+        if self._name_scope is None:
+            self._name_scope = ''
+        self._old_varscope = None
+
+    @property
+    def reuse(self):
+        return self._reuse
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def original_name_scope(self):
+        return self._name_scope
+
+    def get_variable(self, name, shape=None, dtype=None, initializer=None,
+                     trainable=True, collections=None, validate_shape=True, **kwargs):
+        global _VARSTORE
+
+        # get full name
+        from dragon.core.scope import get_tensor_scope
+        full_name = get_tensor_scope() + name
+
+        # create a new variable
+        if not full_name in _VARSTORE:
+            if shape is None:
+                raise ValueError('Must specific a shape for the Variable({}).'.format(full_name))
+            if initializer is None:
+                initializer = self._get_default_initializer(name, shape=shape, dtype=dtype)
+            initial_value = initializer(shape, dtype=dtype)
+            new_var = Variable(initial_value, trainable=trainable, collections=collections,
+                            validate_shape=validate_shape, name=name, dtype=dtype)
+            _VARSTORE[full_name] = new_var
+            return new_var
+        else:
+            # existing ?
+            if self._reuse:
+                return _VARSTORE[full_name]
+            raise ValueError('The Variable({}) already exists.'.format(full_name))
+
+    def __enter__(self):
+        global _VARSCOPE
+        self._old_varscope = _VARSCOPE
+        _VARSCOPE = self
+
+        from dragon.core.scope import get_tensor_scope, set_tensor_scope
+        prefix = self._name_scope + '/' if self._name_scope != '' else ''
+        set_tensor_scope(get_tensor_scope() + prefix)
+        return self
+
+    def __exit__(self, type, value, traceback):
+        global _VARSCOPE
+        _VARSCOPE = self._old_varscope
+
+        from dragon.core.scope import get_tensor_scope, set_tensor_scope
+        prefix = self._name_scope + '/' if self._name_scope != '' else ''
+        assert get_tensor_scope().endswith(prefix)
+        if self._name_scope != '':
+            set_tensor_scope(get_tensor_scope()[:-len(prefix)])
+
+    def _get_default_initializer(self, name, shape=None, dtype=dtypes.float32):
+        if dtype is None: dtype = dtypes.float32
+        if dtype.is_floating:
+            initializer = init_ops.glorot_uniform_initializer()
+        elif dtype.is_integer or dtype.is_unsigned or dtype.is_bool:
+            initializer = init_ops.zeros_initializer()(shape=shape, dtype=dtype.base_dtype)
+        else:
+            raise ValueError('An initializer for Variable({}) of %s is required.'.
+                             format(name, dtype.base_dtype))
+        return initializer
+
+
+def get_variable_scope():
+    global _VARSCOPE
+    if _VARSCOPE is None:
+        _VARSCOPE = VariableScope(False)
+    return _VARSCOPE
+
+
+def variable_scope(name_scope, reuse=None, **kwargs):
+    return VariableScope(reuse, name_scope=name_scope)
+
+
+def get_variable(name, shape=None, dtype=None, initializer=None,
+                 trainable=True, collections=None, validate_shape=True, **kwargs):
+    return get_variable_scope().get_variable(name, shape=shape, dtype=dtype,
+                                             initializer=initializer,trainable=trainable,
+                                             collections=collections, validate_shape=validate_shape)
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/ops/variables.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import copy
+
+from dragon.core.tensor import Tensor
+import dragon.vm.theano as theano
+
+from dragon.vm.tensorflow.framework import ops
+from dragon.vm.tensorflow.framework import dtypes
+from dragon.vm.tensorflow.util.deprecation import deprecated
+
+
+class Variable(Tensor):
+    """
+    Construct a Variable.
+    """
+    def __init__(self, initial_value=None, trainable=True,
+                 collections=None, validate_shape=True,
+                 name=None, dtype=None, **kwargs):
+        super(Variable, self).__init__()
+
+        if initial_value is None:
+            raise ValueError('initial_value must be specified.')
+
+        if collections is None:
+            collections = [ops.GraphKeys.GLOBAL_VARIABLES]
+        if not isinstance(collections, (list, tuple, set)):
+            raise ValueError('collections argument to Variable constructor must be a list, tuple, '
+                             'or set. Got the type {}'.format(type(collections)))
+        if trainable and ops.GraphKeys.TRAINABLE_VARIABLES not in collections:
+            collections = list(collections) + [ops.GraphKeys.TRAINABLE_VARIABLES]
+
+        # initialization
+        if isinstance(initial_value, Tensor):
+            self.clone(initial_value)
+            if name is not None:
+                self.name = name
+                self.expressions.values()[0].output[0] = self.name
+        else:
+            # from ..ops.constant_op import constant
+            # initial_value = constant(initial_value, name=name)
+            # self.clone(initial_value)
+            pass
+
+        # check data type
+        if dtype is not None:
+            if not isinstance(dtype, dtypes.DType):
+                raise TypeError('The dtype should be a valid tf data type.')
+            self.dtype = dtype.name
+
+        # registration
+        self.Variable()
+
+        if validate_shape:
+            initial_value_shape = self.shape
+            if initial_value_shape is None:
+                raise ValueError('initial_value must have a shape specified.')
