Skip to content

SeetaResearch / Dragon

Go to a project
Commit 904b59fd by Ting PAN
Options

Add Contrib ops

1 parent 36f27485
Expand all
Inline Side-by-side
Showing with 795 additions and 146 deletions
Dragon/include/operators/misc/proposal_op.h deleted 100644 → 0
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_MISC_PROPOSAL_OP_H_
#define DRAGON_OPERATORS_MISC_PROPOSAL_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class ProposalOp final : public Operator<Context> {
public:
ProposalOp(const OperatorDef& op_def, Workspace* ws)
: base_size_(OperatorBase::GetSingleArg<int>("base_size", 16)),
min_size_(OperatorBase::GetSingleArg<int>("min_size", 16)),
feat_stride_(OperatorBase::GetSingleArg<int>("feat_stride", -1)),
pre_nms_topn_(OperatorBase::GetSingleArg<int>("pre_nms_topn", 12000)),
post_nms_topn_(OperatorBase::GetSingleArg<int>("post_nms_topn", 2000)),
nms_thresh_(OperatorBase::GetSingleArg<float>("nms_thresh", (float)0.7)),
Operator<Context>(op_def, ws) { Setup(); }
void Setup();
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
int min_size_, base_size_, feat_stride_;
int pre_nms_topn_, post_nms_topn_;
float nms_thresh_;
Tensor anchors_, roi_indices_, proposals_, nms_mask_;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_MISC_COMPARE_OP_H_
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators.rst
......@@ -64,6 +64,7 @@ Custom
operators/custom/data_process
operators/custom/vec_mult
========================================= =====================================================================
List Brief
========================================= =====================================================================
......@@ -73,6 +74,19 @@ List Brief
========================================= =====================================================================
Contrib
-------
.. toctree::
:hidden:
operators/contrib/rcnn
========================================= =====================================================================
List Brief
========================================= =====================================================================
`dragon.operators.contrib.rcnn`_ Contrib ops for R-CNN.
========================================= =====================================================================
.. _dragon.operators.data: operators/data.html
......@@ -91,4 +105,6 @@ List Brief
.. _dragon.io: io.html
.. _dragon.operators.custom.minibatch: operators/custom/minibatch.html
.. _dragon.operators.custom.data_process: operators/custom/data_process.html
.. _dragon.operators.custom.vec_mult: operators/custom/vec_mult.html
\ No newline at end of file
.. _dragon.operators.custom.vec_mult: operators/custom/vec_mult.html
.. _dragon.operators.contrib.rcnn: operators/contrib/rcnn.html
Dragon/python/dragon/docs/contents/operators/contrib/rcnn.rst 0 → 100644
============
:mod:`R-CNN`
============
.. toctree::
:hidden:
.. automodule:: dragon.operators.contrib.rcnn.ops
:members:
\ No newline at end of file
Dragon/python/dragon/docs/contents/ops.rst
......@@ -35,8 +35,8 @@ List Brief
`Conv2d`_ 2d Convolution.
`Conv2dTranspose`_ 2d Deconvolution.
`Pool2d`_ 2d Pooling, MAX or AVG.
`ROIPooling`_ ROIPoolin(MAX), introduced by `[Girshick, 2015] <https://arxiv.org/abs/1504.08083>`_.
`ROIAlign`_ ROIAlign(MAX), introduced by `[He et.al, 2017] <https://arxiv.org/abs/1703.06870>`_.
`ROIPooling`_ ROIPooling(MAX), introduced by `[Girshick, 2015] <https://arxiv.org/abs/1504.08083>`_.
`ROIAlign`_ ROIAlign(AVG), introduced by `[He et.al, 2017] <https://arxiv.org/abs/1703.06870>`_.
`LRN`_ Local Response Normalization, introduced by `[Krizhevsky et.al, 2012] <http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks>`_.
`NNResize`_ Resize the image with Nearest-Neighbor method.
`BilinearResize`_ Resize the image with Bi-linear method.
......@@ -167,6 +167,14 @@ List Brief
`Proposal`_ Generate Regional Proposals, introduced by `[Ren et.al, 2015] <https://arxiv.org/abs/1506.01497>`_.
================= ======================================================================
Contrib
-------
================= ======================================================================
List Brief
================= ======================================================================
`Proposal`_ Generate Regional Proposals, introduced by `[Ren et.al, 2015] <https://arxiv.org/abs/1506.01497>`_.
================= ======================================================================
Cast
----
================= ======================================================================
......@@ -279,7 +287,8 @@ List Brief
.. _Accuracy: operators/misc.html#dragon.operators.misc.Accuracy
.. _StopGradient: operators/misc.html#dragon.operators.misc.StopGradient
.. _MovingAverage: operators/misc.html#dragon.operators.misc.MovingAverage
.. _Proposal: operators/misc.html#dragon.operators.misc.Proposal
.. _Proposal: operators/contrib/rcnn.html#dragon.operators.contrib.rcnn.ops.Proposal
.. _FloatToHalf: operators/cast.html#dragon.operators.misc.FloatToHalf
......
