Skip to content

P
pytensor
提交 3ce8f1cb authored 作者: Frédéric Bastien's avatar Frédéric Bastien 提交者： GitHub
浏览文件
操作

Merge pull request #6362 from affanv14/3dmetafix

fix 3d optimizers for grouped convolutions and add more test cases
上级 926dec89 45f9b15a
显示空白字符变更
内嵌 并排
正在显示 包含 223 行增加 和 20 行删除
theano/gpuarray/dnn.py
浏览文件 @ 3ce8f1cb
...@@ -3258,6 +3258,7 @@ def local_abstractconv3d_cudnn_alt(node): ...@@ -3258,6 +3258,7 @@ def local_abstractconv3d_cudnn_alt(node):
border_mode = node.op.border_mode border_mode = node.op.border_mode
subsample = node.op.subsample subsample = node.op.subsample
filter_dilation = node.op.filter_dilation filter_dilation = node.op.filter_dilation
num_groups = node.op.num_groups
precision = get_precision(None, [inp1, inp2]) precision = get_precision(None, [inp1, inp2])
if node.op.filter_flip: if node.op.filter_flip:
...@@ -3266,7 +3267,7 @@ def local_abstractconv3d_cudnn_alt(node): ...@@ -3266,7 +3267,7 @@ def local_abstractconv3d_cudnn_alt(node):
conv_mode = 'cross' conv_mode = 'cross'
if isinstance(op, AbstractConv3d): if isinstance(op, AbstractConv3d):
if border_mode == 'half' or subsample != (1, 1, 1): if border_mode == 'half' or subsample != (1, 1, 1) or num_groups > 1:
return None return None
if border_mode == 'full': if border_mode == 'full':
direction_hint = 'bprop inputs' direction_hint = 'bprop inputs'
...@@ -3284,7 +3285,7 @@ def local_abstractconv3d_cudnn_alt(node): ...@@ -3284,7 +3285,7 @@ def local_abstractconv3d_cudnn_alt(node):
elif isinstance(op, AbstractConv3d_gradWeights): elif isinstance(op, AbstractConv3d_gradWeights):
if(border_mode == 'valid' and subsample == (1, 1, 1) and if(border_mode == 'valid' and subsample == (1, 1, 1) and
filter_dilation == (1, 1, 1)): filter_dilation == (1, 1, 1) and num_groups == 1):
img = gpu_contiguous(inp1) img = gpu_contiguous(inp1)
topgrad = gpu_contiguous(inp2) topgrad = gpu_contiguous(inp2)
ctx_name = infer_context_name(img, topgrad) ctx_name = infer_context_name(img, topgrad)
...@@ -3315,7 +3316,7 @@ def local_abstractconv3d_cudnn_alt(node): ...@@ -3315,7 +3316,7 @@ def local_abstractconv3d_cudnn_alt(node):
return None return None
elif isinstance(op, AbstractConv3d_gradInputs): elif isinstance(op, AbstractConv3d_gradInputs):
if border_mode == 'valid' and subsample == (1, 1, 1): if border_mode == 'valid' and subsample == (1, 1, 1) and num_groups == 1:
kerns = gpu_contiguous(inp1.dimshuffle(1, 0, 2, 3, 4)) kerns = gpu_contiguous(inp1.dimshuffle(1, 0, 2, 3, 4))
topgrad = gpu_contiguous(inp2) topgrad = gpu_contiguous(inp2)
ctx_name = infer_context_name(kerns, topgrad) ctx_name = infer_context_name(kerns, topgrad)
......