+
+        ops.add_to_collections(collections, copy.deepcopy(self))
+        self.expressions = {}
+
+
+def global_variables():
+    return ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+
+
+def local_variables():
+    return ops.get_collection(ops.GraphKeys.LOCAL_VARIABLES)
+
+
+def model_variables():
+    return ops.get_collection(ops.GraphKeys.MODEL_VARIABLES)
+
+
+def trainable_variables():
+    return ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+
+
+class VariablesInitializer(object):
+    def __init__(self, var_list):
+        self.var_list = var_list
+
+    def run(self):
+        if not hasattr(self, '_init_func'):
+            self._init_func = theano.function(outputs=self.var_list)
+        self._init_func()
+
+
+def variables_initializer(var_list, name="init"):
+    return VariablesInitializer(var_list)
+
+
+def global_variables_initializer():
+    return variables_initializer(global_variables())
+
+
+@deprecated("2017-03-02", "Use `tf.global_variables_initializer` instead.")
+def initialize_all_variables():
+    """
+    See ``tf.global_variables_initializer``.
+    """
+    return global_variables_initializer()

Dragon/python/dragon/vm/tensorflow/protobuf/config.proto

+syntax = "proto2";
+
+package tensorflow;
+
+message GPUOptions {
+  // A value between 0 and 1 that indicates what fraction of the
+  // available GPU memory to pre-allocate for each process.  1 means
+  // to pre-allocate all of the GPU memory, 0.5 means the process
+  // allocates ~50% of the available GPU memory.
+  optional double per_process_gpu_memory_fraction = 1;
+
+  // The type of GPU allocation strategy to use.
+  //
+  // Allowed values:
+  // "": The empty string (default) uses a system-chosen default
+  //     which may change over time.
+  //
+  // "BFC": A "Best-fit with coalescing" algorithm, simplified from a
+  //        version of dlmalloc.
+  optional string allocator_type = 2;
+
+  // Delay deletion of up to this many bytes to reduce the number of
+  // interactions with gpu driver code.  If 0, the system chooses
+  // a reasonable default (several MBs).
+  optional int64 deferred_deletion_bytes = 3;
+
+  // If true, the allocator does not pre-allocate the entire specified
+  // GPU memory region, instead starting small and growing as needed.
+  optional bool allow_growth = 4;
+
+  // A comma-separated list of GPU ids that determines the 'visible'
+  // to 'virtual' mapping of GPU devices.  For example, if TensorFlow
+  // can see 8 GPU devices in the process, and one wanted to map
+  // visible GPU devices 5 and 3 as "/device:GPU:0", and "/device:GPU:1", then one
+  // would specify this field as "5,3".  This field is similar in
+  // spirit to the CUDA_VISIBLE_DEVICES environment variable, except
+  // it applies to the visible GPU devices in the process.
+  //
+  // NOTE: The GPU driver provides the process with the visible GPUs
+  // in an order which is not guaranteed to have any correlation to
+  // the *physical* GPU id in the machine.  This field is used for
+  // remapping "visible" to "virtual", which means this operates only
+  // after the process starts.  Users are required to use vendor
+  // specific mechanisms (e.g., CUDA_VISIBLE_DEVICES) to control the
+  // physical to visible device mapping prior to invoking TensorFlow.
+  optional string visible_device_list = 5;
+
+  // In the event polling loop sleep this many microseconds between
+  // PollEvents calls, when the queue is not empty.  If value is not
+  // set or set to 0, gets set to a non-zero default.
+  optional int32 polling_active_delay_usecs = 6;
+
+  // In the event polling loop sleep this many millisconds between
+  // PollEvents calls, when the queue is empty.  If value is not
+  // set or set to 0, gets set to a non-zero default.
+  optional int32 polling_inactive_delay_msecs = 7;
+
+  // Force all tensors to be gpu_compatible. On a GPU-enabled TensorFlow,
+  // enabling this option forces all CPU tensors to be allocated with Cuda
+  // pinned memory. Normally, TensorFlow will infer which tensors should be
+  // allocated as the pinned memory. But in case where the inference is
+  // incomplete, this option can significantly speed up the cross-device memory
+  // copy performance as long as it fits the memory.
+  // Note that this option is not something that should be
+  // enabled by default for unknown or very large models, since all Cuda pinned
+  // memory is unpageable, having too much pinned memory might negatively impact
+  // the overall host system performance.
+  optional bool force_gpu_compatible = 8;
+}
+
+message GraphOptions {
+  // If true, use control flow to schedule the activation of Recv nodes.
+  // (Currently ignored.)
+  optional bool enable_recv_scheduling = 2;
+
+  // Options controlling how graph is optimized.
+  // OptimizerOptions optimizer_options = 3;
+
+  // The number of steps to run before returning a cost model detailing
+  // the memory usage and performance of each node of the graph. 0 means
+  // no cost model.
+  optional int64 build_cost_model = 4;
+
+  // The number of steps to skip before collecting statistics for the
+  // cost model.
+  optional int64 build_cost_model_after = 9;
+
+  // Annotate each Node with Op output shape data, to the extent it can
+  // be statically inferred.
+  optional bool infer_shapes = 5;
+
+  // Only place the subgraphs that are run, rather than the entire graph.
+  //
+  // This is useful for interactive graph building, where one might
+  // produce graphs that cannot be placed during the debugging
+  // process.  In particular, it allows the client to continue work in
+  // a session after adding a node to a graph whose placement
+  // constraints are unsatisfiable.
+  optional bool place_pruned_graph = 6;
+
+  // If true, transfer float values between processes as bfloat16.