Dragon/python/dragon/docs/contents/vm/caffe/layer.rst
......@@ -282,6 +282,8 @@ API Reference
.. _NormalizeParameter.scale_filler: https://github.com/weiliu89/caffe/blob/f5eac041aafbc8b86954bd161710f65e70042ce6/src/caffe/proto/caffe.proto#L1332
.. _NormalizeParameter.channel_shared: https://github.com/weiliu89/caffe/blob/f5eac041aafbc8b86954bd161710f65e70042ce6/src/caffe/proto/caffe.proto#L1334
.. _NormalizeParameter.eps: https://github.com/weiliu89/caffe/blob/f5eac041aafbc8b86954bd161710f65e70042ce6/src/caffe/proto/caffe.proto#L1336
.. _ReductionParameter.operation: https://github.com/BVLC/caffe/blob/effcdb0b62410b2a6a54f18f23cf90733a115673/src/caffe/proto/caffe.proto#L973
.. _ReductionParameter.axis: https://github.com/BVLC/caffe/blob/effcdb0b62410b2a6a54f18f23cf90733a115673/src/caffe/proto/caffe.proto#L988
.. _TileParameter.multiples: https://github.com/neopenx/Dragon/blob/6eeac5fec58ed3d0d79f0b4003471e4a641c72f4/Dragon/python/dragon/vm/caffe/proto/caffe.proto#L1173
.. _ExpandDimsParameter.axis: https://github.com/neopenx/Dragon/blob/6eeac5fec58ed3d0d79f0b4003471e4a641c72f4/Dragon/python/dragon/vm/caffe/proto/caffe.proto#L1480
.. _ProposalParameter.feat_stride: https://github.com/sanghoon/caffe/blob/6068dd04ea93cca9fcee036628fdb3ea95b4ebcd/src/caffe/proto/caffe.proto#L431
......
Dragon/python/dragon/operators/contrib/__init__.py 0 → 100644
# --------------------------------------------------------
# Dragon
# Copyright(c) 2017 SeetaTech
# Written by Ting Pan
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
# R-CNN ops
from dragon.operators.contrib.rcnn.ops import *
Dragon/python/dragon/operators/contrib/rcnn/__init__.py 0 → 100644
# --------------------------------------------------------
# Dragon
# Copyright(c) 2017 SeetaTech
# Written by Ting Pan
# --------------------------------------------------------
\ No newline at end of file
Dragon/python/dragon/operators/contrib/rcnn/ops.py 0 → 100644
# --------------------------------------------------------
# Dragon
# Copyright(c) 2017 SeetaTech
# Written by Ting Pan
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from dragon.operators import *
def Proposal(inputs, strides, ratios, scales,
pre_nms_top_n=6000, post_nms_top_n=300,
nms_thresh=0.7, min_size=16,
min_level=2, max_level=5,
canonical_scale=224, canonical_level=4, **kwargs):
"""Generate Regional Proposals, introduced by `[Ren et.al, 2015] <https://arxiv.org/abs/1506.01497>`_.
Multi-Level proposals was introduced by `[Lin et.al, 2017] <https://arxiv.org/abs/1612.03144>`_.
For single level proposals(e.g. C4), the inputs should be: [cls_probs, bbox_deltas, im_info].
For multiple level proposals(e.g. FPN), the inputs should be: [cls_score/Px, ...] + [cls_probs, bbox_deltas, im_info].
Parameters
----------
inputs : list of Tensor
The inputs.
strides : list of int
The strides of anchors.
ratios : list of float
The ratios of anchors.
scales : list of float
The scales of anchors.
pre_nms_top_n : int
The number of anchors before nms.
post_nms_top_n : int
The number of anchors after nms.
nms_thresh : float
The threshold of nms.
min_size : int
The min size of anchors.
min_level : int
Finest level of the FPN pyramid.
max_level : int
Coarsest level of the FPN pyramid.
canonical_scale : int
The baseline scale of mapping policy.
canonical_level : int
Heuristic level of the canonical scale.
Returns
-------
Tensor
The proposals.
"""
CheckInputs(inputs, 3, INT_MAX)
arguments = ParseArguments(locals())
num_levels = (max_level - min_level) + 1
num_levels = 1 if len(inputs) == 3 else num_levels
outputs = Tensor.CreateOperator(nout=num_levels, op_type='Proposal', **arguments)
return outputs
\ No newline at end of file
Dragon/python/dragon/operators/misc.py
......@@ -157,44 +157,4 @@ def MovingAverage(inputs, decay, **kwargs):
output = Tensor.CreateOperator(op_type='MovingAverage',
existing_outputs=variable, **arguments)
return output
def Proposal(inputs, ratios, scales,
base_size=16, min_size=16, feat_stride=16,
pre_nms_topn=12000, post_nms_topn=2000, nms_thresh=0.7, **kwargs):
"""Generate Regional Proposals, introduced by `[Ren et.al, 2015] <https://arxiv.org/abs/1506.01497>`_.
Parameters
----------
inputs : list of Tensor
The inputs, represent [input, anchors, im_info].
ratios : list of float
The ratios of anchors.
scales : list of float
The scales of anchors.
base_size : int
The base size of anchors.
min_size : int
The min size of anchors.
feat_stride : int
The stride of input. Default is ``16`` (The 4th down-samples).
pre_nms_topn : int
The number of anchors before nms.
post_nms_topn : int
The number of anchors after nms.
nms_thresh : float
The threshold of nms.
Returns
=------
Tensor
The proposals.
"""
CheckInputs(inputs, 3)
arguments = ParseArguments(locals())
output = Tensor.CreateOperator(nout=1, op_type='Proposal', **arguments)
return output
\ No newline at end of file
Dragon/python/dragon/operators/ndarray.py
......@@ -272,9 +272,14 @@ def Reduce(inputs, axis=-1, operation='NONE', keep_dims=False, **kwargs):
output = Tensor.CreateOperator(nout=1, op_type='Reduce', **arguments)
if inputs.shape is not None:
if axis == -1: output.shape = [1]
output.shape = inputs.shape[:]
if axis == -1:
if keep_dims:
for i in xrange(len(output.shape)):
output.shape[i] = 1
else: output.shape = [1]
else:
output.shape = inputs.shape[:]
if keep_dims: output.shape[axis] = 1
else: del output.shape[axis]
......
