Skip to content

P
pytensor
提交 4747cf44 authored 作者: Frédéric Bastien's avatar Frédéric Bastien 提交者： GitHub
浏览文件
操作

Merge pull request #6267 from affanv14/g3

3D Grouped Convolutions
上级 5c92b461 211f0281
全部展开 显示空白字符变更
内嵌 并排
正在显示 包含 198 行增加 和 87 行删除
theano/gpuarray/blas.py
浏览文件 @ 4747cf44
差异被折叠。
theano/gpuarray/c_code/corr3d_gemm.c
浏览文件 @ 4747cf44
......@@ -411,7 +411,8 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
const size_t dilD = 1,
const size_t padH = 0,
const size_t padW = 0,
const size_t padD = 0)
const size_t padD = 0,
const size_t numgroups = 1)
{
if (PyGpuArray_NDIM(bottom) != 5)
{
......@@ -479,11 +480,16 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
const size_t kH = PyGpuArray_DIMS(weight)[2];
const size_t kW = PyGpuArray_DIMS(weight)[3];
const size_t kD = PyGpuArray_DIMS(weight)[4];
if (nChannels != PyGpuArray_DIMS(weight)[1]) {
if (nChannels != PyGpuArray_DIMS(weight)[1] * numgroups) {
PyErr_SetString(PyExc_ValueError,
"GpuCorr3dMM images and kernel must have the same stack size\n");
return NULL;
}
if ((nFilters % numgroups) != 0) {
PyErr_SetString(PyExc_ValueError,
"CorrMM the number of filters must be divisible by the number of groups\n");
return NULL;
}
// implicit dilated filter
const size_t dil_kH = (kH - 1) * dilH + 1;
const size_t dil_kW = (kW - 1) * dilW + 1;
......@@ -511,7 +517,7 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
" weight shape: %ld %ld %ld %ld %ld\n"
" top shape: %ld %ld %ld %ld %ld (expected %ld %ld %ld %ld %ld)\n",
batchSize, nChannels, bottomHeight, bottomWidth, bottomDepth,
nFilters, nChannels, kH, kW, kD,
nFilters, nChannels / numgroups, kH, kW, kD,
PyGpuArray_DIMS(top)[0], PyGpuArray_DIMS(top)[1],
PyGpuArray_DIMS(top)[2], PyGpuArray_DIMS(top)[3], PyGpuArray_DIMS(top)[4],
batchSize, nFilters, topHeight, topWidth, topDepth);
......@@ -542,11 +548,17 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
}
// Define some useful variables
const size_t bottom_stride = PyGpuArray_STRIDES(bottom)[0] / gpuarray_get_elsize(bottom->ga.typecode);
const size_t top_stride = PyGpuArray_STRIDES(top)[0] / gpuarray_get_elsize(top->ga.typecode);
const size_t K_ = col_dim[0];
const size_t batch_bottom_stride = PyGpuArray_STRIDES(bottom)[0] / gpuarray_get_elsize(bottom->ga.typecode);
const size_t batch_top_stride = PyGpuArray_STRIDES(top)[0] / gpuarray_get_elsize(top->ga.typecode);
const size_t group_bottom_stride = (PyGpuArray_STRIDES(bottom)[1] * nChannels / numgroups) / gpuarray_get_elsize(bottom->ga.typecode);
const size_t group_top_stride = (PyGpuArray_STRIDES(top)[1] * nFilters / numgroups) / gpuarray_get_elsize(top->ga.typecode);
const size_t group_weight_stride = (PyGpuArray_STRIDES(weight)[0] * nFilters / numgroups) / gpuarray_get_elsize(weight->ga.typecode);
const size_t K_ = col_dim[0] / numgroups;
const size_t N_ = col_dim[1];
const size_t M_ = nFilters;
const size_t group_col_stride = (K_ * N_);
const size_t M_ = nFilters / numgroups;
PyGpuArrayObject *output;
if (direction == 0) { // forward pass
......@@ -567,20 +579,22 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
for (size_t n = 0; n < batchSize; n++) {
// First, im3d2col
err = im3d2col(
&bottom->ga, n * bottom_stride, nChannels, bottomHeight,
&bottom->ga, n * batch_bottom_stride, nChannels, bottomHeight,
bottomWidth, bottomDepth, kH, kW, kD, dilH, dilW, dilD,
padH, padW, padD, dH, dW, dD, &col->ga);
if (err != GA_NO_ERROR) {
Py_DECREF(col);
return NULL;
}
for ( size_t g = 0; g < numgroups; ++g){
// Second, gemm
err = rgemm(cb_fortran, cb_no_trans, cb_no_trans,
N_, M_, K_, 1,
&col->ga, 0, N_,
&weight->ga, 0, K_,
&col->ga, g * group_col_stride, N_,
&weight->ga, g * group_weight_stride, K_,
0,
&top->ga, n * top_stride, N_);
&top->ga, n * batch_top_stride + g * group_top_stride, N_);
}
if (err != GA_NO_ERROR) {
PyErr_Format(PyExc_RuntimeError,
"GpuCorr3dMM forward encountered an error running gemm.");
......@@ -607,7 +621,7 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
for (size_t n = 0; n < batchSize; n++) {
// First, im3d2col
err = im3d2col(
&bottom->ga, n * bottom_stride, nChannels, bottomHeight,
&bottom->ga, n * batch_bottom_stride, nChannels, bottomHeight,
bottomWidth, bottomDepth, kH, kW, kD, dilH, dilW, dilD,
padH, padW, padD, dH, dW, dD, &col->ga);
if (err != GA_NO_ERROR) {
......@@ -618,12 +632,14 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
// Note that we accumulate into weight. We do so by setting beta = 0
// for the first iteration and beta = 1 for subsequent ones. (This
// is faster than setting weight to all zeros before the loop.)
