remove default inplace for densenet

3rdparty/lib/README.md

@@ -3,4 +3,4 @@ This directory holds (*after you download them*):
 - libprotobuf.lib (For ``google protobuf``, Windows Only)
 - cudnn.lib (For ``cudnn``, Windows Only)
 - libopenblas.lib (For ``cblas``, Windows Only)
-- python27.lib (For ``python27``, Windows Only)
+- python27.lib/python35.lib/python36.lib (For ``python27/35/36``, Windows Only)

Dragon/CMakeLists.txt

@@ -23,7 +23,7 @@ set(3RDPARTY_DIR  ${PROJECT_SOURCE_DIR}/../3rdparty)
 # set your python environment
 set(PYTHON_DIR /usr/include/python2.7)  # prefer
 #set(PYTHON_DIR /usr/include/python3.x)  # optional, set specific version
-#set(ANACONDA_DIR /xxx/anaconda)  # optional, set specific version below if using py3
+#set(ANACONDA_DIR /xxx/anaconda)  # optional, root folder of anaconda, preset for 2.7, 3.5, and 3.6 
 set(NUMPY_DIR /xxx/numpy)  # require, root folder of numpy package
 
 # set CUDA compiling architecture
@@ -83,7 +83,8 @@ include_directories(${NUMPY_DIR}/core/include)
 include_directories(${NUMPY_DIR})
 include_directories(${NUMPY_DIR}/numpy)
 include_directories(${ANACONDA_DIR}/include/python2.7)
-include_directories(${ANACONDA_DIR}/include/python3.x)
+include_directories(${ANACONDA_DIR}/include/python3.5)
+include_directories(${ANACONDA_DIR}/include/python3.6)
 include_directories(${PYTHON_DIR})
 include_directories(${ANACONDA_DIR}/include)
 

Dragon/include/operators/norm/batch_norm_op.h

@@ -19,7 +19,7 @@ class BatchNormOp : public Operator<Context> {
           momentum(OperatorBase::GetSingleArg<float>("momentum", float(0.9))),
           eps(OperatorBase::GetSingleArg<float>("eps", float(1e-3))),
           use_stats(OperatorBase::GetSingleArg<int>("use_stats", -1)),
-          inplace(OperatorBase::GetSingleArg<bool>("inplace", true)) {}
+          inplace(OperatorBase::GetSingleArg<bool>("inplace", false)) {}
 
     void RunOnDevice() override;
     template <typename T> void RunWithType();

Dragon/include/operators/norm/batch_renorm_op.h

@@ -22,7 +22,7 @@ class BatchRenormOp : public Operator<Context> {
           d_max(OperatorBase::GetSingleArg<float>("d_max", float(5.0))),
           t_delta(OperatorBase::GetSingleArg<float>("t_delta", float(1.0))),
           use_stats(OperatorBase::GetSingleArg<int>("use_stats", -1)),
-          inplace(OperatorBase::GetSingleArg<bool>("inplace", true)),
+          inplace(OperatorBase::GetSingleArg<bool>("inplace", false)),
           t_r_max(float(1.0)), t_d_max(float(0.0)), t_val(float(0.0)) {}
 
     void RunOnDevice() override;

Dragon/include/operators/norm/instance_norm_op.h

@@ -17,7 +17,7 @@ class InstanceNormOp : public Operator<Context> {
     InstanceNormOp(const OperatorDef& op_def, Workspace* ws) 
         : Operator<Context>(op_def, ws),
           eps(OperatorBase::GetSingleArg<float>("eps", float(1e-3))),
-          inplace(OperatorBase::GetSingleArg<bool>("inplace", true)) {}
+          inplace(OperatorBase::GetSingleArg<bool>("inplace", false)) {}
 
     void RunOnDevice() override;
     template <typename T> void RunWithType();

Dragon/modules/python/CMakeLists.txt

@@ -20,6 +20,9 @@ endforeach()
 if (UNIX AND WITH_CUDNN)
 	TARGET_LINK_LIBRARIES(${PROJECT_NAME}_python cudnn)
 endif()
+if (UNIX AND WITH_BLAS)
+	TARGET_LINK_LIBRARIES(${PROJECT_NAME}_python openblas)
+endif()
 