+  optional bool enable_bfloat16_sendrecv = 7;
+
+  // If > 0, record a timeline every this many steps.
+  // EXPERIMENTAL: This currently has no effect in MasterSession.
+  optional int32 timeline_step = 8;
+
+  // Options that control the type and amount of graph rewriting.
+  // Not currently configurable via the public Python API (i.e. there is no API
+  // stability guarantee if you import RewriterConfig explicitly).
+  // RewriterConfig rewrite_options = 10;
+}
+
+message ConfigProto {
+  // Map from device type name (e.g., "CPU" or "GPU" ) to maximum
+  // number of devices of that type to use.  If a particular device
+  // type is not found in the map, the system picks an appropriate
+  // number.
+  // map<string, int32> device_count = 1;
+
+  // The execution of an individual op (for some op types) can be
+  // parallelized on a pool of intra_op_parallelism_threads.
+  // 0 means the system picks an appropriate number.
+  optional int32 intra_op_parallelism_threads = 2;
+
+  // Nodes that perform blocking operations are enqueued on a pool of
+  // inter_op_parallelism_threads available in each process.
+  //
+  // 0 means the system picks an appropriate number.
+  //
+  // Note that the first Session created in the process sets the
+  // number of threads for all future sessions unless use_per_session_threads is
+  // true or session_inter_op_thread_pool is configured.
+  optional int32 inter_op_parallelism_threads = 5;
+
+  // If true, use a new set of threads for this session rather than the global
+  // pool of threads. Only supported by direct sessions.
+  //
+  // If false, use the global threads created by the first session, or the
+  // per-session thread pools configured by session_inter_op_thread_pool.
+  //
+  // This option is deprecated. The same effect can be achieved by setting
+  // session_inter_op_thread_pool to have one element, whose num_threads equals
+  // inter_op_parallelism_threads.
+  optional bool use_per_session_threads = 9;
+
+  // This option is experimental - it may be replaced with a different mechanism
+  // in the future.
+  //
+  // Configures session thread pools. If this is configured, then RunOptions for
+  // a Run call can select the thread pool to use.
+  //
+  // The intended use is for when some session invocations need to run in a
+  // background pool limited to a small number of threads:
+  // - For example, a session may be configured to have one large pool (for
+  // regular compute) and one small pool (for periodic, low priority work);
+  // using the small pool is currently the mechanism for limiting the inter-op
+  // parallelism of the low priority work.  Note that it does not limit the
+  // parallelism of work spawned by a single op kernel implementation.
+  // - Using this setting is normally not needed in training, but may help some
+  // serving use cases.
+  // - It is also generally recommended to set the global_name field of this
+  // proto, to avoid creating multiple large pools. It is typically better to
+  // run the non-low-priority work, even across sessions, in a single large
+  // pool.
+  // repeated ThreadPoolOptionProto session_inter_op_thread_pool = 12;
+
+  // Assignment of Nodes to Devices is recomputed every placement_period
+  // steps until the system warms up (at which point the recomputation
+  // typically slows down automatically).
+  optional int32 placement_period = 3;
+
+  // When any filters are present sessions will ignore all devices which do not
+  // match the filters. Each filter can be partially specified, e.g. "/job:ps"
+  // "/job:worker/replica:3", etc.
+  repeated string device_filters = 4;
+
+  // Options that apply to all GPUs.
+  optional GPUOptions gpu_options = 6;
+
+  // Whether soft placement is allowed. If allow_soft_placement is true,
+  // an op will be placed on CPU if
+  //   1. there's no GPU implementation for the OP
+  // or
+  //   2. no GPU devices are known or registered
+  // or
+  //   3. need to co-locate with reftype input(s) which are from CPU.
+  optional bool allow_soft_placement = 7;
+
+  // Whether device placements should be logged.
+  optional bool log_device_placement = 8;
+
+  // Options that apply to all graphs.
+  optional GraphOptions graph_options = 10;
+
+  // Global timeout for all blocking operations in this session.  If non-zero,
+  // and not overridden on a per-operation basis, this value will be used as the
+  // deadline for all blocking operations.
+  optional int64 operation_timeout_in_ms = 11;
+
+  // Options that apply when this session uses the distributed runtime.
+  // RPCOptions rpc_options = 13;
+
+  // Optional list of all workers to use in this session.