for ( size_t g = 0; g < numgroups; ++g){
err = rgemm(cb_fortran, cb_trans, cb_no_trans,
K_, M_, N_, 1,
&col->ga, 0, N_,
&top->ga, n * top_stride, N_,
&col->ga, g * group_col_stride, N_,
&top->ga, n * batch_top_stride + g * group_top_stride, N_,
(n == 0) ? 0 : 1,
&weight->ga, 0, K_);
&weight->ga, g * group_weight_stride, K_);
}
if (err != GA_NO_ERROR) {
PyErr_Format(PyExc_RuntimeError,
"GpuCorr3dMM grad weights encountered an error running gemm.");
......@@ -658,12 +674,14 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
// Iterate over batch
for (size_t n = 0; n < batchSize; n++) {
// gemm into columns
for ( size_t g = 0; g < numgroups; ++g){
err = rgemm(cb_fortran, cb_no_trans, cb_trans,
N_, K_, M_, 1,
&top->ga, n * top_stride, N_,
&weight->ga, 0, K_,
&top->ga, n * batch_top_stride + g * group_top_stride, N_,
&weight->ga, g * group_weight_stride, K_,
0,
&col->ga, 0, N_);
&col->ga, g * group_col_stride, N_);
}
if (err != GA_NO_ERROR) {
PyErr_Format(PyExc_RuntimeError,
"GpuCorr3dMM grad inputs encountered an error running gemm.");
......@@ -674,7 +692,7 @@ PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
err = col2im3d(&col->ga, nChannels,
bottomHeight, bottomWidth, bottomDepth,
kH, kW, kD, dilH, dilW, dilD, padH, padW, padD,
dH, dW, dD, &bottom->ga, n * bottom_stride);
dH, dW, dD, &bottom->ga, n * batch_bottom_stride);
if (err != GA_NO_ERROR) {
Py_DECREF(col);
return NULL;
......
theano/gpuarray/dnn.py
浏览文件 @ 4747cf44
......@@ -2790,6 +2790,8 @@ def local_abstractconv_cudnn_graph(op, context_name, inputs, outputs):
if version(raises=False) < 6000 and op.filter_dilation != (1, 1):
return None
if op.num_groups > 1:
return None
inp1 = inputs[0]
inp2 = inputs[1]
......@@ -2839,6 +2841,8 @@ def local_abstractconv3d_cudnn_graph(op, context_name, inputs, outputs):
if version(raises=False) < 6000 and op.filter_dilation != (1, 1, 1):
return None
if op.num_groups > 1:
return None
inp1 = inputs[0]
inp2 = inputs[1]
......