theano/gpuarray/opt.py
浏览文件 @ 3ce8f1cb
...@@ -1842,8 +1842,10 @@ def local_abstractconv3d_alt(node): ...@@ -1842,8 +1842,10 @@ def local_abstractconv3d_alt(node):
border_mode = node.op.border_mode border_mode = node.op.border_mode
subsample = node.op.subsample subsample = node.op.subsample
filter_dilation = node.op.filter_dilation filter_dilation = node.op.filter_dilation
num_groups = node.op.num_groups
if ((border_mode == 'full') and (subsample == (1, 1, 1))): if((border_mode == 'full') and (subsample == (1, 1, 1)) and
(num_groups == 1)):
if not node.op.filter_flip: if not node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1, ::-1] kern = kern[:, :, ::-1, ::-1, ::-1]
kern = kern.dimshuffle(1, 0, 2, 3, 4) kern = kern.dimshuffle(1, 0, 2, 3, 4)
...@@ -1853,7 +1855,7 @@ def local_abstractconv3d_alt(node): ...@@ -1853,7 +1855,7 @@ def local_abstractconv3d_alt(node):
gpu_contiguous(kern), gpu_contiguous(img)) gpu_contiguous(kern), gpu_contiguous(img))
elif(subsample == (1, 1, 1) and filter_dilation == (1, 1, 1) and elif(subsample == (1, 1, 1) and filter_dilation == (1, 1, 1) and
border_mode == 'valid'): border_mode == 'valid' and num_groups == 1):
if node.op.filter_flip: if node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1, ::-1] kern = kern[:, :, ::-1, ::-1, ::-1]
rval = GpuCorr3dMM_gradWeights(border_mode, rval = GpuCorr3dMM_gradWeights(border_mode,
...@@ -1881,8 +1883,10 @@ def local_abstractconv3d2d(node): ...@@ -1881,8 +1883,10 @@ def local_abstractconv3d2d(node):
border_mode = node.op.border_mode border_mode = node.op.border_mode
subsample = node.op.subsample subsample = node.op.subsample
filter_dilation = node.op.filter_dilation filter_dilation = node.op.filter_dilation
num_groups = node.op.num_groups
if subsample == (1, 1, 1) and filter_dilation == (1, 1, 1): if(subsample == (1, 1, 1) and filter_dilation == (1, 1, 1) and
num_groups == 1):
reorder_array = [0, 2, 1, 3, 4] reorder_array = [0, 2, 1, 3, 4]
rval = conv3d2d.conv3d(gpu_contiguous(img.dimshuffle(*reorder_array)), rval = conv3d2d.conv3d(gpu_contiguous(img.dimshuffle(*reorder_array)),
gpu_contiguous(kern.dimshuffle(*reorder_array)), gpu_contiguous(kern.dimshuffle(*reorder_array)),
...@@ -1968,8 +1972,10 @@ def local_abstractconv3d_gemm_gradweights_alt(node): ...@@ -1968,8 +1972,10 @@ def local_abstractconv3d_gemm_gradweights_alt(node):
border_mode = node.op.border_mode border_mode = node.op.border_mode
subsample = node.op.subsample subsample = node.op.subsample
filter_dilation = node.op.filter_dilation filter_dilation = node.op.filter_dilation
num_groups = node.op.num_groups
if border_mode == 'valid' and subsample == (1, 1, 1) and filter_dilation == (1, 1, 1): if(border_mode == 'valid' and subsample == (1, 1, 1) and
filter_dilation == (1, 1, 1) and num_groups == 1):
rval = GpuCorr3dMM(border_mode, rval = GpuCorr3dMM(border_mode,
subsample, subsample,
filter_dilation)( filter_dilation)(
...@@ -2091,8 +2097,10 @@ def local_abstractconv3d_gradinputs_gemm_alt(node): ...@@ -2091,8 +2097,10 @@ def local_abstractconv3d_gradinputs_gemm_alt(node):
border_mode = node.op.border_mode border_mode = node.op.border_mode
subsample = node.op.subsample subsample = node.op.subsample
filter_dilation = node.op.filter_dilation filter_dilation = node.op.filter_dilation
num_groups = node.op.num_groups
if border_mode == 'valid' and subsample == (1, 1, 1): if(border_mode == 'valid' and subsample == (1, 1, 1) and
num_groups == 1):
if not node.op.filter_flip: if not node.op.filter_flip:
kern = kern[:, :, ::-1, ::-1, ::-1] kern = kern[:, :, ::-1, ::-1, ::-1]
rval = GpuCorr3dMM(border_mode='full', rval = GpuCorr3dMM(border_mode='full',
......