Dragon/python/dragon/operators/vision.py
......@@ -329,7 +329,7 @@ def ROIPooling(inputs, pool_h, pool_w, spatial_scale, **kwargs):
def ROIAlign(inputs, pool_h=0, pool_w=0, spatial_scale=1.0, sampling_ratio=2, **kwargs):
"""Max ROIAlign, introduced by `[He et.al, 2017] <https://arxiv.org/abs/1703.06870>`_.
"""AVG ROIAlign, introduced by `[He et.al, 2017] <https://arxiv.org/abs/1703.06870>`_.
The first dimension of input must be ``1``.
......
Dragon/python/dragon/ops.py
......@@ -21,6 +21,7 @@ from .operators import mpi
from .operators import ndarray
from .operators import norm
from .operators import recurrent
from .operators import contrib
# data
LMDBData = data.LMDBData
......@@ -128,11 +129,13 @@ Template = misc.Template
Accuracy = misc.Accuracy
StopGradient = misc.StopGradient
MovingAverage = misc.MovingAverage
Proposal = misc.Proposal
# cast
FloatToHalf = cast.FloatToHalf
# mpi
MPIBroadcast = mpi.MPIBroadcast
MPIGather = mpi.MPIGather
\ No newline at end of file
MPIGather = mpi.MPIGather
# contrib
Proposal = contrib.Proposal # R-CNN
\ No newline at end of file
Dragon/python/dragon/vm/caffe/layers/common.py
......@@ -272,7 +272,7 @@ class GatherLayer(Layer):
Parameters
----------
axis : int
The axis for gathering. Refer `GatherParameter.axis`_.
The axis for gathering. Refer ``GatherParameter.axis``.
"""
def __init__(self, LayerParameter):
......@@ -623,35 +623,44 @@ class ProposalLayer(Layer):
Parameters
----------
feat_stride : int
The stride of input. Refer `ProposalParameter.feat_stride`_.
base_size : int
The base size of anchors. Refer `ProposalParameter.base_size`_.
min_size : int
The min size of anchors. Refer `ProposalParameter.min_size`_.
ratio : list of float
The ratios of anchors. Refer `ProposalParameter.ratio`_.
stride : list of int
The stride of anchors. Refer ``ProposalParameter.stride``.
scale : list of float
The scales of anchors. Refer `ProposalParameter.scale`_.
pre_nms_topn : int
The num of anchors before nms. Refer `ProposalParameter.pre_nms_topn`_.
post_nms_topn : int
ratio : list of float
The ratios of anchors. Refer `ProposalParameter.ratio`_.
pre_nms_top_n : int
The num of anchors before nms. Refer `ProposalParameter.pre_nms_topn`_.
post_nms_top_n : int
The num of anchors after nms. Refer `ProposalParameter.post_nms_topn`_.
nms_thresh : float
The threshold of nms. Refer `ProposalParameter.nms_thresh`_.
min_size : int
The min size of anchors. Refer `ProposalParameter.min_size`_.
min_level : int
Finest level of the FPN pyramid. Refer ``ProposalParameter.min_level``.
max_level : int
Coarsest level of the FPN pyramid. Refer ``ProposalParameter.max_level``.
canonical_scale : int
The baseline scale of mapping policy. Refer ``ProposalParameter.canonical_scale``.
canonical_level : int
Heuristic level of the canonical scale. Refer ``ProposalParameter.canonical_level``.
"""
def __init__(self, LayerParameter):
super(ProposalLayer, self).__init__(LayerParameter)
param = LayerParameter.proposal_param
self._param = {'base_size': param.base_size,
'min_size': param.min_size,
'feat_stride': param.feat_stride,
'pre_nms_topn': param.pre_nms_topn,
'post_nms_topn': param.post_nms_topn,
'nms_thresh': param.nms_thresh,
self._param = {'strides': param.stride,
'ratios': param.ratio,
'scales': param.scale}
'scales': param.scale,
'pre_nms_top_n': param.pre_nms_top_n,
'post_nms_top_n': param.post_nms_top_n,
'nms_thresh': param.nms_thresh,
'min_size': param.min_size,
'min_level': param.min_level,
'max_level': param.max_level,
'canonical_scale': param.canonical_scale,
'canonical_level': param.canonical_level}
def Setup(self, bottom):
super(ProposalLayer, self).Setup(bottom)
......
Dragon/python/dragon/vm/caffe/proto/caffe.proto
......@@ -1474,14 +1474,17 @@ message ExpandDimsParameter {
}
message ProposalParameter {
optional uint32 feat_stride = 1 [default = 16];
optional uint32 base_size = 2 [default = 16];
optional uint32 min_size = 3 [default = 16];
repeated float ratio = 4;
repeated float scale = 5;
optional uint32 pre_nms_topn = 6 [default = 6000];
optional uint32 post_nms_topn = 7 [default = 300];
optional float nms_thresh = 8 [default = 0.7];
repeated int32 stride = 1;
repeated float ratio = 2;
repeated float scale = 3;
optional uint32 pre_nms_top_n = 4 [default = 6000];
optional uint32 post_nms_top_n = 5 [default = 300];
optional float nms_thresh = 6 [default = 0.7];
optional uint32 min_size = 7 [default = 16];
optional int32 min_level = 8 [default = 2];
optional int32 max_level = 9 [default = 5];
optional int32 canonical_scale = 10 [default = 224];
optional int32 canonical_level = 11 [default = 4];
}
message BatchRenormParameter {
......
