Skip to content

P
pytensor
提交 1a919959 authored 作者: affanv14's avatar affanv14
浏览文件
操作

fix tests

上级 04750bbf
隐藏空白字符变更
内嵌 并排
正在显示 包含 112 行增加 和 109 行删除
theano/gpuarray/tests/test_opt.py
浏览文件 @ 1a919959
...@@ -23,29 +23,8 @@ from ..linalg import GpuCusolverSolve, cusolver_available, GpuCholesky ...@@ -23,29 +23,8 @@ from ..linalg import GpuCusolverSolve, cusolver_available, GpuCholesky
from .config import mode_with_gpu, mode_without_gpu, test_ctx_name, SkipTest from .config import mode_with_gpu, mode_without_gpu, test_ctx_name, SkipTest
import unittest import unittest
from theano.tensor.nnet.abstract_conv import (AbstractConv2d, from theano.tensor.nnet import abstract_conv
AbstractConv2d_gradInputs, from theano.gpuarray import dnn, blas
AbstractConv2d_gradWeights,
AbstractConv3d,
AbstractConv3d_gradWeights,
AbstractConv3d_gradInputs,
conv2d,
conv2d_grad_wrt_weights,
conv2d_grad_wrt_inputs,
conv3d,
conv3d_grad_wrt_weights,
conv3d_grad_wrt_inputs)
from theano.gpuarray.opt import (local_abstractconv_gemm_alternative,
local_abstractconv_gemm_gradweights_alt,
local_abstractconv_gradinputs_gemm_alt,
local_abstractconv_cudnn_alternative,
local_abstractconv3d2d,
local_abstractconv3d_alt,
local_abstractconv3d_gemm_gradweights_alt,
local_abstractconv3d_gradinputs_gemm_alt,
local_abstractconv3d_cudnn_alternative,
local_conv_gpu_conv)
def test_local_assert(): def test_local_assert():
...@@ -727,96 +706,86 @@ def test_crossentropycategorical1hot_lifter(): ...@@ -727,96 +706,86 @@ def test_crossentropycategorical1hot_lifter():
class Conv_opt_test(unittest.TestCase): class Conv_opt_test(unittest.TestCase):
def optimizer_2d(self, input_shapes, direction, optimizer, border_mode='valid', def optimizer_2d(self, input_shapes, direction, include_tags, exclude_tags,
subsample=(1, 1), filter_dilation=(1, 1)): op, border_mode='valid', subsample=(1, 1), filter_dilation=(1, 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):
abstract_op = AbstractConv2d conv_op = abstract_conv.conv2d(inp1,
conv_op = conv2d(inp1, inp2,
inp2, border_mode=border_mode,
border_mode=border_mode, subsample=subsample,
subsample=subsample, filter_dilation=filter_dilation)
filter_dilation=filter_dilation)
if(direction == 1): if(direction == 1):
abstract_op = AbstractConv2d_gradWeights conv_op = abstract_conv.conv2d_grad_wrt_weights(inp1,
conv_op = conv2d_grad_wrt_weights(inp1, inp2,
inp2, input_shapes[2],
input_shapes[2], input_shapes[0],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation)
if(direction == 2): if(direction == 2):
abstract_op = AbstractConv2d_gradInputs conv_op = abstract_conv.conv2d_grad_wrt_inputs(inp1,
conv_op = conv2d_grad_wrt_inputs(inp1, inp2,
inp2, input_shapes[2],
input_shapes[2], input_shapes[1],
border_mode=border_mode, border_mode=border_mode,
subsample=subsample, subsample=subsample,
filter_dilation=filter_dilation) filter_dilation=filter_dilation)
theano.config.metaopt.optimizer_including = include_tags
theano.config.metaopt.optimizer_excluding = exclude_tags
mode = theano.Mode().including('conv_meta')
ref_func = theano.function([], conv_op) ref_func = theano.function([], conv_op)
conv_node = conv_op.owner conv_func = theano.function([], conv_op, mode=mode)
assert any([isinstance(node.op, op)
if isinstance(conv_node.op, abstract_op): for node in conv_func.maker.fgraph.toposort()])
conv_op = local_conv_gpu_conv.transform(conv_node)
if isinstance(conv_node.op, abstract_op):