theano/gpuarray/tests/test_opt.py
浏览文件 @ 3ce8f1cb
...@@ -709,7 +709,8 @@ def test_crossentropycategorical1hot_lifter(): ...@@ -709,7 +709,8 @@ def test_crossentropycategorical1hot_lifter():
class Conv_opt_test(unittest.TestCase): class Conv_opt_test(unittest.TestCase):
def optimizer_2d(self, input_shapes, direction, include_tags, exclude_tags, def optimizer_2d(self, input_shapes, direction, include_tags, exclude_tags,
op, border_mode='valid', subsample=(1, 1), filter_dilation=(1, 1)): op, border_mode='valid', subsample=(1, 1),
filter_dilation=(1, 1), num_groups=1):
inp1 = theano.shared(np.random.random(input_shapes[0]).astype(theano.config.floatX)) inp1 = theano.shared(np.random.random(input_shapes[0]).astype(theano.config.floatX))
inp2 = theano.shared(np.random.random(input_shapes[1]).astype(theano.config.floatX)) inp2 = theano.shared(np.random.random(input_shapes[1]).astype(theano.config.floatX))
...@@ -720,7 +721,8 @@ class Conv_opt_test(unittest.TestCase): ...@@ -720,7 +721,8 @@ class Conv_opt_test(unittest.TestCase):
input_shapes[1], input_shapes[1],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation,
num_groups=num_groups)
if(direction == 1): if(direction == 1):
conv_op = abstract_conv.conv2d_grad_wrt_weights(inp1, conv_op = abstract_conv.conv2d_grad_wrt_weights(inp1,
...@@ -729,7 +731,8 @@ class Conv_opt_test(unittest.TestCase): ...@@ -729,7 +731,8 @@ class Conv_opt_test(unittest.TestCase):
input_shapes[0], input_shapes[0],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation,
num_groups=num_groups)
if(direction == 2): if(direction == 2):
conv_op = abstract_conv.conv2d_grad_wrt_inputs(inp1, conv_op = abstract_conv.conv2d_grad_wrt_inputs(inp1,
...@@ -738,16 +741,24 @@ class Conv_opt_test(unittest.TestCase): ...@@ -738,16 +741,24 @@ class Conv_opt_test(unittest.TestCase):
input_shapes[1], input_shapes[1],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation,
num_groups=num_groups)
theano.config.metaopt.optimizer_including = include_tags theano.config.metaopt.optimizer_including = include_tags
theano.config.metaopt.optimizer_excluding = exclude_tags theano.config.metaopt.optimizer_excluding = exclude_tags
mode = mode_with_gpu.including('conv_meta') mode = mode_with_gpu.including('conv_meta').excluding('conv_dnn').excluding('conv_gemm')
ref_func = theano.function([], conv_op, mode=mode_with_gpu) ref_func = theano.function([], conv_op, mode=mode_with_gpu)
# All meta optimizer compile a new function. This need to know # All meta optimizer compile a new function. This need to know
# the current linker, but this information is not available, # the current linker, but this information is not available,
# so it use the default mode. # so it use the default mode.
if op is None:
# No convolutions optimization takes place
with theano.change_flags(mode=mode):
with self.assertRaises(AssertionError):
theano.function([], conv_op, mode=mode)
return
else:
with theano.change_flags(mode=mode): with theano.change_flags(mode=mode):
conv_func = theano.function([], conv_op, mode=mode) conv_func = theano.function([], conv_op, mode=mode)
assert any([isinstance(node.op, op) assert any([isinstance(node.op, op)
...@@ -756,7 +767,7 @@ class Conv_opt_test(unittest.TestCase): ...@@ -756,7 +767,7 @@ class Conv_opt_test(unittest.TestCase):
def optimizer_3d(self, input_shapes, direction, include_tags, exclude_tags, def optimizer_3d(self, input_shapes, direction, include_tags, exclude_tags,
op, border_mode='valid', subsample=(1, 1, 1), op, border_mode='valid', subsample=(1, 1, 1),
filter_dilation=(1, 1, 1)): filter_dilation=(1, 1, 1), num_groups=1):
inp1 = theano.shared(np.random.random(input_shapes[0]).astype(theano.config.floatX)) inp1 = theano.shared(np.random.random(input_shapes[0]).astype(theano.config.floatX))