Dragon/python/setup.py
......@@ -36,7 +36,7 @@ find_packages('dragon')
find_modules()
setup(name = 'dragon',
version='0.2.1.3',
version='0.2.1.4',
description = 'Dragon: A Computation Graph Virtual Machine Based Deep Learning Framework',
url='https://github.com/neopenx/Dragon',
author='Ting Pan',
......
Dragon/src/contrib/rcnn/bbox_utils.cc 0 → 100644
#include "core/context.h"
#include "contrib/rcnn/bbox_utils.h"
namespace dragon {
namespace rcnn {
/******************** Proposal ********************/
template <> void GenerateProposals<float, CPUContext>(const int A,
const int feat_h,
const int feat_w,
const int stride,
const float im_h, const float im_w,
const float min_box_h, const float min_box_w,
const float* scores,
const float* bbox_deltas,
const float* anchors,
float* proposals) {
float* proposal = proposals;
const int K = feat_h * feat_w;
for (int h = 0; h < feat_h; ++h) {
for (int w = 0; w < feat_w; ++w) {
const float x = w * stride;
const float y = h * stride;
// bbox_deltas: [1, A, 4, K]
const float* bbox_delta = bbox_deltas + h * feat_w + w;
// scores: [1, A, K]
const float* score = scores + h * feat_w + w;
for (int a = 0; a < A; ++a) {
const float dx = bbox_delta[(a * 4 + 0) * K];
const float dy = bbox_delta[(a * 4 + 1) * K];
const float d_log_w = bbox_delta[(a * 4 + 2) * K];
const float d_log_h = bbox_delta[(a * 4 + 3) * K];
proposal[0] = x + anchors[a * 4 + 0];
proposal[1] = y + anchors[a * 4 + 1];
proposal[2] = x + anchors[a * 4 + 2];
proposal[3] = y + anchors[a * 4 + 3];
proposal[4] = BBoxTransform<float>(dx, dy,
d_log_w, d_log_h,
im_w, im_h,
min_box_w, min_box_h,
proposal) * score[a * K];
proposal += 5;
}
}
}
}
template <> void GenerateProposals_v2<float, CPUContext>(const int total_anchors,
const float im_h, const float im_w,
const float min_box_h, const float min_box_w,
const float* scores,
const float* bbox_deltas,
float* proposals) {
float* proposal = proposals;
for (int i = 0; i < total_anchors; ++i) {
// bbox_deltas: [1, 4, total_anchors]
// scores: [1, total_anchors]
const float dx = bbox_deltas[i];
const float dy = bbox_deltas[total_anchors + i];
const float d_log_w = bbox_deltas[2 * total_anchors + i];
const float d_log_h = bbox_deltas[3 * total_anchors + i];
proposal[4] = BBoxTransform<float>(dx, dy,
d_log_w, d_log_h,
im_w, im_h,
min_box_w, min_box_h,
proposal) * scores[i];
proposal += 5;
}
}
/******************** NMS ********************/
template <typename T>
T iou(const T A[], const T B[]) {
if (A[0] > B[2] || A[1] > B[3] || A[2] < B[0] || A[3] < B[1]) return 0;
const T x1 = std::max(A[0], B[0]);
const T y1 = std::max(A[1], B[1]);
const T x2 = std::min(A[2], B[2]);
const T y2 = std::min(A[3], B[3]);
const T width = std::max((T)0, x2 - x1 + (T)1);
const T height = std::max((T)0, y2 - y1 + (T)1);
const T area = width * height;
const T A_area = (A[2] - A[0] + (T)1) * (A[3] - A[1] + (T)1);
const T B_area = (B[2] - B[0] + (T)1) * (B[3] - B[1] + (T)1);
return area / (A_area + B_area - area);
}
template <> void NMS<float, CPUContext>(const int num_boxes,
const int max_keeps,
const float thresh,
const float* proposals,
int* roi_indices,
int& num_rois,
Tensor* mask) {
int count = 0;
std::vector<char> is_dead(num_boxes);
for (int i = 0; i < num_boxes; ++i) is_dead[i] = 0;
for (int i = 0; i < num_boxes; ++i) {
if (is_dead[i]) continue;
roi_indices[count++] = i;
if (count == max_keeps) break;
for (int j = i + 1; j < num_boxes; ++j)
if (!is_dead[j] && iou(&proposals[i * 5], &proposals[j * 5]) > thresh) is_dead[j] = 1;
}
num_rois = count;
}
} // namespace rcnn
} // namespace dragon
\ No newline at end of file
Dragon/src/contrib/rcnn/bbox_utils.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_CONTRIB_RCNN_BBOX_UTILS_H_
#define DRAGON_CONTRIB_RCNN_BBOX_UTILS_H_
#include "core/context.h"
#include "core/operator.h"
namespace dragon {
namespace rcnn {
#define ROUND(x) ((int)((x) + (T)0.5))
/******************** BBox ********************/
template <typename T>
int BBoxTransform(const T dx, const T dy,
const T d_log_w, const T d_log_h,
const T im_w, const T im_h,
const T min_box_w, const T min_box_h,
T* bbox) {
const T w = bbox[2] - bbox[0] + (T)1;
const T h = bbox[3] - bbox[1] + (T)1;
const T ctr_x = bbox[0] + (T)0.5 * w;
const T ctr_y = bbox[1] + (T)0.5 * h;
const T pred_ctr_x = dx * w + ctr_x;
const T pred_ctr_y = dy * h + ctr_y;
const T pred_w = exp(d_log_w) * w;
const T pred_h = exp(d_log_h) * h;
bbox[0] = pred_ctr_x - (T)0.5 * pred_w;