 # ---[ link platforms
 if(UNIX)

Dragon/python/dragon/vm/caffe/layers/common.py

@@ -177,9 +177,9 @@ class BatchNormLayer(Layer):
                        'momentum': param.moving_average_fraction,
                        'eps': param.eps}
         # mean, var, factor are set to 0 in order to do statistics
-        mean = Tensor(LayerParameter.name + '@param0').Constant()
-        var  = Tensor(LayerParameter.name + '@param1').Constant()
-        factor = Tensor(LayerParameter.name + '@param2').Constant()
+        mean = Tensor(LayerParameter.name + '@param0').Constant(value=0.0)
+        var  = Tensor(LayerParameter.name + '@param1').Constant(value=0.0)
+        factor = Tensor(LayerParameter.name + '@param2').Constant(value=0.0)
         # in dragon, set diff as None will ignore computing grad automatically
         # but in bvlc-caffe1, you must set lr_mult = 0 manually
         self._blobs.append({'data': mean, 'diff': None})
@@ -202,9 +202,9 @@ class BatchRenormLayer(Layer):
                        'r_max': float(param.r_max),
                        'd_max': float(param.d_max),
                        't_delta': float(param.t_delta)}
-        mean = Tensor(LayerParameter.name + '@param0').Constant()
-        var  = Tensor(LayerParameter.name + '@param1').Constant()
-        factor = Tensor(LayerParameter.name + '@param2').Constant()
+        mean = Tensor(LayerParameter.name + '@param0').Constant(value=0.0)
+        var  = Tensor(LayerParameter.name + '@param1').Constant(value=0.0)
+        factor = Tensor(LayerParameter.name + '@param2').Constant(value=0.0)
         self._blobs.append({'data': mean, 'diff': None})
         self._blobs.append({'data': var, 'diff': None})
         self._blobs.append({'data': factor, 'diff': None})
@@ -253,19 +253,27 @@ class BNLayer(Layer):
 
     def __init__(self, LayerParameter):
         super(BNLayer, self).__init__(LayerParameter)
-        param = LayerParameter.batch_norm_param
-        self._param = {'use_stats': int(param.use_global_stats)
-                                        if param.HasField('use_global_stats') else -1,
-                       'momentum': param.moving_average_fraction,
-                       'eps': param.eps}
-        mean = Tensor(LayerParameter.name + '@param0').Constant()
-        var = Tensor(LayerParameter.name + '@param1').Constant()
-        scale = Tensor(LayerParameter.name + '@param2').Constant(value=1.0)
-        bias = Tensor(LayerParameter.name + '@param3').Constant(value=0.0)
+        bn_param = LayerParameter.batch_norm_param
+        scale_param = LayerParameter.scale_param
+        self._param = {'use_stats': int(bn_param.use_global_stats)
+                                        if bn_param.HasField('use_global_stats') else -1,
+                       'momentum': bn_param.moving_average_fraction,
+                       'eps': bn_param.eps}
+        mean = Tensor(LayerParameter.name + '@param0').Constant(value=0.0)
+        var = Tensor(LayerParameter.name + '@param1').Constant(value=0.0)
+        scale = Tensor(LayerParameter.name + '@param2')
+        scale_diff = Tensor(LayerParameter.name + '@param2_grad')
+        bias = Tensor(LayerParameter.name + '@param3')
+        bias_diff = Tensor(LayerParameter.name + '@param3_grad')
+
+        if scale_param.HasField('filler'):
+            self.Fill(scale, scale_param, 'filler')
+        else: scale.Constant(value=1.0)
+        self.Fill(bias, scale_param, 'bias_filler')
         self.norm_blobs = [{'data': mean, 'diff': None},
                            {'data': var, 'diff': None}]
-        self.scale_blobs = [{'data': scale, 'diff': Tensor(scale.name + '_grad')},
-                            {'data': bias, 'diff': Tensor(bias.name + '_grad')}]
+        self.scale_blobs = [{'data': scale, 'diff': scale_diff},
+                            {'data': bias, 'diff': bias_diff}]
         self._blobs.extend(self.norm_blobs)
         self._blobs.extend(self.scale_blobs)
 