inp2 = theano.shared(np.random.random(input_shapes[1]).astype(theano.config.floatX)) inp2 = theano.shared(np.random.random(input_shapes[1]).astype(theano.config.floatX))
if(direction == 0): if(direction == 0):
...@@ -766,7 +777,8 @@ class Conv_opt_test(unittest.TestCase): ...@@ -766,7 +777,8 @@ class Conv_opt_test(unittest.TestCase):
input_shapes[1], input_shapes[1],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation,
num_groups=num_groups)
if(direction == 1): if(direction == 1):
conv_op = abstract_conv.conv3d_grad_wrt_weights(inp1, conv_op = abstract_conv.conv3d_grad_wrt_weights(inp1,
...@@ -775,7 +787,8 @@ class Conv_opt_test(unittest.TestCase): ...@@ -775,7 +787,8 @@ class Conv_opt_test(unittest.TestCase):
input_shapes[0], input_shapes[0],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation,
num_groups=num_groups)
if(direction == 2): if(direction == 2):
conv_op = abstract_conv.conv3d_grad_wrt_inputs(inp1, conv_op = abstract_conv.conv3d_grad_wrt_inputs(inp1,
...@@ -784,21 +797,34 @@ class Conv_opt_test(unittest.TestCase): ...@@ -784,21 +797,34 @@ class Conv_opt_test(unittest.TestCase):
input_shapes[1], input_shapes[1],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation,
num_groups=num_groups)
theano.config.metaopt.optimizer_including = include_tags theano.config.metaopt.optimizer_including = include_tags
theano.config.metaopt.optimizer_excluding = exclude_tags theano.config.metaopt.optimizer_excluding = exclude_tags
mode = mode_with_gpu.including('conv_meta') mode = mode_with_gpu.including('conv_meta').excluding('conv_dnn').excluding('conv_gemm')
ref_func = theano.function([], conv_op, mode=mode_with_gpu) ref_func = theano.function([], conv_op, mode=mode_with_gpu)
# All meta optimizer compile a new function. This need to know # All meta optimizer compile a new function. This need to know
# the current linker, but this information is not available, # the current linker, but this information is not available,
# so it use the default mode. # so it use the default mode.
if op is None:
# No convolutions optimization takes place
with theano.change_flags(mode=mode):
with self.assertRaises(AssertionError):
theano.function([], conv_op, mode=mode)
return
elif op != 'conv3d2d':
with theano.change_flags(mode=mode): with theano.change_flags(mode=mode):
conv_func = theano.function([], conv_op, mode=mode) conv_func = theano.function([], conv_op, mode=mode)
if op is not None:
assert any([isinstance(node.op, op) assert any([isinstance(node.op, op)
for node in conv_func.maker.fgraph.toposort()]) for node in conv_func.maker.fgraph.toposort()])
else:
with theano.change_flags(mode=mode):
conv_func = theano.function(
[], conv_op,
mode=mode_with_gpu.including('conv_meta'))
utt.assert_allclose(conv_func(), ref_func()) utt.assert_allclose(conv_func(), ref_func())
def test_optimizers_2d(self): def test_optimizers_2d(self):
...@@ -883,7 +909,7 @@ class Conv_opt_test(unittest.TestCase): ...@@ -883,7 +909,7 @@ class Conv_opt_test(unittest.TestCase):
self.optimizer_3d([imshp, kshp, tshp], 0, self.optimizer_3d([imshp, kshp, tshp], 0,
'conv3d2d', 'conv3d2d',
'default', 'default',
None) 'conv3d2d')
self.optimizer_3d([imshp, kshp, tshp], 0, self.optimizer_3d([imshp, kshp, tshp], 0,
'alternative', 'alternative',
'conv_gemm:default:conv3d2d', 'conv_gemm:default:conv3d2d',
...@@ -989,3 +1015,171 @@ class Conv_opt_test(unittest.TestCase): ...@@ -989,3 +1015,171 @@ class Conv_opt_test(unittest.TestCase):
dnn.GpuDnnConvGradI, dnn.GpuDnnConvGradI,
border_mode='full', border_mode='full',
filter_dilation=fdil) filter_dilation=fdil)
# test non default num_groups for default optimizers
imshp2d = [(2, 6, 5, 5), (2, 4, 5, 5)]
kshp2d = [(3, 2, 3, 3), (2, 2, 3, 3)]
tshp2d = [(2, 3, 3, 3), (2, 2, 3, 3)]
num_groups = [3, 2]
for imshp, kshp, tshp, groups in zip(imshp2d, kshp2d, tshp2d, num_groups):
# forward pass
self.optimizer_2d([imshp, kshp, tshp], 0,
'',
'conv_dnn:alternative',