+  // ClusterDef cluster_def = 14;
+
+  // Next: 15
+}
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/protobuf/config_pb2.py

+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# source: config.proto
+
+import sys
+_b=sys.version_info[0]<3 and (lambda x:x) or (lambda x:x.encode('latin1'))
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import message as _message
+from google.protobuf import reflection as _reflection
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf import descriptor_pb2
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+
+
+
+DESCRIPTOR = _descriptor.FileDescriptor(
+  name='config.proto',
+  package='tensorflow',
+  serialized_pb=_b('\n\x0c\x63onfig.proto\x12\ntensorflow\"\x89\x02\n\nGPUOptions\x12\'\n\x1fper_process_gpu_memory_fraction\x18\x01 \x01(\x01\x12\x16\n\x0e\x61llocator_type\x18\x02 \x01(\t\x12\x1f\n\x17\x64\x65\x66\x65rred_deletion_bytes\x18\x03 \x01(\x03\x12\x14\n\x0c\x61llow_growth\x18\x04 \x01(\x08\x12\x1b\n\x13visible_device_list\x18\x05 \x01(\t\x12\"\n\x1apolling_active_delay_usecs\x18\x06 \x01(\x05\x12$\n\x1cpolling_inactive_delay_msecs\x18\x07 \x01(\x05\x12\x1c\n\x14\x66orce_gpu_compatible\x18\x08 \x01(\x08\"\xd3\x01\n\x0cGraphOptions\x12\x1e\n\x16\x65nable_recv_scheduling\x18\x02 \x01(\x08\x12\x18\n\x10\x62uild_cost_model\x18\x04 \x01(\x03\x12\x1e\n\x16\x62uild_cost_model_after\x18\t \x01(\x03\x12\x14\n\x0cinfer_shapes\x18\x05 \x01(\x08\x12\x1a\n\x12place_pruned_graph\x18\x06 \x01(\x08\x12 \n\x18\x65nable_bfloat16_sendrecv\x18\x07 \x01(\x08\x12\x15\n\rtimeline_step\x18\x08 \x01(\x05\"\xe7\x02\n\x0b\x43onfigProto\x12$\n\x1cintra_op_parallelism_threads\x18\x02 \x01(\x05\x12$\n\x1cinter_op_parallelism_threads\x18\x05 \x01(\x05\x12\x1f\n\x17use_per_session_threads\x18\t \x01(\x08\x12\x18\n\x10placement_period\x18\x03 \x01(\x05\x12\x16\n\x0e\x64\x65vice_filters\x18\x04 \x03(\t\x12+\n\x0bgpu_options\x18\x06 \x01(\x0b\x32\x16.tensorflow.GPUOptions\x12\x1c\n\x14\x61llow_soft_placement\x18\x07 \x01(\x08\x12\x1c\n\x14log_device_placement\x18\x08 \x01(\x08\x12/\n\rgraph_options\x18\n \x01(\x0b\x32\x18.tensorflow.GraphOptions\x12\x1f\n\x17operation_timeout_in_ms\x18\x0b \x01(\x03')
+)
+_sym_db.RegisterFileDescriptor(DESCRIPTOR)
+
+
+
+
+_GPUOPTIONS = _descriptor.Descriptor(
+  name='GPUOptions',
+  full_name='tensorflow.GPUOptions',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='per_process_gpu_memory_fraction', full_name='tensorflow.GPUOptions.per_process_gpu_memory_fraction', index=0,
+      number=1, type=1, cpp_type=5, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='allocator_type', full_name='tensorflow.GPUOptions.allocator_type', index=1,
+      number=2, type=9, cpp_type=9, label=1,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='deferred_deletion_bytes', full_name='tensorflow.GPUOptions.deferred_deletion_bytes', index=2,
+      number=3, type=3, cpp_type=2, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='allow_growth', full_name='tensorflow.GPUOptions.allow_growth', index=3,
+      number=4, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='visible_device_list', full_name='tensorflow.GPUOptions.visible_device_list', index=4,
+      number=5, type=9, cpp_type=9, label=1,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='polling_active_delay_usecs', full_name='tensorflow.GPUOptions.polling_active_delay_usecs', index=5,
+      number=6, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='polling_inactive_delay_msecs', full_name='tensorflow.GPUOptions.polling_inactive_delay_msecs', index=6,
+      number=7, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='force_gpu_compatible', full_name='tensorflow.GPUOptions.force_gpu_compatible', index=7,
+      number=8, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=29,
+  serialized_end=294,
+)
+
+
+_GRAPHOPTIONS = _descriptor.Descriptor(
+  name='GraphOptions',
+  full_name='tensorflow.GraphOptions',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='enable_recv_scheduling', full_name='tensorflow.GraphOptions.enable_recv_scheduling', index=0,
+      number=2, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='build_cost_model', full_name='tensorflow.GraphOptions.build_cost_model', index=1,
+      number=4, type=3, cpp_type=2, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='build_cost_model_after', full_name='tensorflow.GraphOptions.build_cost_model_after', index=2,
+      number=9, type=3, cpp_type=2, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='infer_shapes', full_name='tensorflow.GraphOptions.infer_shapes', index=3,
+      number=5, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='place_pruned_graph', full_name='tensorflow.GraphOptions.place_pruned_graph', index=4,
+      number=6, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='enable_bfloat16_sendrecv', full_name='tensorflow.GraphOptions.enable_bfloat16_sendrecv', index=5,
+      number=7, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='timeline_step', full_name='tensorflow.GraphOptions.timeline_step', index=6,
+      number=8, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=297,
+  serialized_end=508,
+)
+
+
+_CONFIGPROTO = _descriptor.Descriptor(
+  name='ConfigProto',
+  full_name='tensorflow.ConfigProto',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='intra_op_parallelism_threads', full_name='tensorflow.ConfigProto.intra_op_parallelism_threads', index=0,
+      number=2, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='inter_op_parallelism_threads', full_name='tensorflow.ConfigProto.inter_op_parallelism_threads', index=1,