theano/gpuarray/opt.py
浏览文件 @ 4747cf44
......@@ -1707,7 +1707,8 @@ def local_abstractconv3d_gemm(node):
border_mode = node.op.border_mode
subsample = node.op.subsample
filter_dilation = node.op.filter_dilation
if ((border_mode == 'full') and (subsample == (1, 1, 1))):
num_groups = node.op.num_groups
if ((border_mode == 'full') and (subsample == (1, 1, 1)) and num_groups == 1):
if not node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1, ::-1]
# need to dimshuffle the kernel for full convolution
......@@ -1724,7 +1725,8 @@ def local_abstractconv3d_gemm(node):
# By default use GpuCorr3dMM
rval = GpuCorr3dMM(border_mode,
subsample,
filter_dilation)(gpu_contiguous(img),
filter_dilation,
num_groups)(gpu_contiguous(img),
gpu_contiguous(kern))
# call GpuCorr3dMM_gradWeights if good
......@@ -1737,7 +1739,8 @@ def local_abstractconv3d_gemm(node):
(None not in node.op.imshp[-3:]) and
(node.op.kshp is not None) and
(None not in node.op.kshp) and
border_mode != "half"):
border_mode != "half" and
num_groups == 1):
# we know the kernel and output size
prod1 = node.op.kshp[0] * node.op.kshp[1] * node.op.kshp[2]
prod2 = ((node.op.imshp[-3] - node.op.kshp[0] + 1) *
......@@ -1929,7 +1932,8 @@ def local_abstractconv3d_gradweights_gemm(node):
rval = GpuCorr3dMM_gradWeights(border_mode=node.op.border_mode,
subsample=node.op.subsample,
filter_dilation=node.op.filter_dilation)(
filter_dilation=node.op.filter_dilation,
num_groups=node.op.num_groups)(
gpu_contiguous(img), gpu_contiguous(topgrad), shape)
if node.op.filter_flip:
rval = rval[:, :, ::-1, ::-1, ::-1]
......@@ -1999,7 +2003,8 @@ def local_abstractconv3d_gradinputs_gemm(node):
rval = GpuCorr3dMM_gradInputs(border_mode=node.op.border_mode,
subsample=node.op.subsample,
filter_dilation=node.op.filter_dilation)(
filter_dilation=node.op.filter_dilation,
num_groups=node.op.num_groups)(
gpu_contiguous(kern), gpu_contiguous(topgrad), shape)
return [rval]
......
theano/gpuarray/tests/test_dnn.py
浏览文件 @ 4747cf44
......@@ -2292,11 +2292,11 @@ def dconv2di(border_mode, subsample, filter_dilation, num_groups):
class Cudnn_grouped_conv(Grouped_conv_noOptim):
mode = mode_with_gpu
conv2d = staticmethod(dconv2d)
conv2d_gradw = staticmethod(dconv2dw)
conv2d_gradi = staticmethod(dconv2di)
conv2d_op = dnn.GpuDnnConv
conv2d_gradw_op = dnn.GpuDnnConvGradW
conv2d_gradi_op = dnn.GpuDnnConvGradI
conv = staticmethod(dconv2d)
conv_gradw = staticmethod(dconv2dw)
conv_gradi = staticmethod(dconv2di)
conv_op = dnn.GpuDnnConv
conv_gradw_op = dnn.GpuDnnConvGradW
conv_gradi_op = dnn.GpuDnnConvGradI
flip_filter = False
is_dnn = True
theano/gpuarray/tests/test_gemmcorr.py
浏览文件 @ 4747cf44
......@@ -224,11 +224,11 @@ class TestCorrMM(unittest.TestCase):
class TestGroupGpuCorr2d(Grouped_conv_noOptim):
mode = theano.compile.get_mode("FAST_RUN")
conv2d = GpuCorrMM
conv2d_gradw = GpuCorrMM_gradWeights
conv2d_gradi = GpuCorrMM_gradInputs
conv2d_op = GpuCorrMM
conv2d_gradw_op = GpuCorrMM_gradWeights
conv2d_gradi_op = GpuCorrMM_gradInputs
conv = GpuCorrMM
conv_gradw = GpuCorrMM_gradWeights
conv_gradi = GpuCorrMM_gradInputs
conv_op = GpuCorrMM
conv_gradw_op = GpuCorrMM_gradWeights
conv_gradi_op = GpuCorrMM_gradInputs
flip_filter = True
is_dnn = False
theano/gpuarray/tests/test_gemmcorr3d.py
浏览文件 @ 4747cf44
......@@ -11,6 +11,7 @@ from theano.tensor.nnet.corr3d import Corr3dMM, Corr3dMM_gradWeights, Corr3dMM_g
from ..type import gpuarray_shared_constructor
from ..blas import GpuCorr3dMM, GpuCorr3dMM_gradWeights, GpuCorr3dMM_gradInputs
from .config import mode_with_gpu, mode_without_gpu, ref_cast
from theano.tensor.nnet.tests.test_abstract_conv import Grouped_conv3d_noOptim
class TestCorr3dMM(unittest.TestCase):
......@@ -218,3 +219,15 @@ class TestCorr3dMM(unittest.TestCase):
verify_grad=False)
self.run_gradinput(inputs_shape=(1, 1024, 3, 3, 1),
filters_shape=(1, 1, 1, 1, 1024))
class TestGroupGpuCorr3d(Grouped_conv3d_noOptim):
mode = theano.compile.get_mode("FAST_RUN")
conv = GpuCorr3dMM
conv_gradw = GpuCorr3dMM_gradWeights