bbox[1] = pred_ctr_y - (T)0.5 * pred_h;
bbox[2] = pred_ctr_x + (T)0.5 * pred_w;
bbox[3] = pred_ctr_y + (T)0.5 * pred_h;
bbox[0] = std::max((T)0, std::min(bbox[0], im_w - (T)1));
bbox[1] = std::max((T)0, std::min(bbox[1], im_h - (T)1));
bbox[2] = std::max((T)0, std::min(bbox[2], im_w - (T)1));
bbox[3] = std::max((T)0, std::min(bbox[3], im_h - (T)1));
const T bbox_w = bbox[2] - bbox[0] + (T)1;
const T bbox_h = bbox[3] - bbox[1] + (T)1;
return (bbox_w >= min_box_w) * (bbox_h >= min_box_h);
}
/******************** Anchor ********************/
template <typename T>
void GenerateAnchors(int base_size,
const int num_ratios,
const int num_scales,
const T* ratios,
const T* scales,
T* anchors) {
const T base_area = (T)(base_size * base_size);
const T center = (T)0.5 * (base_size - (T)1);
T* offset_anchors = anchors;
for (int i = 0; i < num_ratios; ++i) {
const T ratio_w = (T)ROUND(sqrt(base_area / ratios[i]));
const T ratio_h = (T)ROUND(ratio_w * ratios[i]);
for (int j = 0; j < num_scales; ++j) {
const T scale_w = (T)0.5 * (ratio_w * scales[j] - (T)1);
const T scale_h = (T)0.5 * (ratio_h * scales[j] - (T)1);
offset_anchors[0] = center - scale_w;
offset_anchors[1] = center - scale_h;
offset_anchors[2] = center + scale_w;
offset_anchors[3] = center + scale_h;
offset_anchors += 4;
}
}
}
template <typename T>
void GenerateGridAnchors(const int A, const int feat_h, const int feat_w,
const int stride,
const T* anchors,
T* proposals) {
T* proposal = proposals;
for (int a = 0; a < A; ++a) {
for (int h = 0; h < feat_h; ++h) {
for (int w = 0; w < feat_w; ++w) {
const T x = w * stride;
const T y = h * stride;
proposal[0] = x + anchors[a * 4 + 0];
proposal[1] = y + anchors[a * 4 + 1];
proposal[2] = x + anchors[a * 4 + 2];
proposal[3] = y + anchors[a * 4 + 3];
proposal += 5;
}
}
}
}
/******************** Proposal ********************/
template <typename T, class Context>
void GenerateProposals(const int A, const int feat_h, const int feat_w,
const int stride,
const float im_h, const float im_w,
const float min_box_h, const float min_box_w,
const T* scores,
const T* bbox_deltas,
const T* anchors,
T* proposals);
template <typename T, class Context>
void GenerateProposals_v2(const int total_anchors,
const float im_h, const float im_w,
const float min_box_h, const float min_box_w,
const T* scores,
const T* bbox_deltas,
T* proposals);
template <typename T>
void SortProposals(const int start,
const int end,
const int num_top,
T* proposals) {
const T pivot_score = proposals[start * 5 + 4];
int left = start + 1, right = end;
while (left <= right) {
while (left <= end && proposals[left * 5 + 4] >= pivot_score) ++left;
while (right > start && proposals[right * 5 + 4] <= pivot_score) --right;
if (left <= right) {
for (int i = 0; i < 5; ++i)
std::swap(proposals[left * 5 + i], proposals[right * 5 + i]);
++left;
--right;
}
}
if (right > start) {
for (int i = 0; i < 5; ++i)
std::swap(proposals[start * 5 + i], proposals[right * 5 + i]);
}
if (start < right - 1) SortProposals(start, right - 1, num_top, proposals);
if (right + 1 < num_top && right + 1 < end) SortProposals(right + 1, end, num_top, proposals);
}
template <typename T>
void RetrieveRoIs(const int num_rois,
const int roi_batch_ind,
const T* proposals,
const int* roi_indices,
T* rois) {
for (int i = 0; i < num_rois; ++i) {
const T* proposal = proposals + roi_indices[i] * 5;
rois[i * 5 + 0] = roi_batch_ind;
rois[i * 5 + 1] = proposal[0];
rois[i * 5 + 2] = proposal[1];
rois[i * 5 + 3] = proposal[2];
rois[i * 5 + 4] = proposal[3];
}
}
template <typename T>
int roi_level(const int min_level, // e.g. 2
const int max_level, // e.g. 5
const int canonical_level, // e.g. 4
const int canonical_scale, // e.g. 224
T* roi) {
T w = roi[3] - roi[1] + 1;
T h = roi[4] - roi[2] + 1;
// reference the settings of paper
int level = canonical_level + std::log(std::max(std::sqrt(w * h), T(1)) / T(canonical_scale));
return std::min(max_level, std::max(min_level, level));
}
template <typename T>
void CollectRoIs(const int num_rois,
const int min_level,
const int max_level,
const int canonical_level,
const int canonical_scale,
const T* rois,
vector< vector<TIndex> >& roi_bins) {
const T* roi = rois;
for (int i = 0; i < num_rois; ++i) {
int bin_idx = roi_level(min_level, max_level,
canonical_level, canonical_scale,
roi);
bin_idx = std::max(bin_idx - min_level, 0);
roi_bins[bin_idx].push_back(i);
roi += 5;
}
}
template <typename T>
void DistributeRoIs(const vector< vector<TIndex> >& roi_bins,