README.md

@@ -18,10 +18,16 @@
 
 3. (Optional) Download 3rdparty.zip and unzip to Dragon/3rdparty (Out of source code dir)
 
-    [*Win64*](https://pan.baidu.com/s/1pLmGOLt) (OpenBLAS / Protobuf2.6 for VS2013 / CUDNN v6 / Microsoft MPI)
+    [*Win64-VS2013*](https://pan.baidu.com/s/1miGAZl2) (OpenBLAS / Protobuf2.6 for VS2013 / CUDNN v7 / Microsoft MPI)
+
+    [*Win64-VS2015*](https://pan.baidu.com/s/1c2eX6lq) (OpenBLAS / Protobuf2.6 for VS2015 / CUDNN v7 / Microsoft MPI)
 
     [*Linux64*](https://pan.baidu.com/s/1qXPEOWG) (OpenMPI)
 
+    For Windows, ``python27/35/36.lib`` should be copied to ``Dragon/3rdparty/lib``, it depends on the version of Python.
+
+    For Linux, ``libpython-dev``, ``libprotobuf-dev``, ``libopenblas-dev`` and ``cuDNN`` should be installed by yourself.
+
 4. Install Python Requirements
 
     ```Shell
@@ -30,12 +36,12 @@
 	```
 
 5. Configure Dragon/CMakeLists.txt
-	- Select optional libraries [PYTHON3 / CUDA / CUDNN / BLAS / SSE / MPI / MPI_CUDA_AWARE / CUDA_FP16]
+	- Select optional libraries [PYTHON3 / CUDA / CUDNN / BLAS / SSE / MPI]
 	- Set 3rdparty path (recommend to keep defualt)
 	- Set Python include path & Numpy root path
 	- Set CUDA compiling architectures if necessary
 	- GCC version(4.8+, 5.0-) should add ``-std=c++11`` to ``CUDA_NVCC_FLAGS``, if ``nullptr`` is not found
-	- We pre-generated files under the ``Dragon/src/protos`` with protobuf-2.6, run protoc by yourself if higher are required
+	- We pre-generated files under ``Dragon/src/protos`` with protobuf-2.6, run ``protoc`` by yourself if higher are required
 
 6. Environment Variables
     ### Linux(Only for OpenMPI):

examples/cifar10/gen_lmdb.py

@@ -32,12 +32,6 @@ def wrapper_str(raw_str):
 
 def extract_images():
     prefix = 'data/cifar-10-batches-py'
-    extract_path = 'data/extract'
-    if not os.path.exists(os.path.join(extract_path, 'JPEGImages')):
-        os.makedirs(os.path.join(extract_path, 'JPEGImages'))
-    if not os.path.exists(os.path.join(extract_path, 'ImageSets')):
-        os.makedirs(os.path.join(extract_path, 'ImageSets'))
-
     batches = [os.path.join(prefix, 'data_batch_{}'.format(i)) for i in xrange(1, 6)]
     batches += [os.path.join(prefix, 'test_batch')]
 
@@ -60,28 +54,13 @@ def extract_images():
                 label = dict[wrapper_str('labels')][item_idx]
                 im = im.transpose((1, 2, 0))
                 im = im[:, :, ::-1]
-                filename = str(total_idx).zfill(ZFILL) + '.jpg'
-                cv2.imwrite(os.path.join(extract_path, 'JPEGImages', filename), im)
-                images_list.append((filename, str(label)))
+                images_list.append((im, str(label)))
                 total_idx += 1
 
-    # make list
-    with open(os.path.join(extract_path, 'ImageSets', 'train.txt'), 'w') as f:
-        for i in xrange(50000):
-            item = images_list[i][0] + ' ' + images_list[i][1]
-            if i != 49999: item += '\n'
-            f.write(item)
-
-    with open(os.path.join(extract_path, 'ImageSets', 'test.txt'), 'w') as f:
-        for i in xrange(50000, 60000):
-            item = images_list[i][0] + ' ' + images_list[i][1]
-            if i != 59999: item += '\n'
-            f.write(item)
+    return images_list
 
 
-def make_db(image_path, label_path, database_path, pad=0):
-    if os.path.isfile(label_path) is False:
-        raise ValueError('input path is empty or wrong.')
+def make_db(images_list, database_path, pad=0):
