Skip to content

P
pytensor
提交 906f4eac authored 作者: --global's avatar --global
浏览文件
操作

Add test cases for fwd and bwd 3d convolution

上级 8ff1685f
显示空白字符变更
内嵌 并排
正在显示 包含 138 行增加 和 105 行删除
theano/sandbox/cuda/tests/test_dnn.py
浏览文件 @ 906f4eac
...@@ -2,7 +2,7 @@ import logging ...@@ -2,7 +2,7 @@ import logging
from nose.plugins.skip import SkipTest from nose.plugins.skip import SkipTest
import numpy import numpy
from itertools import product from itertools import chain, product
import theano import theano
from six import StringIO from six import StringIO
...@@ -922,13 +922,49 @@ def test_dnn_conv_grad(): ...@@ -922,13 +922,49 @@ def test_dnn_conv_grad():
utt.verify_grad(dconvw, [img_val, kern_val, out_val]) utt.verify_grad(dconvw, [img_val, kern_val, out_val])
def get_conv3d_test_cases():
# Every element of test_shapes follows the format
# [input_shape, filter_shape, subsample]
test_shapes = [# Test with standard size inputs and kernels
[(128, 3, 5, 5, 5), (64, 3, 1, 2, 4), (1, 1, 1)],
[(8, 4, 20, 12, 15), (5, 4, 6, 12, 4), (2, 2, 2)],
[(8, 1, 20, 12, 15), (5, 1, 6, 12, 4), (3, 3, 3)],
[(8, 1, 20, 12, 15), (5, 1, 6, 12, 4), (3, 2, 1)],
[(8, 1, 20, 12, 15), (5, 1, 6, 12, 4), (3, 2, 1)],
# Test with 1x1x1 filters
[(8, 1, 10, 10, 10), (10, 1, 1, 1, 1), (1, 1, 1)],
# Test with dimensions larger than 1024 (thread block dim)
[(1025, 1, 2, 3, 4), (5, 1, 1, 2, 3), (1, 1, 1)],
[(8, 1, 2, 3, 4), (1025, 1, 1, 2, 3), (1, 1, 1)],
[(8, 1025, 2, 3, 4), (5, 1025, 1, 2, 3), (1, 1, 1)],
[(8, 1, 1030, 3, 4), (5, 1, 1025, 2, 3), (1, 1, 1)],
[(8, 1, 2, 1030, 4), (5, 1, 1, 1025, 3), (1, 1, 1)],
[(8, 1, 2, 3, 1030), (5, 1, 1, 2, 1025), (1, 1, 1)],
# The equivalent of this caused a crash with conv2d
[(1, 1, 1, 44800, 1), (6, 1, 1, 1, 1), (1, 1, 1)]]
# With border mode 'full', test with kernel bigger than image in some/all
# dimensions
test_shapes_full = [[(6, 2, 2, 2, 2), (4, 2, 3, 1, 1), (1, 1, 1)],
[(6, 2, 2, 2, 2), (4, 2, 1, 3, 1), (1, 1, 1)],
[(6, 2, 2, 2, 2), (4, 2, 1, 1, 3), (1, 1, 1)],
[(6, 2, 2, 2, 2), (4, 2, 5, 5, 5), (1, 1, 1)],
]
border_modes = ['valid', 'full', (1, 2, 3), (3, 2, 1)]
conv_modes = ['conv', 'cross']
itt = chain(product(test_shapes, border_modes, conv_modes),
product(test_shapes_full, ['full'], conv_modes))
return itt
def test_conv3d_fwd(): def test_conv3d_fwd():
if not cuda.dnn.dnn_available() and dnn.version()[0] >= 3000: if not cuda.dnn.dnn_available() and dnn.version()[0] >= 3000:
raise SkipTest('"3D conv not supported in cudnn v1') raise SkipTest('"CuDNN 3D convolution requires CuDNN v3')
def run_conv3d_fwd(inputs_shape, filters_shape, def run_conv3d_fwd(inputs_shape, filters_shape, subsample,
subsample=(1, 1, 1)): border_mode, conv_mode):
inputs_val = numpy.random.random(inputs_shape).astype('float32') inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32') filters_val = numpy.random.random(filters_shape).astype('float32')
...@@ -936,132 +972,129 @@ def test_conv3d_fwd(): ...@@ -936,132 +972,129 @@ def test_conv3d_fwd():
inputs = shared(inputs_val) inputs = shared(inputs_val)
filters = shared(filters_val) filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32')) bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
conv_ref = theano.tensor.nnet.conv3D(V=inputs.dimshuffle(0, 2, 3, 4, 1),
W=filters.dimshuffle(0, 2, 3, 4, 1), # Compile a theano function for the CuDNN implementation
b=bias, d=subsample)
conv = dnn.dnn_conv3d(img=inputs, kerns=filters, conv = dnn.dnn_conv3d(img=inputs, kerns=filters,
border_mode="valid", subsample=subsample, conv_mode='cross') border_mode=border_mode, subsample=subsample,
f_ref = theano.function([], conv_ref.dimshuffle(0, 4, 1, 2, 3)) conv_mode=conv_mode)