const T* rois,
vector<T*> outputs) {
for (int i = 0; i < roi_bins.size(); i++) {
auto* y = outputs[i];
if (roi_bins[i].size() == 0) {
// fake a tiny roi to avoid empty roi pooling
y[0] = 0, y[1] = 0, y[2] = 0, y[3] = 1, y[4] = 1;
} else {
for (int j = 0; j < roi_bins[i].size(); ++j) {
const T* roi = rois + roi_bins[i][j] * 5;
for (int k = 0; k < 5; ++k) y[k] = roi[k];
y += 5;
}
}
}
}
/******************** NMS ********************/
template <typename T, class Context>
void NMS(const int num_boxes,
const int max_keeps,
const T thresh,
const T* proposals,
int* roi_indices,
int& num_rois,
Tensor* mask);
} // namespace rcnn
} // namespace dragon
#endif // DRAGON_OPERATORS_CONTRIB_BBOX_UTILS_H_
\ No newline at end of file
Dragon/src/contrib/rcnn/proposal_op.cc 0 → 100644
#include "contrib/rcnn/proposal_op.h"
#include "contrib/rcnn/bbox_utils.h"
namespace dragon {
template <class Context> template <typename T>
void ProposalOp<Context>::RunWithType() {
TIndex total_rois = 0;
auto* im_info = input(-1).template data<T, CPUContext>();
auto* Ydata = output(0)->template mutable_data<T, CPUContext>();
for (int n = 0; n < num_images; ++n) {
const T im_height = im_info[0];
const T im_width = im_info[1];
const T scale = im_info[2];
const T min_box_h = min_size * scale;
const T min_box_w = min_size * scale;
int num_rois = 0;
if (strides.size() == 1) {
// case 1: single stride (Faster R-CNN)
const TIndex feat_height = input(0).dim(2);
const TIndex feat_width = input(0).dim(3);
const TIndex K = feat_height * feat_width;
const TIndex A = ratios.size() * scales.size();
const TIndex num_proposals = K * A;
const TIndex pre_nms_topn = std::min(num_proposals, pre_nms_top_n);
anchors_.Reshape(vector<TIndex>({ A, 4 }));
proposals_.Reshape(vector<TIndex>({ num_proposals, 5 }));
rcnn::GenerateAnchors<T>(strides[0], (int)ratios.size(), (int)scales.size(),
&ratios[0], &scales[0],
anchors_.template mutable_data<T, CPUContext>());
rcnn::GenerateProposals<T, Context>(A, feat_height, feat_width, strides[0],
im_height, im_width, min_box_h, min_box_w,
input(0).template data<T, Context>() + num_proposals,
input(1).template data<T, Context>(),
anchors_.template mutable_data<T, Context>(),
proposals_.template mutable_data<T, Context>());
rcnn::SortProposals(0, num_proposals - 1, pre_nms_top_n,
proposals_.template mutable_data<T, CPUContext>());
rcnn::NMS<T, Context>(pre_nms_topn, post_nms_top_n, nms_thresh,
proposals_.template mutable_data<T, Context>(),
roi_indices_.template mutable_data<int, CPUContext>(),
num_rois,
&nms_mask_);
rcnn::RetrieveRoIs<T>(num_rois, n, proposals_.template mutable_data<T, CPUContext>(),
roi_indices_.template mutable_data<int, CPUContext>(),
Ydata);
} else if (strides.size() > 1) {
// case 2: multiple stride (FPN / Mask R-CNN / RetinaNet)
CHECK_EQ(strides.size(), (int)InputSize() - 3)
<< "\nGiven " << strides.size() << " strides and "
<< InputSize() - 3 << " feature inputs";
CHECK_EQ(strides.size(), scales.size())
<< "\nGiven " << strides.size() << " strides and "
<< scales.size() << " scales";
// cls_probs: [1, 2, total_proposals]
// bbox_deltas: [1, 4, total_proposals]
TIndex total_proposals = input(-3).dim(2), acc_proposals = 0;
const TIndex pre_nms_topn = std::min(total_proposals, pre_nms_top_n);;
proposals_.Reshape(vector<TIndex>({ total_proposals, 5 }));
auto* proposals = proposals_.template mutable_data<T, CPUContext>();
for (int i = 0; i < strides.size(); i++) {
const TIndex feat_height = input(i).dim(2);
const TIndex feat_width = input(i).dim(3);
const TIndex K = feat_height * feat_width;
const TIndex A = ratios.size();
const TIndex num_proposals = K * A;
anchors_.Reshape(vector<TIndex>({ A, 4 }));
rcnn::GenerateAnchors<T>(strides[i], (int)ratios.size(), 1,
&ratios[0], &scales[0],
anchors_.template mutable_data<T, CPUContext>());
rcnn::GenerateGridAnchors<T>(A, feat_height, feat_width, strides[i],
anchors_.template mutable_data<T, CPUContext>(),
proposals);
acc_proposals += num_proposals;
proposals += (num_proposals * 5);
}
CHECK_EQ(acc_proposals, total_proposals)
<< "\nExcepted " << total_proposals << " proposals from the network, "
<< "but generated " << acc_proposals << " proposals.";
rcnn::GenerateProposals_v2<T, Context>(total_proposals, im_height, im_width,
min_box_h, min_box_w,
input(-3).template data<T, Context>() + total_proposals,
input(-2).template data<T, Context>(),
proposals_.template mutable_data<T, Context>());