     if os.path.isdir(database_path) is True:
         raise ValueError('the database path is already exist.')
 
@@ -90,42 +69,35 @@ def make_db(image_path, label_path, database_path, pad=0):
     db = LMDB(max_commit=10000)
     db.open(database_path, mode='w')
 
-    total_line = sum(1 for line in open(label_path))
+    total_line = len(images_list)
     count = 0
     zfill_flag = '{0:0%d}' % (ZFILL)
 
-    encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), 95]
-
     start_time = time.time()
 
-    with open(label_path, 'r') as input_file:
-        for record in input_file:
-            count += 1
-            if count % 10000 == 0:
-                now_time = time.time()
-                print('{0} / {1} in {2:.2f} sec'.format(
-                    count, total_line, now_time - start_time))
-                db.commit()
-
-            record = record.split()
-            path = record[0]
-            label = record[1]
-
-            img = cv2.imread(os.path.join(image_path ,path))
-            if pad > 0:
-                pad_img = np.zeros((img.shape[0] + 2 * pad,
-                                    img.shape[1] + 2 * pad, 3), dtype=np.uint8)
-                pad_img[pad : pad + img.shape[0],
+    for record in images_list:
+        count += 1
+        if count % 10000 == 0:
+            now_time = time.time()
+            print('{0} / {1} in {2:.2f} sec'.format(
+                count, total_line, now_time - start_time))
+            db.commit()
+
+        img = record[0]
+        label = record[1]
+        if pad > 0:
+            pad_img = np.zeros((img.shape[0] + 2 * pad,
+                                img.shape[1] + 2 * pad, 3), dtype=np.uint8)
+            pad_img[pad : pad + img.shape[0],
                         pad : pad + img.shape[1], :] = img
-                img = pad_img
-            result, imgencode = cv2.imencode('.jpg', img, encode_param)
+            img = pad_img
 
-            datum = caffe_pb2.Datum()
-            datum.height, datum.width, datum.channels = img.shape
-            datum.label = int(label)
-            datum.encoded = True
-            datum.data = imgencode.tostring()
-            db.put(zfill_flag.format(count - 1), datum.SerializeToString())
+        datum = caffe_pb2.Datum()
+        datum.height, datum.width, datum.channels = img.shape
+        datum.label = int(label)
+        datum.encoded = False
+        datum.data = img.tostring()
+        db.put(zfill_flag.format(count - 1), datum.SerializeToString())
 
     now_time = time.time()
     print('{0} / {1} in {2:.2f} sec'.format(count, total_line, now_time - start_time))
@@ -134,7 +106,6 @@ def make_db(image_path, label_path, database_path, pad=0):
     db.commit()
     db.close()
 
-    shutil.copy(label_path, database_path + '/image_list.txt')
     end_time = time.time()
     print('{0} images have been stored in the database.'.format(total_line))
     print('This task finishes within {0:.2f} seconds.'.format(
@@ -147,12 +118,8 @@ if __name__ == '__main__':
 
     untar('data/cifar-10-python.tar.gz')
 
-    extract_images()
+    images_list = extract_images()
 
-    make_db('data/extract/JPEGImages',
-            'data/extract/ImageSets/train.txt',
-            'data/train_lmdb')
+    make_db(images_list[0:50000], 'data/train_lmdb')
 
-    make_db('data/extract/JPEGImages',
-            'data/extract/ImageSets/test.txt',
-            'data/test_lmdb')
+    make_db(images_list[50000:60000], 'data/test_lmdb')