+      number=5, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='use_per_session_threads', full_name='tensorflow.ConfigProto.use_per_session_threads', index=2,
+      number=9, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='placement_period', full_name='tensorflow.ConfigProto.placement_period', index=3,
+      number=3, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='device_filters', full_name='tensorflow.ConfigProto.device_filters', index=4,
+      number=4, type=9, cpp_type=9, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='gpu_options', full_name='tensorflow.ConfigProto.gpu_options', index=5,
+      number=6, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='allow_soft_placement', full_name='tensorflow.ConfigProto.allow_soft_placement', index=6,
+      number=7, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='log_device_placement', full_name='tensorflow.ConfigProto.log_device_placement', index=7,
+      number=8, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='graph_options', full_name='tensorflow.ConfigProto.graph_options', index=8,
+      number=10, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='operation_timeout_in_ms', full_name='tensorflow.ConfigProto.operation_timeout_in_ms', index=9,
+      number=11, type=3, cpp_type=2, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=511,
+  serialized_end=870,
+)
+
+_CONFIGPROTO.fields_by_name['gpu_options'].message_type = _GPUOPTIONS
+_CONFIGPROTO.fields_by_name['graph_options'].message_type = _GRAPHOPTIONS
+DESCRIPTOR.message_types_by_name['GPUOptions'] = _GPUOPTIONS
+DESCRIPTOR.message_types_by_name['GraphOptions'] = _GRAPHOPTIONS
+DESCRIPTOR.message_types_by_name['ConfigProto'] = _CONFIGPROTO
+
+GPUOptions = _reflection.GeneratedProtocolMessageType('GPUOptions', (_message.Message,), dict(
+  DESCRIPTOR = _GPUOPTIONS,
+  __module__ = 'config_pb2'
+  # @@protoc_insertion_point(class_scope:tensorflow.GPUOptions)
+  ))
+_sym_db.RegisterMessage(GPUOptions)
+
+GraphOptions = _reflection.GeneratedProtocolMessageType('GraphOptions', (_message.Message,), dict(
+  DESCRIPTOR = _GRAPHOPTIONS,
+  __module__ = 'config_pb2'
+  # @@protoc_insertion_point(class_scope:tensorflow.GraphOptions)
+  ))
+_sym_db.RegisterMessage(GraphOptions)
+
+ConfigProto = _reflection.GeneratedProtocolMessageType('ConfigProto', (_message.Message,), dict(
+  DESCRIPTOR = _CONFIGPROTO,
+  __module__ = 'config_pb2'
+  # @@protoc_insertion_point(class_scope:tensorflow.ConfigProto)
+  ))
+_sym_db.RegisterMessage(ConfigProto)
+
+
+# @@protoc_insertion_point(module_scope)

Dragon/python/dragon/vm/tensorflow/protobuf/types.proto

+syntax = "proto2";
+
+package tensorflow;
+
+enum DataType {
+  // Not a legal value for DataType.  Used to indicate a DataType field
+  // has not been set.
+  DT_INVALID = 0;
+
+  // Data types that all computation devices are expected to be
+  // capable to support.
+  DT_FLOAT = 1;
+  DT_DOUBLE = 2;
+  DT_INT32 = 3;
+  DT_UINT8 = 4;
+  DT_INT16 = 5;
+  DT_INT8 = 6;
+  DT_STRING = 7;
+  DT_COMPLEX64 = 8;  // Single-precision complex
+  DT_INT64 = 9;
+  DT_BOOL = 10;
+  DT_QINT8 = 11;     // Quantized int8
+  DT_QUINT8 = 12;    // Quantized uint8
+  DT_QINT32 = 13;    // Quantized int32
+  DT_BFLOAT16 = 14;  // Float32 truncated to 16 bits.  Only for cast ops.
+  DT_QINT16 = 15;    // Quantized int16
+  DT_QUINT16 = 16;   // Quantized uint16
+  DT_UINT16 = 17;
+  DT_COMPLEX128 = 18;  // Double-precision complex
+  DT_HALF = 19;
+  DT_RESOURCE = 20;
+  DT_VARIANT = 21;  // Arbitrary C++ data types
+  DT_UINT32 = 22;
+  DT_UINT64 = 23;
+
+  // Do not use!  These are only for parameters.  Every enum above
+  // should have a corresponding value below (verified by types_test).
+  DT_FLOAT_REF = 101;
+  DT_DOUBLE_REF = 102;
+  DT_INT32_REF = 103;
+  DT_UINT8_REF = 104;
+  DT_INT16_REF = 105;
+  DT_INT8_REF = 106;
+  DT_STRING_REF = 107;
+  DT_COMPLEX64_REF = 108;
+  DT_INT64_REF = 109;
+  DT_BOOL_REF = 110;
+  DT_QINT8_REF = 111;
+  DT_QUINT8_REF = 112;
+  DT_QINT32_REF = 113;
+  DT_BFLOAT16_REF = 114;
+  DT_QINT16_REF = 115;
+  DT_QUINT16_REF = 116;
+  DT_UINT16_REF = 117;
+  DT_COMPLEX128_REF = 118;
+  DT_HALF_REF = 119;
+  DT_RESOURCE_REF = 120;
+  DT_VARIANT_REF = 121;
+  DT_UINT32_REF = 122;
+  DT_UINT64_REF = 123;
+}
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/protobuf/types_pb2.py

+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# source: types.proto
+
+import sys
+_b=sys.version_info[0]<3 and (lambda x:x) or (lambda x:x.encode('latin1'))
+from google.protobuf.internal import enum_type_wrapper
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import message as _message
+from google.protobuf import reflection as _reflection
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf import descriptor_pb2
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+
+
+
+DESCRIPTOR = _descriptor.FileDescriptor(
+  name='types.proto',
+  package='tensorflow',