conv_gradi = GpuCorr3dMM_gradInputs
conv_op = GpuCorr3dMM
conv_gradw_op = GpuCorr3dMM_gradWeights
conv_gradi_op = GpuCorr3dMM_gradInputs
flip_filter = True
is_dnn = False
theano/tensor/nnet/c_code/corr3d_gemm.c
浏览文件 @ 4747cf44
......@@ -127,7 +127,8 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
const int dilD = 1,
const int padH = 0,
const int padW = 0,
const int padD = 0)
const int padD = 0,
const int numgroups=1)
{
if (PyArray_NDIM(bottom) != 5)
{
......@@ -178,11 +179,16 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
const int kH = PyArray_DIMS(weight)[2];
const int kW = PyArray_DIMS(weight)[3];
const int kD = PyArray_DIMS(weight)[4];
if (nChannels != PyArray_DIMS(weight)[1]) {
if (nChannels != PyArray_DIMS(weight)[1] * numgroups) {
PyErr_SetString(PyExc_ValueError,
"Corr3dMM images and kernel must have the same stack size\n");
return NULL;
}
if ((nFilters %% numgroups) != 0) {
PyErr_SetString(PyExc_ValueError,
"CorrMM the number of filters must be divisible by the number of groups\n");
return NULL;
}
// implicit dilated filter
const int dil_kH = (kH - 1) * dilH + 1;
const int dil_kW = (kW - 1) * dilW + 1;
......@@ -210,7 +216,7 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
" weight shape: %%d %%d %%d %%d %%d\n"
" top shape: %%ld %%ld %%ld %%ld %%ld (expected %%d %%d %%d %%d %%d)\n",
batchSize, nChannels, bottomHeight, bottomWidth, bottomDepth,
nFilters, nChannels, kH, kW, kD,
nFilters, nChannels / numgroups, kH, kW, kD,
PyArray_DIMS(top)[0], PyArray_DIMS(top)[1],
PyArray_DIMS(top)[2], PyArray_DIMS(top)[3], PyArray_DIMS(top)[4],
batchSize, nFilters, topHeight, topWidth, topDepth);
......@@ -241,12 +247,16 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
}
// Define some useful variables
const int bottom_stride = PyArray_STRIDES(bottom)[0]/%(n_bytes)f;
const int top_stride = PyArray_STRIDES(top)[0]/%(n_bytes)f;
const int K_ = col_dim[1];
const int batch_bottom_stride = PyArray_STRIDES(bottom)[0]/%(n_bytes)f;
const int group_bottom_stride = (PyArray_STRIDES(bottom)[1] * nChannels / numgroups)/%(n_bytes)f;
const int batch_top_stride = PyArray_STRIDES(top)[0]/%(n_bytes)f;
const int group_top_stride = (PyArray_STRIDES(top)[1] * nFilters / numgroups)/%(n_bytes)f;
const int K_ = col_dim[1] / numgroups;
const int N_ = col_dim[2];
const int col_stride = (K_ * N_);
const int M_ = nFilters;
const int col_stride = (K_ * N_ * numgroups);
const int group_col_stride = (K_ * N_);
const int group_weight_stride = (PyArray_STRIDES(weight)[0] * nFilters / numgroups)/%(n_bytes)f;
const int M_ = nFilters / numgroups;
const %(c_float_type)s one = 1.0;
const %(c_float_type)s zero = 0.0;
char NTrans = 'N';
......@@ -280,18 +290,21 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
for (int n = 0; n < batchSize; ++n) {
int tid = %(omp_get_thread_num)s;
// First, im3d2col
im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * bottom_stride, nChannels,
bottomHeight, bottomWidth, bottomDepth,
im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * batch_bottom_stride,
nChannels, bottomHeight, bottomWidth, bottomDepth,
kH, kW, kD, dilH, dilW, dilD, padH, padW, padD, dH, dW, dD,
(%(float_type)s*)PyArray_DATA(col)+ tid * col_stride);
for ( int g = 0; g < numgroups; ++g){
// Second, gemm
%(gemm)s(&NTrans, &NTrans,
&N_, &M_, &K_,
&one,
(%(float_type)s*)PyArray_DATA(col)+ tid * col_stride, &N_,
(%(float_type)s*)PyArray_DATA(weight), &K_,
(%(float_type)s*)PyArray_DATA(col)+ tid * col_stride + g * group_col_stride, &N_,
(%(float_type)s*)PyArray_DATA(weight) + g * group_weight_stride, &K_,
&zero,
(%(float_type)s*)PyArray_DATA(top) + n * top_stride, &N_);
(%(float_type)s*)PyArray_DATA(top) + n * batch_top_stride + g * group_top_stride, &N_);
}
}
// Restore to previous blas threads
%(blas_set_num_threads)s(blas_threads_saved);
......@@ -300,7 +313,7 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
output = weight;
npy_intp weight_dim[2];
weight_dim[0] = (npy_intp)max_threads;
weight_dim[1] = (npy_intp)(M_ * K_);
weight_dim[1] = (npy_intp)(M_ * K_ * numgroups);