f = theano.function([], conv, mode=mode_with_gpu) f = theano.function([], conv, mode=mode_with_gpu)
res_ref = f_ref() # If conv_mode is 'conv' the reference implementation should use
res = f() # filters filpped according to the width, height and time axis
if conv_mode == 'conv':
flipped_filters = filters[:,:,::-1,::-1,::-1]
else:
flipped_filters = filters
utt.assert_allclose(res_ref, res) # If border mode is anything but 'valid', the reference implementation
# should operate on padded inputs
if border_mode == 'valid':
padded_inputs = inputs
else:
if border_mode == 'full':
pad_per_dim = [filters_shape[i] - 1 for i in range(2,5)]
else:
pad_per_dim = border_mode
run_conv3d_fwd(inputs_shape=(128, 3, 5, 5, 5), pad_before_after = ([(0, 0), (0, 0)] +
filters_shape=(64, 3, 1, 2, 4)) [(p, p) for p in pad_per_dim])
run_conv3d_fwd(inputs_shape=(16, 4, 20, 12, 15), padded_inputs_val = numpy.pad(inputs_val, pad_before_after,
filters_shape=(10, 4, 6, 12, 4), 'constant')
subsample=(2, 2, 2)) padded_inputs = shared(padded_inputs_val)
run_conv3d_fwd(inputs_shape=(16, 4, 20, 12, 15),
filters_shape=(10, 4, 6, 12, 4),
subsample=(2, 2, 2))
run_conv3d_fwd(inputs_shape=(16, 1, 20, 12, 15),
filters_shape=(10, 1, 6, 12, 4),
subsample=(3, 3, 3))
run_conv3d_fwd(inputs_shape=(16, 2, 20, 12, 15),
filters_shape=(10, 2, 6, 12, 4),
subsample=(3, 3, 3))
run_conv3d_fwd(inputs_shape=(16, 1, 20, 12, 15),
filters_shape=(10, 1, 6, 12, 4),
subsample=(3, 2, 1))
run_conv3d_fwd(inputs_shape=(16, 1, 20, 12, 15),
filters_shape=(10, 1, 6, 12, 4),
subsample=(1, 2, 3))
def test_conv3d_gradweight():
if not cuda.dnn.dnn_available() and dnn.version()[0] >= 3000: # Compile a theano function for the reference implementation
raise SkipTest('"3D conv not supported in cudnn v1') conv_ref = theano.tensor.nnet.conv3D(V=padded_inputs.dimshuffle(0, 2, 3, 4, 1),
W=flipped_filters.dimshuffle(0, 2, 3, 4, 1),
b=bias, d=subsample)
f_ref = theano.function([], conv_ref.dimshuffle(0, 4, 1, 2, 3))
def run_gradweight(inputs_shape, filters_shape, dCdH_shape, # Compare the results of the two implementations
subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
kern_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
kern = shared(kern_val)
filters_shape_s = (filters_shape[0], filters_shape[2],
filters_shape[3], filters_shape[4],
filters_shape[1])
conv = theano.tensor.nnet.convGrad3D(V=inputs.dimshuffle(0, 2, 3, 4, 1),
dCdH=dCdH.dimshuffle(0, 2, 3, 4, 1),
WShape=filters_shape_s,
d=subsample)
desc = dnn.GpuDnnConvDesc(border_mode='valid', subsample=subsample,
conv_mode='cross')(inputs.shape, kern.shape)
gradW = dnn.GpuDnnConv3dGradW()(inputs, dCdH, kern, desc)
f_ref = theano.function([], conv.dimshuffle(0, 4, 1, 2, 3))
f = theano.function([], gradW, mode=mode_with_gpu)
res_ref = f_ref() res_ref = f_ref()
res = f() res = f()
utt.assert_allclose(res_ref, res) utt.assert_allclose(res_ref, res)
run_gradweight(inputs_shape=(16, 1, 10, 12, 16), test_cases = get_conv3d_test_cases()
filters_shape=(10, 1, 6, 12, 4), for (i_shape, f_shape, subsample), border_mode, conv_mode in test_cases:
dCdH_shape=(16, 10, 5, 1, 13), yield (run_conv3d_fwd, i_shape, f_shape, subsample, border_mode,
subsample=(1, 1, 1)) conv_mode)
run_gradweight(inputs_shape=(16, 1, 20, 10, 16),
filters_shape=(10, 1, 6, 4, 4),
dCdH_shape=(16, 10, 8, 4, 7), def test_conv3d_bwd():
subsample=(2, 2, 2))
run_gradweight(inputs_shape=(16, 1, 20, 10, 16),
filters_shape=(10, 1, 6, 3, 4),
dCdH_shape=(16, 10, 5, 3, 5),
subsample=(3, 3, 3))
run_gradweight(inputs_shape=(16, 1, 20, 12, 16),
filters_shape=(10, 1, 6, 12, 4),
dCdH_shape=(16, 10, 8, 1, 5),
subsample=(2, 1, 3))
def test_conv3d_gradinput():
if not cuda.dnn.dnn_available() and dnn.version()[0] >= 3000: if not cuda.dnn.dnn_available() and dnn.version()[0] >= 3000:
raise SkipTest('"3D conv not supported in cudnn v1') raise SkipTest('"CuDNN 3D convolution requires CuDNN v3')
def run_gradinput(inputs_shape, filters_shape, def run_conv3d_bwd(inputs_shape, filters_shape, subsample,
subsample=(1, 1, 1)): border_mode, conv_mode):
inputs_val = numpy.random.random(inputs_shape).astype('float32') inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32') filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val) inputs = shared(inputs_val)
filters = shared(filters_val) filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
bias = shared(numpy.zeros(filters_shape[1]).astype('float32')) # Compile a theano function for the CuDNN implementation
conv = theano.tensor.nnet.convTransp3D(W=filters.dimshuffle(0, 2, 3, 4, 1), conv = dnn.dnn_conv3d(img=inputs, kerns=filters,
b=bias, d=subsample, border_mode=border_mode, subsample=subsample,
H=inputs.dimshuffle(0, 2, 3, 4, 1)) conv_mode=conv_mode)
f_ref = theano.function([], conv.dimshuffle(0, 4, 1, 2, 3)) grad_i, grad_w = theano.tensor.grad(conv.sum(), [inputs, filters])
res_ref = f_ref()
bottom_shape = res_ref.shape f = theano.function([], [grad_i, grad_w], mode=mode_with_gpu)
bottom_val = numpy.random.random(bottom_shape).astype('float32')
bottom = shared(bottom_val)
desc = dnn.GpuDnnConvDesc(border_mode='valid', subsample=subsample, # If conv_mode is 'conv' the reference implementation should use
conv_mode='cross')(bottom.shape, filters.shape) # filters filpped according to the width, height and time axis
gradI = dnn.GpuDnnConv3dGradI()(filters, inputs, bottom, desc) if conv_mode == 'conv':
f = theano.function([], gradI, mode=mode_with_gpu) flipped_filters = filters[:,:,::-1,::-1,::-1]
res = f() else:
flipped_filters = filters
utt.assert_allclose(res_ref, res) # If border mode is anything but 'valid', the reference implementation
# should operate on padded inputs
if border_mode == 'valid':
padded_inputs = inputs
else:
if border_mode == 'full':
pad_per_dim = [filters_shape[i] - 1 for i in range(2,5)]
else:
pad_per_dim = border_mode
pad_before_after = ([(0, 0), (0, 0)] +
[(p, p) for p in pad_per_dim])
padded_inputs_val = numpy.pad(inputs_val, pad_before_after,
'constant')
padded_inputs = shared(padded_inputs_val)
# Compile a theano function for the reference implementation
conv_ref = theano.tensor.nnet.conv3D(V=padded_inputs.dimshuffle(0, 2, 3, 4, 1),
W=flipped_filters.dimshuffle(0, 2, 3, 4, 1),
b=bias, d=subsample)
(grad_padded_i_ref,
grad_w_ref) = theano.tensor.grad(conv_ref.sum(), [padded_inputs, filters])
# Recover grad_i_ref from grad_padded_i_ref
if border_mode == 'valid':
grad_i_ref = grad_padded_i_ref
else:
shp = grad_padded_i_ref.shape
grad_i_ref = grad_padded_i_ref[:, :,
pad_per_dim[0]:shp[2] - pad_per_dim[0],
pad_per_dim[1]:shp[3] - pad_per_dim[1],
pad_per_dim[2]:shp[4] - pad_per_dim[2]]
f_ref = theano.function([], [grad_i_ref, grad_w_ref])
# Compare the results of the two implementations
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref[0], res[0])
utt.assert_allclose(res_ref[1], res[1])
run_gradinput(inputs_shape=(16, 10, 15, 12, 12), test_cases = get_conv3d_test_cases()
filters_shape=(10, 1, 6, 12, 4)) for (i_shape, f_shape, subsample), border_mode, conv_mode in test_cases:
run_gradinput(inputs_shape=(16, 10, 15, 12, 12), yield (run_conv3d_bwd, i_shape, f_shape, subsample, border_mode,
filters_shape=(10, 1, 6, 12, 4), conv_mode)
subsample=(2, 2, 2))
run_gradinput(inputs_shape=(16, 10, 15, 12, 12),
filters_shape=(10, 1, 6, 12, 4),
subsample=(3, 3, 3))
run_gradinput(inputs_shape=(16, 10, 15, 12, 12),
filters_shape=(10, 1, 6, 12, 4),
subsample=(3, 1, 2))
def test_version(): def test_version():
......
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