+  serialized_pb=_b('\n\x0btypes.proto\x12\ntensorflow*\xaa\x06\n\x08\x44\x61taType\x12\x0e\n\nDT_INVALID\x10\x00\x12\x0c\n\x08\x44T_FLOAT\x10\x01\x12\r\n\tDT_DOUBLE\x10\x02\x12\x0c\n\x08\x44T_INT32\x10\x03\x12\x0c\n\x08\x44T_UINT8\x10\x04\x12\x0c\n\x08\x44T_INT16\x10\x05\x12\x0b\n\x07\x44T_INT8\x10\x06\x12\r\n\tDT_STRING\x10\x07\x12\x10\n\x0c\x44T_COMPLEX64\x10\x08\x12\x0c\n\x08\x44T_INT64\x10\t\x12\x0b\n\x07\x44T_BOOL\x10\n\x12\x0c\n\x08\x44T_QINT8\x10\x0b\x12\r\n\tDT_QUINT8\x10\x0c\x12\r\n\tDT_QINT32\x10\r\x12\x0f\n\x0b\x44T_BFLOAT16\x10\x0e\x12\r\n\tDT_QINT16\x10\x0f\x12\x0e\n\nDT_QUINT16\x10\x10\x12\r\n\tDT_UINT16\x10\x11\x12\x11\n\rDT_COMPLEX128\x10\x12\x12\x0b\n\x07\x44T_HALF\x10\x13\x12\x0f\n\x0b\x44T_RESOURCE\x10\x14\x12\x0e\n\nDT_VARIANT\x10\x15\x12\r\n\tDT_UINT32\x10\x16\x12\r\n\tDT_UINT64\x10\x17\x12\x10\n\x0c\x44T_FLOAT_REF\x10\x65\x12\x11\n\rDT_DOUBLE_REF\x10\x66\x12\x10\n\x0c\x44T_INT32_REF\x10g\x12\x10\n\x0c\x44T_UINT8_REF\x10h\x12\x10\n\x0c\x44T_INT16_REF\x10i\x12\x0f\n\x0b\x44T_INT8_REF\x10j\x12\x11\n\rDT_STRING_REF\x10k\x12\x14\n\x10\x44T_COMPLEX64_REF\x10l\x12\x10\n\x0c\x44T_INT64_REF\x10m\x12\x0f\n\x0b\x44T_BOOL_REF\x10n\x12\x10\n\x0c\x44T_QINT8_REF\x10o\x12\x11\n\rDT_QUINT8_REF\x10p\x12\x11\n\rDT_QINT32_REF\x10q\x12\x13\n\x0f\x44T_BFLOAT16_REF\x10r\x12\x11\n\rDT_QINT16_REF\x10s\x12\x12\n\x0e\x44T_QUINT16_REF\x10t\x12\x11\n\rDT_UINT16_REF\x10u\x12\x15\n\x11\x44T_COMPLEX128_REF\x10v\x12\x0f\n\x0b\x44T_HALF_REF\x10w\x12\x13\n\x0f\x44T_RESOURCE_REF\x10x\x12\x12\n\x0e\x44T_VARIANT_REF\x10y\x12\x11\n\rDT_UINT32_REF\x10z\x12\x11\n\rDT_UINT64_REF\x10{')
+)
+_sym_db.RegisterFileDescriptor(DESCRIPTOR)
+
+_DATATYPE = _descriptor.EnumDescriptor(
+  name='DataType',
+  full_name='tensorflow.DataType',
+  filename=None,
+  file=DESCRIPTOR,
+  values=[
+    _descriptor.EnumValueDescriptor(
+      name='DT_INVALID', index=0, number=0,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_FLOAT', index=1, number=1,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_DOUBLE', index=2, number=2,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT32', index=3, number=3,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT8', index=4, number=4,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT16', index=5, number=5,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT8', index=6, number=6,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_STRING', index=7, number=7,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_COMPLEX64', index=8, number=8,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT64', index=9, number=9,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_BOOL', index=10, number=10,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QINT8', index=11, number=11,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QUINT8', index=12, number=12,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QINT32', index=13, number=13,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_BFLOAT16', index=14, number=14,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QINT16', index=15, number=15,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QUINT16', index=16, number=16,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT16', index=17, number=17,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_COMPLEX128', index=18, number=18,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_HALF', index=19, number=19,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_RESOURCE', index=20, number=20,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_VARIANT', index=21, number=21,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT32', index=22, number=22,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT64', index=23, number=23,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_FLOAT_REF', index=24, number=101,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_DOUBLE_REF', index=25, number=102,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT32_REF', index=26, number=103,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT8_REF', index=27, number=104,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT16_REF', index=28, number=105,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT8_REF', index=29, number=106,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_STRING_REF', index=30, number=107,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_COMPLEX64_REF', index=31, number=108,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_INT64_REF', index=32, number=109,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_BOOL_REF', index=33, number=110,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QINT8_REF', index=34, number=111,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QUINT8_REF', index=35, number=112,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QINT32_REF', index=36, number=113,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_BFLOAT16_REF', index=37, number=114,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QINT16_REF', index=38, number=115,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_QUINT16_REF', index=39, number=116,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT16_REF', index=40, number=117,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_COMPLEX128_REF', index=41, number=118,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_HALF_REF', index=42, number=119,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_RESOURCE_REF', index=43, number=120,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_VARIANT_REF', index=44, number=121,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT32_REF', index=45, number=122,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='DT_UINT64_REF', index=46, number=123,
+      options=None,
+      type=None),
+  ],
+  containing_type=None,
+  options=None,
+  serialized_start=28,
+  serialized_end=838,
+)
+_sym_db.RegisterEnumDescriptor(_DATATYPE)
+
+DataType = enum_type_wrapper.EnumTypeWrapper(_DATATYPE)
+DT_INVALID = 0
+DT_FLOAT = 1
+DT_DOUBLE = 2
+DT_INT32 = 3
+DT_UINT8 = 4
+DT_INT16 = 5
+DT_INT8 = 6
+DT_STRING = 7
+DT_COMPLEX64 = 8
+DT_INT64 = 9
+DT_BOOL = 10
+DT_QINT8 = 11
+DT_QUINT8 = 12
+DT_QINT32 = 13
+DT_BFLOAT16 = 14
+DT_QINT16 = 15