PyArrayObject* local_weight = (PyArrayObject*)PyArray_ZEROS(2,
weight_dim, PyArray_TYPE(weight), 0);
......@@ -322,10 +335,12 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
for (int n = 0; n < batchSize; ++n) {
int tid = %(omp_get_thread_num)s;
// First, im2col
im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * bottom_stride, nChannels,
bottomHeight, bottomWidth, bottomDepth,
im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * batch_bottom_stride,
nChannels, bottomHeight, bottomWidth, bottomDepth,
kH, kW, kD, dilH, dilW, dilD, padH, padW, padD, dH, dW, dD,
(%(float_type)s*)PyArray_DATA(col)+ tid * col_stride);
for ( int g = 0; g < numgroups; ++g){
// Second, gemm
// Note that we accumulate into weight. We do so by setting beta = 0
// for the first iteration and beta = 1 for subsequent ones. (This
......@@ -333,12 +348,13 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
%(gemm)s(&Trans, &NTrans,
&K_, &M_, &N_,
&one,
(%(float_type)s*)PyArray_DATA(col) + tid * col_stride, &N_,
(%(float_type)s*)PyArray_DATA(top) + n * top_stride, &N_,
(%(float_type)s*)PyArray_DATA(col) + tid * col_stride + g * group_col_stride, &N_,
(%(float_type)s*)PyArray_DATA(top) + n * batch_top_stride + g * group_top_stride, &N_,
(n == 0) ? &zero : &one,
(%(float_type)s*)PyArray_DATA(local_weight) +
(%(float_type)s*)PyArray_DATA(local_weight) + g * group_weight_stride +
tid * weight_dim[1], &K_);
}
}
// Restore to previous blas threads
%(blas_set_num_threads)s(blas_threads_saved);
......@@ -370,20 +386,23 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
%(blas_set_num_threads)s(1);
%(omp_flags)s
for (int n = 0; n < batchSize; ++n) {
// gemm into columns
int tid = %(omp_get_thread_num)s;
for ( int g = 0; g < numgroups; ++g){
// gemm into columns
%(gemm)s(&NTrans, &Trans,
&N_, &K_, &M_,
&one,
(%(float_type)s*)PyArray_DATA(top) + n * top_stride, &N_,
(%(float_type)s*)PyArray_DATA(weight), &K_,
(%(float_type)s*)PyArray_DATA(top) + n * batch_top_stride + g * group_top_stride, &N_,
(%(float_type)s*)PyArray_DATA(weight) + g * group_weight_stride, &K_,
&zero,
(%(float_type)s*)PyArray_DATA(col) + tid * col_stride, &N_);
(%(float_type)s*)PyArray_DATA(col) + tid * col_stride + g * group_col_stride, &N_);
}
// col2im back to the data
col2im3d((%(float_type)s*)PyArray_DATA(col) + tid * col_stride, nChannels,
bottomHeight, bottomWidth, bottomDepth,
kH, kW, kD, dilH, dilW, dilD, padH, padW, padD, dH, dW, dD,
(%(float_type)s*)PyArray_DATA(bottom) + n * bottom_stride);
(%(float_type)s*)PyArray_DATA(bottom) + n * batch_bottom_stride);
}
// Restore to previous blas threads
%(blas_set_num_threads)s(blas_threads_saved);
......
theano/tensor/nnet/corr3d.py
浏览文件 @ 4747cf44
......@@ -40,9 +40,11 @@ class BaseCorr3dMM(gof.OpenMPOp):
Perform subsampling of the output (default: (1, 1, 1)).
filter_dilation
Perform dilated correlation (default: (1, 1, 1))
num_groups
Perform grouped convolutions (default: 1)
"""
check_broadcast = False
__props__ = ('border_mode', 'subsample', 'filter_dilation')
__props__ = ('border_mode', 'subsample', 'filter_dilation', 'num_groups')
_direction = None
......@@ -51,10 +53,11 @@ class BaseCorr3dMM(gof.OpenMPOp):
('DIRECTION_BACKPROP_INPUTS', 'backprop inputs')), # 2
dH=int64, dW=int64, dD=int64,
dilH=int64, dilW=int64, dilD=int64,
padH=int64, padW=int64, padD=int64)
padH=int64, padW=int64, padD=int64,
num_groups=int64)
def __init__(self, border_mode="valid", subsample=(1, 1, 1),
filter_dilation=(1, 1, 1), openmp=None):
filter_dilation=(1, 1, 1), openmp=None, num_groups=1):
super(BaseCorr3dMM, self).__init__(openmp=openmp)
if isinstance(border_mode, integer_types):
if border_mode < 0:
......@@ -82,6 +85,9 @@ class BaseCorr3dMM(gof.OpenMPOp):
raise ValueError("filter_dilation must have three elements")
self.subsample = tuple(subsample)
self.filter_dilation = tuple(filter_dilation)
if num_groups < 1:
raise ValueError("Number of groups should be greater than 0")
self.num_groups = num_groups
if not theano.config.blas.ldflags:
# Theano will use a NumPy C implementation of [sd]gemm_ instead.
......@@ -127,11 +133,12 @@ class BaseCorr3dMM(gof.OpenMPOp):
padD = property(lambda self: self.pad[2])
def __str__(self):
return '%s{%s, %s, %s}' % (
return '%s{%s, %s, %s, %s}' % (
self.__class__.__name__,
self.border_mode,
str(self.subsample),
str(self.filter_dilation))
str(self.filter_dilation),
str(self.num_groups))
@staticmethod
def as_common_dtype(in1, in2):
......@@ -141,6 +148,11 @@ class BaseCorr3dMM(gof.OpenMPOp):
dtype = theano.scalar.upcast(in1.dtype, in2.dtype)
return in1.astype(dtype), in2.astype(dtype)
def __setstate__(self, d):
self.__dict__.update(d)
if not hasattr(self, 'num_groups'):
self.num_groups = 1
def c_support_code(self):
ccodes = blas_headers.blas_header_text()
if self.blas_type == 'openblas':
......@@ -170,7 +182,7 @@ class BaseCorr3dMM(gof.OpenMPOp):
def c_code_cache_version(self):
# raise this whenever modifying any of the support_code_files
return (7, self.openmp, blas_header_version())
return (8, self.openmp, blas_header_version())
def c_support_code_apply(self, node, nodename):
# REMEMBER TO RAISE c_code_cache_version when changing any of
......@@ -293,6 +305,7 @@ class BaseCorr3dMM(gof.OpenMPOp):
int padH = %(params)s->padH;
int padW = %(params)s->padW;
int padD = %(params)s->padD;
int numgroups = %(params)s->num_groups;
PyArrayObject * bottom = %(bottom)s;
PyArrayObject * weights = %(weights)s;
......@@ -428,7 +441,7 @@ class BaseCorr3dMM(gof.OpenMPOp):
// output is weights: (num_filters, num_channels, height, width, depth)
// height and width: weights = (bottom + 2*pad - (top - 1) * sample - 1) / dil + 1
out_dim[0] = (npy_intp)PyArray_DIMS(top)[1];
out_dim[1] = (npy_intp)PyArray_DIMS(bottom)[1];
out_dim[1] = (npy_intp)PyArray_DIMS(bottom)[1] / numgroups;
out_dim[2] = (npy_intp)kH; // already inferred further above
out_dim[3] = (npy_intp)kW; // how convenient
out_dim[4] = (npy_intp)kD;
......@@ -454,7 +467,7 @@ class BaseCorr3dMM(gof.OpenMPOp):
// output is bottom: (batchsize, num_channels, height, width, depth)
// height and width: bottom = (top - 1) * sample + (weights-1)*dil + 1 - 2*pad
out_dim[0] = (npy_intp)PyArray_DIMS(top)[0];
out_dim[1] = (npy_intp)PyArray_DIMS(weights)[1];
out_dim[1] = (npy_intp)PyArray_DIMS(weights)[1] * numgroups;
out_dim[2] = (npy_intp)((%(height)s != -1) ? %(height)s : (PyArray_DIMS(top)[2] - 1) * dH + (PyArray_DIMS(weights)[2]-1)*dilH + 1 - 2*padH);
out_dim[3] = (npy_intp)((%(width)s != -1) ? %(width)s : (PyArray_DIMS(top)[3] - 1) * dW + (PyArray_DIMS(weights)[3]-1)*dilW + 1 - 2*padW);
out_dim[4] = (npy_intp)((%(depth)s != -1) ? %(depth)s : (PyArray_DIMS(top)[4] - 1) * dD + (PyArray_DIMS(weights)[4]-1)*dilD + 1 - 2*padD);
......@@ -516,7 +529,8 @@ class BaseCorr3dMM(gof.OpenMPOp):
// Call corr3dMM code
out2 = corr3dMM(%(bottom)s, %(weights)s, %(top)s, direction,
dH, dW, dD, dilH, dilW, dilD, padH, padW, padD);
dH, dW, dD, dilH, dilW, dilD, padH, padW, padD,
numgroups);
if (out2==NULL){
%(fail)s
}
......@@ -552,7 +566,8 @@ class Corr3dMM(BaseCorr3dMM):
The filter dilation operation applied to each input image.