rcnn::SortProposals(0, total_proposals - 1, pre_nms_top_n,
proposals_.template mutable_data<T, CPUContext>());
rcnn::NMS<T, Context>(pre_nms_topn, post_nms_top_n, nms_thresh,
proposals_.template mutable_data<T, Context>(),
roi_indices_.template mutable_data<int, CPUContext>(),
num_rois,
&nms_mask_);
rcnn::RetrieveRoIs<T>(num_rois, n, proposals_.template mutable_data<T, CPUContext>(),
roi_indices_.template mutable_data<int, CPUContext>(),
Ydata);
} else {
LOG(FATAL) << "There should be given at least one stride for proposals.";
}
total_rois += num_rois;
Ydata += (num_rois * 5);
im_info += 3;
}
output(0)->Reshape(vector<TIndex>({ total_rois, 5 }));
// distribute rois into K bins
if (OutputSize() > 1) {
CHECK_EQ(max_level - min_level + 1, (int)OutputSize())
<< "Excepted " << OutputSize() << " outputs for levels between "
<< "[" << min_level << ", " << max_level << "].";
vector< vector<TIndex> > roi_bins(OutputSize(), vector<TIndex>());
vector<T*> outputs;
Tensor collective_rois;
collective_rois.ReshapeLike(*output(0));
auto* rois = collective_rois.template mutable_data<T, CPUContext>();
CPUContext::template Copy<T, CPUContext, CPUContext>(collective_rois.count(),
rois,
output(0)->template data<T, CPUContext>());
rcnn::CollectRoIs<T>(total_rois, min_level, max_level,
canonical_level, canonical_scale,
rois,
roi_bins);
for (int i = 0; i < OutputSize(); i++) {
output(i)->Reshape(vector<TIndex>({ std::max((int)roi_bins[i].size(), 1), 5 }));
outputs.push_back(output(i)->template mutable_data<T, CPUContext>());
}
rcnn::DistributeRoIs(roi_bins, rois, outputs);
}
}
template <class Context>
void ProposalOp<Context>::RunOnDevice() {
num_images = input(0).dim(0);
CHECK_EQ(input(-1).count(), num_images * 3)
<< "Excepted " << num_images * 3 << " groups image info, "
<< "but got " << input(-1).count() / 3 << ".";
roi_indices_.Reshape(vector<TIndex>(1, post_nms_top_n));
output(0)->Reshape(vector<TIndex>({ num_images * post_nms_top_n, 5 }));
if (TypeMeta::Id<Context>() == TypeMeta::Id<CPUContext>()) {
if (input(0).template IsType<float>()) RunWithType<float>();
else LOG(FATAL) << "Unsupported input types.";
} else if (TypeMeta::Id<Context>() == TypeMeta::Id<CUDAContext>()) {
if (input(0).template IsType<float>()) RunWithType<float>();
else LOG(FATAL) << "Unsupported input types.";
}
}
DEPLOY_CPU(Proposal);
#ifdef WITH_CUDA
DEPLOY_CUDA(Proposal);
#endif
OPERATOR_SCHEMA(Proposal).NumInputs(3, INT_MAX).NumOutputs(1, INT_MAX);
} // namespace dragon
\ No newline at end of file
Dragon/src/contrib/rcnn/proposal_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_CONTRIB_RCNN_PROPOSAL_OP_H_
#define DRAGON_CONTRIB_RCNN_PROPOSAL_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class ProposalOp final : public Operator<Context> {
public:
ProposalOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
strides(OperatorBase::GetRepeatedArg<int>("strides")),
ratios(OperatorBase::GetRepeatedArg<float>("ratios")),
scales(OperatorBase::GetRepeatedArg<float>("scales")),
pre_nms_top_n(OperatorBase::GetSingleArg<int>("pre_nms_top_n", 6000)),
post_nms_top_n(OperatorBase::GetSingleArg<int>("post_nms_top_n", 300)),
nms_thresh(OperatorBase::GetSingleArg<float>("nms_thresh", (float)0.7)),
min_size(OperatorBase::GetSingleArg<int>("min_size", 16)),
min_level(OperatorBase::GetSingleArg<int>("min_level", 2)),
max_level(OperatorBase::GetSingleArg<int>("max_level", 5)),
canonical_level(OperatorBase::GetSingleArg<int>("canonical_level", 4)),
canonical_scale(OperatorBase::GetSingleArg<int>("canonical_scale", 224)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
vector<int> strides;
vector<float> ratios, scales;
TIndex pre_nms_top_n, post_nms_top_n, min_size, num_images;
TIndex min_level, max_level, canonical_level, canonical_scale;
float nms_thresh;
Tensor anchors_, proposals_, roi_indices_, nms_mask_;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_CONTRIB_RCNN_PROPOSAL_OP_H_
\ No newline at end of file
Dragon/src/operators/norm/cudnn_batch_norm_op.cc
......@@ -244,23 +244,46 @@ void CuDNNBatchNormGradientOp<Context>::TrainingRunWithType() {
template <class Context> template <typename T>
void CuDNNBatchNormGradientOp<Context>::InferenceRunWithType() {
INIT_MULTIPLIER(multiplier, NS);
INIT_MULTIPLIER(num_multiplier, N);
INIT_MULTIPLIER(spatial_multiplier, S);
auto* dYdata = input(-1).template data<T, Context>();
auto* Sdata = input(3).template data<T, Context>();
auto* hVar_data = input(2).template data<T, Context>();