+DT_QUINT16 = 16
+DT_UINT16 = 17
+DT_COMPLEX128 = 18
+DT_HALF = 19
+DT_RESOURCE = 20
+DT_VARIANT = 21
+DT_UINT32 = 22
+DT_UINT64 = 23
+DT_FLOAT_REF = 101
+DT_DOUBLE_REF = 102
+DT_INT32_REF = 103
+DT_UINT8_REF = 104
+DT_INT16_REF = 105
+DT_INT8_REF = 106
+DT_STRING_REF = 107
+DT_COMPLEX64_REF = 108
+DT_INT64_REF = 109
+DT_BOOL_REF = 110
+DT_QINT8_REF = 111
+DT_QUINT8_REF = 112
+DT_QINT32_REF = 113
+DT_BFLOAT16_REF = 114
+DT_QINT16_REF = 115
+DT_QUINT16_REF = 116
+DT_UINT16_REF = 117
+DT_COMPLEX128_REF = 118
+DT_HALF_REF = 119
+DT_RESOURCE_REF = 120
+DT_VARIANT_REF = 121
+DT_UINT32_REF = 122
+DT_UINT64_REF = 123
+
+
+DESCRIPTOR.enum_types_by_name['DataType'] = _DATATYPE
+
+
+# @@protoc_insertion_point(module_scope)

Dragon/python/dragon/vm/tensorflow/training/__init__.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/training/optimizer.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import dragon.updaters as updaters
+import dragon.vm.theano as theano
+import dragon.vm.theano.tensor as T
+
+from dragon.vm.tensorflow.framework import ops
+from dragon.vm.tensorflow.ops import variables
+
+
+class Optimizer(object):
+    def __init__(self, use_locking, name):
+        if not name:
+            raise ValueError('Must specify the optimizer name.')
+        self._use_locking = use_locking
+        self._name = name
+        self._slots = {}
+        self.loss = self.updater = None
+        self.train = self.update = None
+
+    def get_name(self):
+        return self._name
+
+    def minimize(self, loss, global_step=None, var_list=None, **kwargs):
+        grads_and_vars = self.compute_gradients(loss, var_list)
+        return self.apply_gradients(grads_and_vars, global_step=global_step)
+
+    def compute_gradients(self, loss, var_list=None, **kwargs):
+        if var_list is None:
+            var_list = variables.trainable_variables() + \
+                ops.get_collection(ops.GraphKeys.TRAINABLE_RESOURCE_VARIABLES)
+
+        self.loss = loss
+        grads = T.grad(loss, var_list)
+        grads_and_vars = list(zip(grads, var_list))
+        return grads_and_vars
+
+    def apply_gradients(self, grads_and_vars, global_step=None, **kwargs):
+        objs = set()
+        for grad_var in grads_and_vars:
+            self.updater.append((grad_var[1], grad_var[0])) # (var, grad)
+            for obj in grad_var[1].grad_objs: objs.add(obj)
+        self.objs = list(objs)
+        return self
+
+    def run(self, feed_dict=None):
+        # objective function
+        if not hasattr(self, '_objective_func'):
+            # find minimum solving targets
+            targets = set()
+            for t in self.objs: targets.add(t)
+            if feed_dict is not None:
+                self._objective_func = theano.function(inputs=feed_dict.keys(),
+                                                       outputs=list(targets))
+            else:
+                self._objective_func = theano.function(outputs=list(targets))
+        if feed_dict is not None:
+            self._objective_func(*feed_dict.values())
+        else:
+            self._objective_func()
+
+        # update function
+        if not hasattr(self, '_update_func'):
+            self._update_func = theano.function(updater=self.updater)
+        self._update_func()
+
+
+class GradientDescentOptimizer(Optimizer):
+    def __init__(self, learning_rate, use_locking=False, name='GradientDescent'):
+        super(GradientDescentOptimizer, self).__init__(use_locking, name)
+        self.updater = updaters.SGDUpdater(learning_rate, 0.0)
+
+
+class MomentumOptimizer(Optimizer):
+    def __init__(self, learning_rate, momentum,
+                 use_locking=False, name='Momentum', use_nesterov=False):
+        super(MomentumOptimizer, self).__init__(use_locking, name)
+        if not use_nesterov:
+            self.updater = updaters.SGDUpdater(learning_rate, momentum)
+        else:
+            self.updater = updaters.NesterovUpdater(learning_rate, momentum)
+
+
+class AdamOptimizer(Optimizer):
+    def __init__(self, learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-8,
+                 use_locking=False, name='Adam'):
+        super(AdamOptimizer, self).__init__(use_locking, name)
+        self.updater = updaters.AdamUpdater(learning_rate, beta1, beta2, epsilon)
+
+
+class RMSPropOptimizer(Optimizer):
+    def __init__(self, learning_rate, decay, momentum, epsilon=1e-10,
+                 use_locking=False, centered=False, name='RMSProp'):
+        super(RMSPropOptimizer, self).__init__(use_locking, name)
+        self.updater = updaters.RMSPropUpdater(learning_rate, decay, epsilon)

Dragon/python/dragon/vm/tensorflow/training/saver.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+__all__ = ['Saver']
+
+import dragon.core.workspace as ws
+from dragon.core.tensor import Tensor
+
+class Saver(object):
+    def __init__(self,
+                 var_list=None,
+                 max_to_keep=5,
+                 name=None,):
+        self.var_list = var_list
+
+    def save(self,
+             sess,
+             save_path,
+             global_step=None):
+        from ..core.variables import VARIABLES
+        global VARIABLES
+        var_list = VARIABLES if self.var_list is None else self.var_list
+        filename = save_path
+        if global_step is not None:
+            if isinstance(global_step, Tensor):
+                __ndarray__global_step = ws.FetchTensor(global_step)
+                if __ndarray__global_step.size != 1:
+                    raise ValueError('global step must be a scalar of length 1.')