Should be a tuple with 3 elements.
Set to `(1, 1, 1)` to disable filter dilation.
num_groups
Perform grouped convolutions (default: 1)
"""
_direction = "forward"
......@@ -592,11 +607,13 @@ class Corr3dMM(BaseCorr3dMM):
top, = grads
d_bottom = Corr3dMM_gradInputs(self.border_mode,
self.subsample,
self.filter_dilation)(weights, top,
self.filter_dilation,
num_groups=self.num_groups)(weights, top,
bottom.shape[-3:])
d_weights = Corr3dMM_gradWeights(self.border_mode,
self.subsample,
self.filter_dilation)(bottom, top,
self.filter_dilation,
num_groups=self.num_groups)(bottom, top,
weights.shape[-3:])
return d_bottom, d_weights
......@@ -653,6 +670,7 @@ class Corr3dMM_gradWeights(BaseCorr3dMM):
imshp = input_shape[0]
topshp = input_shape[1]
ssize, imshp = imshp[1], list(imshp[2:])
ssize = ssize // self.num_groups
nkern, topshp = topshp[1], list(topshp[2:])
height_width_depth = node.inputs[-3:]
if ((dH != 1) or (padH == -1)):
......@@ -691,11 +709,13 @@ class Corr3dMM_gradWeights(BaseCorr3dMM):
weights, = grads
d_bottom = Corr3dMM_gradInputs(self.border_mode,
self.subsample,
self.filter_dilation)(weights, top,
self.filter_dilation,
num_groups=self.num_groups)(weights, top,
bottom.shape[-3:])
d_top = Corr3dMM(self.border_mode,
self.subsample,
self.filter_dilation)(bottom, weights)
self.filter_dilation,
num_groups=self.num_groups)(bottom, weights)
d_height_width_depth = ((theano.gradient.DisconnectedType()(),) * 3
if len(inp) == 5 else ())
return (d_bottom, d_top) + d_height_width_depth
......@@ -738,6 +758,10 @@ class Corr3dMM_gradInputs(BaseCorr3dMM):
as_tensor_variable(shape[1]).astype('int64'),
as_tensor_variable(shape[2]).astype('int64')]
if self.num_groups > 1:
broadcastable = [topgrad.type.broadcastable[0], False,
False, False, False]
else:
broadcastable = [topgrad.type.broadcastable[0], kern.type.broadcastable[1],
False, False, False]
dtype = kern.type.dtype
......@@ -758,6 +782,7 @@ class Corr3dMM_gradInputs(BaseCorr3dMM):
kshp = input_shape[0]
topshp = input_shape[1]
ssize, kshp = kshp[1], list(kshp[2:])
ssize = ssize * self.num_groups
bsize, topshp = topshp[0], list(topshp[2:])
height_width_depth = node.inputs[-3:]
if padH == -1:
......@@ -807,12 +832,14 @@ class Corr3dMM_gradInputs(BaseCorr3dMM):
bottom, = grads
d_weights = Corr3dMM_gradWeights(self.border_mode,
self.subsample,
self.filter_dilation)(bottom,
self.filter_dilation,
num_groups=self.num_groups)(bottom,
top,
weights.shape[-3:])
d_top = Corr3dMM(self.border_mode,
self.subsample,
self.filter_dilation)(bottom, weights)
self.filter_dilation,
num_groups=self.num_groups)(bottom, weights)
d_height_width_depth = ((theano.gradient.DisconnectedType()(),) * 3
if len(inp) == 5 else ())
return (d_weights, d_top) + d_height_width_depth
......