blas.GpuCorrMM,
num_groups=groups)
self.optimizer_2d([imshp, kshp, tshp], 0,
'',
'conv_gemm:alternative',
dnn.GpuDnnConv,
num_groups=groups)
# grad with respect to weights
self.optimizer_2d([imshp, tshp, kshp], 1,
'',
'conv_dnn:alternative',
blas.GpuCorrMM_gradWeights,
num_groups=groups)
self.optimizer_2d([imshp, tshp, kshp], 1,
'',
'conv_gemm:alternative',
dnn.GpuDnnConvGradW,
num_groups=groups)
# grad with respect to inputs
self.optimizer_2d([tshp, kshp, imshp], 2,
'',
'conv_dnn:alternative',
blas.GpuCorrMM_gradInputs,
num_groups=groups)
self.optimizer_2d([tshp, kshp, imshp], 2,
'',
'conv_gemm:alternative',
dnn.GpuDnnConvGradI,
num_groups=groups)
imshp3d = [(2, 6, 5, 5, 5), (2, 4, 5, 5, 5)]
kshp3d = [(3, 2, 3, 3, 3), (2, 2, 3, 3, 3)]
tshp3d = [(2, 3, 3, 3, 3), (2, 2, 3, 3, 3)]
num_groups = [3, 2]
for imshp, kshp, tshp, groups in zip(imshp3d, kshp3d, tshp3d, num_groups):
# forward pass
self.optimizer_3d([imshp, kshp, tshp], 0,
'',
'conv_dnn:alternative:conv3d2d',
blas.GpuCorr3dMM,
num_groups=groups)
self.optimizer_3d([imshp, kshp, tshp], 0,
'',
'conv_gemm:alternative:conv3d2d',
dnn.GpuDnnConv,
num_groups=groups)
# grad with respect to weights
self.optimizer_3d([imshp, tshp, kshp], 1,
'',
'conv_dnn:alternative:conv3d2d',
blas.GpuCorr3dMM_gradWeights,
num_groups=groups)
self.optimizer_3d([imshp, tshp, kshp], 1,
'',
'conv_gemm:alternative:conv3d2d',
dnn.GpuDnnConvGradW,
num_groups=groups)
# grad with respect to inputs
self.optimizer_3d([tshp, kshp, imshp], 2,
'',
'conv_dnn:alternative:conv3d2d',
blas.GpuCorr3dMM_gradInputs,
num_groups=groups)
self.optimizer_3d([tshp, kshp, imshp], 2,
'',
'conv_gemm:alternative:conv3d2d',
dnn.GpuDnnConvGradI,
num_groups=groups)
def test_returns_none(self):
if theano.config.cxx == "":
raise SkipTest("Need a c compiler.")
# values given dont matter since it returns None
imshp = (2, 3, 5, 5)
kshp = (4, 3, 3, 3)
tshp = (2, 4, 3, 3)
exclude_string = ['conv_dnn:default', 'conv_gemm:default']
conv_direction = [0, 1, 2]
# test that non default subsample returns None
for string in exclude_string:
for direction in conv_direction:
self.optimizer_2d([imshp, kshp, tshp],
direction,
'alternative',
string,
None,
subsample=(2, 2))
# test that non default num_groups returns None
for string in exclude_string:
for direction in conv_direction:
self.optimizer_2d([imshp, kshp, tshp],
direction,
'alternative',
string,
None,
num_groups=3)
# test that border_mode=half returns None
for string in exclude_string:
for direction in conv_direction:
self.optimizer_2d([imshp, kshp, tshp],
direction,
'alternative',
string,
None,
border_mode='half')
# test that Non-default filter dilation return None for
# direction 1
for string in exclude_string:
direction = 1
self.optimizer_2d([imshp, kshp, tshp],
direction,
'alternative',
'conv_dnn:default',
None,
filter_dilation=(2, 2))
imshp = (2, 3, 5, 5, 5)
kshp = (4, 3, 3, 3, 3)
tshp = (2, 4, 3, 3, 3)
exclude_string = ['conv_dnn:default', 'conv_gemm:default']
# test that non default subsample returns None
for string in exclude_string:
for direction in conv_direction:
self.optimizer_3d([imshp, kshp, tshp],
direction,
'alternative',
string,
None,
subsample=(2, 2, 2))
# test that non default num_groups returns None
for string in exclude_string:
for direction in conv_direction:
self.optimizer_3d([imshp, kshp, tshp],
direction,
'alternative',
string,
None,
num_groups=3)
# test that border_mode=half returns None
for string in exclude_string:
for direction in conv_direction:
self.optimizer_3d([imshp, kshp, tshp],
direction,
'alternative',
string,
None,
border_mode='half')
# test that Non-default filter dilation return None for
# direction 1
for string in exclude_string:
direction = 1
self.optimizer_3d([imshp, kshp, tshp],
direction,
'alternative',
string,
None,
filter_dilation=(2, 2, 2))
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