+                filename += '-' + str(__ndarray__global_step.flatten()[0])
+        ws.Snapshot(var_list.values(), filename=filename, suffix='')
+
+    def restore(self, sess, save_path):
+        ws.Restore(save_path)
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/training/train.py

+# --------------------------------------------------------
+# TensorFlow for Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from .optimizer import *
+from .saver import *
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/training/training.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+from dragon.vm.tensorflow.training.optimizer import GradientDescentOptimizer, \
+                                                    MomentumOptimizer, \
+                                                    RMSPropOptimizer, \
+                                                    AdamOptimizer
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/util/deprecation.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright(c) 2017 SeetaTech
+# Written by Ting Pan
+# --------------------------------------------------------
+
+import functools
+import re
+
+_PRINT_DEPRECATION_WARNINGS = True
+
+def _validate_callable(func, decorator_name):
+  if not hasattr(func, '__call__'):
+    raise ValueError(
+        '%s is not a function. If this is a property, make sure'
+        ' @property appears before @%s in your source code:'
+        '\n\n@property\n@%s\ndef method(...)' % (
+            func, decorator_name, decorator_name))
+
+
+def _validate_deprecation_args(date, instructions):
+  if date is not None and not re.match(r'20\d\d-[01]\d-[0123]\d', date):
+    raise ValueError('Date must be YYYY-MM-DD.')
+  if not instructions:
+    raise ValueError('Don\'t deprecate things without conversion instructions!')
+
+
+def _get_qualified_name(function):
+    # Python 3
+    if hasattr(function, '__qualname__'):
+        return function.__qualname__
+
+    # Python 2
+    if hasattr(function, 'im_class'):
+        return function.im_class.__name__ + '.' + function.__name__
+    return function.__name__
+
+
+def deprecated(date, instructions):
+   _validate_deprecation_args(date, instructions)
+
+   def deprecated_wrapper(func):
+       _validate_callable(func, 'deprecated')
+       @functools.wraps(func)
+       def new_func(*args, **kwargs):
+           from dragon.config import logger
+           if _PRINT_DEPRECATION_WARNINGS:
+               logger.warning(
+                   '{} (from {}) is deprecated and will be removed {}.\n'
+                   'Instructions for updating:\n{}'.
+                       format(_get_qualified_name(func),
+                              func.__module__,
+                              'in a future version' if date is None else ('after %s' % date),
+                              instructions))
+               return func(*args, **kwargs)
+       return new_func
+   return deprecated_wrapper
\ No newline at end of file

Dragon/python/dragon/vm/tensorflow/util/nest.py

+# --------------------------------------------------------
+# TensorFlow @ Dragon
+# Copyright 2016 The TensorFlow Authors
+# --------------------------------------------------------
+
+import collections as _collections
+import six as _six
+
+def is_sequence(seq):
+    if isinstance(seq, dict):
+        return True
+    return (isinstance(seq, _collections.Sequence)
+            and not isinstance(seq, _six.string_types))
+
+
+def _yield_value(iterable):
+  if isinstance(iterable, dict):
+    for key in sorted(_six.iterkeys(iterable)):
+      yield iterable[key]
+  else:
+    for value in iterable:
+      yield value
+
+
+def _yield_flat_nest(nest):
+  for n in _yield_value(nest):
+    if is_sequence(n):
+      for ni in _yield_flat_nest(n):
+        yield ni
+    else:
+      yield n
+
+
+def flatten(nest):
+    if is_sequence(nest):
+        return list(_yield_flat_nest(nest))
+    else:
+        return [nest]

Dragon/python/setup.py

@@ -36,7 +36,7 @@ find_packages('dragon')
 find_modules()
 find_modules()
 
 
 setup(name = 'dragon',
 setup(name = 'dragon',
-      version='0.2.1.5',
+      version='0.2.1.6',
       description = 'Dragon: A Computation Graph Virtual Machine Based Deep Learning Framework',
       description = 'Dragon: A Computation Graph Virtual Machine Based Deep Learning Framework',
       url='https://github.com/neopenx/Dragon',
       url='https://github.com/neopenx/Dragon',
       author='Ting Pan',
       author='Ting Pan',