theano/tensor/nnet/opt.py
浏览文件 @ 4747cf44
......@@ -114,7 +114,8 @@ def local_abstractconv3d_gemm(node):
kern = kern[:, :, ::-1, ::-1, ::-1]
rval = Corr3dMM(border_mode=node.op.border_mode,
subsample=node.op.subsample,
filter_dilation=node.op.filter_dilation)(img, kern)
filter_dilation=node.op.filter_dilation,
num_groups=node.op.num_groups)(img, kern)
copy_stack_trace(node.outputs[0], rval)
return [rval]
......@@ -163,7 +164,8 @@ def local_abstractconv3d_gradweight_gemm(node):
rval = Corr3dMM_gradWeights(border_mode=node.op.border_mode,
subsample=node.op.subsample,
filter_dilation=node.op.filter_dilation)(img, topgrad, shape)
filter_dilation=node.op.filter_dilation,
num_groups=node.op.num_groups)(img, topgrad, shape)
copy_stack_trace(node.outputs[0], rval)
# need to flip the kernel if necessary
......@@ -219,7 +221,8 @@ def local_abstractconv3d_gradinputs_gemm(node):
kern = kern[:, :, ::-1, ::-1, ::-1]
rval = Corr3dMM_gradInputs(border_mode=node.op.border_mode,
subsample=node.op.subsample,
filter_dilation=node.op.filter_dilation)(kern, topgrad,
filter_dilation=node.op.filter_dilation,
num_groups=node.op.num_groups)(kern, topgrad,
shape)
copy_stack_trace(node.outputs[0], rval)
......@@ -267,6 +270,8 @@ def local_conv3d_cpu(node):
return None
if node.op.filter_dilation != (1, 1, 1):
return None
if node.op.num_groups > 1:
return None
bias = theano.tensor.zeros_like(kern[:, 0, 0, 0, 0])
......@@ -419,6 +424,8 @@ def local_conv3d_gradweight_cpu(node):
return None
if node.op.filter_dilation != (1, 1, 1):
return None
if node.op.num_groups > 1:
return None
# conv3D expects shape (batch, row, column, time, channel)
img = img.dimshuffle(0, 2, 3, 4, 1)
......@@ -544,6 +551,8 @@ def local_conv3d_gradinputs_cpu(node):
return None
if node.op.filter_dilation != (1, 1, 1):
return None
if node.op.num_groups > 1:
return None
# need to flip the kernel if necessary (conv3D does not flip)
if node.op.filter_flip:
......
theano/tensor/nnet/tests/test_corr.py
浏览文件 @ 4747cf44
......@@ -422,12 +422,12 @@ class TestGroupCorr2d(Grouped_conv_noOptim):
mode = theano.compile.get_mode("FAST_RUN")
else:
mode = None
conv2d = corr.CorrMM
conv2d_gradw = corr.CorrMM_gradWeights
conv2d_gradi = corr.CorrMM_gradInputs
conv2d_op = corr.CorrMM
conv2d_gradw_op = corr.CorrMM_gradWeights
conv2d_gradi_op = corr.CorrMM_gradInputs
conv = corr.CorrMM
conv_gradw = corr.CorrMM_gradWeights
conv_gradi = corr.CorrMM_gradInputs
conv_op = corr.CorrMM
conv_gradw_op = corr.CorrMM_gradWeights
conv_gradi_op = corr.CorrMM_gradInputs
flip_filter = True
is_dnn = False
......@@ -440,13 +440,13 @@ class TestGroupCorr2d(Grouped_conv_noOptim):
kern_sym = T.tensor4('kern')
# grouped convolution graph
conv_group = self.conv2d(num_groups=groups)(bottom_sym, kern_sym)
conv_group = self.conv(num_groups=groups)(bottom_sym, kern_sym)
gconv_func = theano.function([bottom_sym, kern_sym], conv_group, mode=self.mode)
# Graph for the normal hard way
kern_offset = kern_sym.shape[0] // groups
bottom_offset = bottom_sym.shape[1] // groups
split_conv_output = [self.conv2d()(bottom_sym[:, i * bottom_offset:(i + 1) * bottom_offset, :, :],
split_conv_output = [self.conv()(bottom_sym[:, i * bottom_offset:(i + 1) * bottom_offset, :, :],
kern_sym[i * kern_offset:(i + 1) * kern_offset, :, :, :])
for i in range(groups)]
concatenated_output = T.concatenate(split_conv_output, axis=1)
......
theano/tensor/nnet/tests/test_corr3d.py
浏览文件 @ 4747cf44
......@@ -12,6 +12,7 @@ import theano
import theano.tensor as T
from theano.tests import unittest_tools as utt
from theano.tensor.nnet import corr3d, conv
from theano.tensor.nnet.tests.test_abstract_conv import Grouped_conv3d_noOptim
class TestCorr3D(utt.InferShapeTester):
......@@ -418,6 +419,21 @@ class TestCorr3D(utt.InferShapeTester):
self.validate((3, 1, 7, 5, 5), (2, 1, 2, 3, 3), (2, 1, 1), non_contiguous=True)
class TestGroupCorr3d(Grouped_conv3d_noOptim):
if theano.config.mode == "FAST_COMPILE":
mode = theano.compile.get_mode("FAST_RUN")
else:
mode = None
conv = corr3d.Corr3dMM
conv_gradw = corr3d.Corr3dMM_gradWeights
conv_gradi = corr3d.Corr3dMM_gradInputs
conv_op = corr3d.Corr3dMM
conv_gradw_op = corr3d.Corr3dMM_gradWeights
conv_gradi_op = corr3d.Corr3dMM_gradInputs
flip_filter = True
is_dnn = False
if __name__ == '__main__':
t = TestCorr3D('setUp')
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论