auto* tVar_data = var->template mutable_data<T, Context>();
auto* NSMul_data = multiplier->template data<T, Context>();
auto* SMul_data = spatial_multiplier->template data<T, Context>();
auto* NMul_data = num_multiplier->template data<T, Context>();
auto* NC_data = num_by_chans.template mutable_data<T, Context>();
// gradient w.r.t. scale
if (output(1)->name() != "ignore")
LOG(FATAL) << "The gamma should be fixed if using global stats.";
// gradient w.r.t. bias
if (output(2)->name() != "ignore") {
auto* dBdata = output(2)->template mutable_data<T, Context>();
if (data_format == "NCHW") {
math::Gemv<T, Context>(CblasNoTrans, NC, S,
1.0, dYdata, SMul_data,
0.0, NC_data);
math::Gemv<T, Context>(CblasTrans, N, C,
1.0, NC_data, NMul_data,
1.0, dBdata);
} else if (data_format == "NHWC") {
math::Gemv<T, Context>(CblasTrans, NS, C,
1.0, dYdata, NSMul_data,
1.0, dBdata);
}
}
// gradient w.r.t. x
if (output(0)->name() != "ignore") {
INIT_MULTIPLIER(multiplier, NS);
INIT_MULTIPLIER(num_multiplier, N);
INIT_MULTIPLIER(spatial_multiplier, S);
stddev = ws()->GetBuffer();
stddev->ReshapeLike(input(0));
auto* dYdata = input(-1).template data<T, Context>();
auto* dXdata = output(0)->template mutable_data<T, Context>();
auto* Std_data = stddev->template mutable_data<T, Context>();
auto* Sdata = input(3).template data<T, Context>();
auto* hVar_data = input(2).template data<T, Context>();
auto* tVar_data = var->template mutable_data<T, Context>();
auto* NSMul_data = multiplier->template data<T, Context>();
auto* SMul_data = spatial_multiplier->template data<T, Context>();
auto* NMul_data = num_multiplier->template data<T, Context>();
auto* NC_data = num_by_chans.template mutable_data<T, Context>();
// compute stddev
ctx().template Copy<T, Context, Context>(var->count(), tVar_data, hVar_data);
......
Dragon/src/operators/norm/fused_batch_norm.cc
......@@ -429,21 +429,43 @@ void FusedBatchNormGradientOp<Context>::TrainingRunWithType() {
template <class Context> template <typename T>
void FusedBatchNormGradientOp<Context>::InferenceRunWithType() {
INIT_MULTIPLIER(multiplier, NS);
INIT_MULTIPLIER(num_multiplier, N);
INIT_MULTIPLIER(spatial_multiplier, S);
auto* dYdata = input(-1).template data<T, Context>();
auto* Sdata = input(3).template data<T, Context>();
auto* tVar_data = var->template mutable_data<T, Context>();
auto* NSMul_data = multiplier->template data<T, Context>();
auto* SMul_data = spatial_multiplier->template data<T, Context>();
auto* NMul_data = num_multiplier->template data<T, Context>();
auto* NC_data = num_by_chans.template mutable_data<T, Context>();
// gradient w.r.t. scale
if (output(1)->name() != "ignore")
LOG(FATAL) << "The gamma should be fixed if using global stats.";
// gradient w.r.t. bias
if (output(2)->name() != "ignore") {
auto* dBdata = output(2)->template mutable_data<T, Context>();
if (data_format == "NCHW") {
math::Gemv<T, Context>(CblasNoTrans, NC, S,
1.0, dYdata, SMul_data,
0.0, NC_data);
math::Gemv<T, Context>(CblasTrans, N, C,
1.0, NC_data, NMul_data,
1.0, dBdata);
} else if (data_format == "NHWC") {
math::Gemv<T, Context>(CblasTrans, NS, C,
1.0, dYdata, NSMul_data,
1.0, dBdata);
}
}
// gradient w.r.t. x
if (output(0)->name() != "ignore") {
INIT_MULTIPLIER(multiplier, NS);
INIT_MULTIPLIER(num_multiplier, N);
INIT_MULTIPLIER(spatial_multiplier, S);
auto* dYdata = input(-1).template data<T, Context>();
auto* dXdata = output(0)->template mutable_data<T, Context>();
auto* Std_data = stddev->template mutable_data<T, Context>();
auto* Sdata = input(3).template data<T, Context>();
auto* hVar_data = input(2).template data<T, Context>();
auto* tVar_data = var->template mutable_data<T, Context>();
auto* NSMul_data = multiplier->template data<T, Context>();
auto* SMul_data = spatial_multiplier->template data<T, Context>();
auto* NMul_data = num_multiplier->template data<T, Context>();
auto* NC_data = num_by_chans.template mutable_data<T, Context>();
// divide scale by stddev
math::Div<T, Context>(var->count(), Sdata, tVar_data, tVar_data);
......@@ -492,7 +514,9 @@ void FusedBatchNormGradientOp<Context>::Setup() {
// reshape
num_by_chans.Reshape(vector<TIndex>(1, NC));
output(0)->ReshapeLike(input(0));
output(0)->ReshapeLike(input(0)); // dX
output(1)->ReshapeLike(input(3)); // dScale
output(2)->ReshapeLike(input(3)); // dBias
}
template <class Context>
......
Markdown is supported
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!