conv_op = optimizer.transform(conv_op[0].owner.inputs[0].owner)
conv_func = theano.function([], conv_op[0])
assert not any([isinstance(node.op, abstract_op)
for node in conv_func.maker.fgraph.toposort()])
utt.assert_allclose(conv_func(), ref_func()) utt.assert_allclose(conv_func(), ref_func())
def optimizer_3d(self, input_shapes, direction, optimizer, border_mode='valid', def optimizer_3d(self, input_shapes, direction, include_tags, exclude_tags,
subsample=(1, 1, 1), filter_dilation=(1, 1, 1)): op, border_mode='valid', subsample=(1, 1, 1),
filter_dilation=(1, 1, 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):
abstract_op = AbstractConv3d conv_op = abstract_conv.conv3d(inp1,
conv_op = conv3d(inp1, inp2,
inp2, input_shapes[0],
input_shapes[0], 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)
if(direction == 1): if(direction == 1):
abstract_op = AbstractConv3d_gradWeights conv_op = abstract_conv.conv3d_grad_wrt_weights(inp1,
conv_op = conv3d_grad_wrt_weights(inp1, inp2,
inp2, input_shapes[2],
input_shapes[2], 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)
if(direction == 2): if(direction == 2):
abstract_op = AbstractConv3d_gradInputs conv_op = abstract_conv.conv3d_grad_wrt_inputs(inp1,
conv_op = conv3d_grad_wrt_inputs(inp1, inp2,
inp2, input_shapes[2],
input_shapes[2], 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)
theano.config.metaopt.optimizer_including = include_tags
theano.config.metaopt.optimizer_excluding = exclude_tags
mode = theano.Mode().including('conv_meta')
ref_func = theano.function([], conv_op) ref_func = theano.function([], conv_op)
conv_node = conv_op.owner conv_func = theano.function([], conv_op, mode=mode)
if op is not None:
if isinstance(conv_node.op, abstract_op): assert any([isinstance(node.op, op)
conv_op = local_conv_gpu_conv.transform(conv_node)
if isinstance(conv_node.op, abstract_op):
conv_op = optimizer.transform(conv_op[0].owner.inputs[0].owner)
conv_func = theano.function([], conv_op[0])
assert not any([isinstance(node.op, abstract_op)
for node in conv_func.maker.fgraph.toposort()]) for node in conv_func.maker.fgraph.toposort()])
utt.assert_allclose(conv_func(), ref_func()) utt.assert_allclose(conv_func(), ref_func())
...@@ -828,19 +797,31 @@ class Conv_opt_test(unittest.TestCase): ...@@ -828,19 +797,31 @@ class Conv_opt_test(unittest.TestCase):
for imshp, kshp, tshp in zip(imshp2d, kshp2d, tshp2d): for imshp, kshp, tshp in zip(imshp2d, kshp2d, tshp2d):
# forward passes # forward passes
self.optimizer_2d([imshp, kshp, tshp], 0, self.optimizer_2d([imshp, kshp, tshp], 0,
local_abstractconv_gemm_alternative) 'alternative',
'conv_dnn:default',
blas.GpuCorrMM_gradWeights)
self.optimizer_2d([imshp, kshp, tshp], 0, self.optimizer_2d([imshp, kshp, tshp], 0,
local_abstractconv_cudnn_alternative) 'alternative',
'conv_gemm:default',
dnn.GpuDnnConvGradW)
# backwards wrt weights # backwards wrt weights
self.optimizer_2d([imshp, tshp, kshp], 1, self.optimizer_2d([imshp, tshp, kshp], 1,
local_abstractconv_gemm_gradweights_alt) 'alternative',
'conv_dnn:default',
blas.GpuCorrMM)
self.optimizer_2d([imshp, tshp, kshp], 1, self.optimizer_2d([imshp, tshp, kshp], 1,
local_abstractconv_cudnn_alternative) 'alternative',
'conv_gemm:default',
dnn.GpuDnnConv)
# backwards wrt to inputs # backwards wrt to inputs
self.optimizer_2d([tshp, kshp, imshp], 2, self.optimizer_2d([tshp, kshp, imshp], 2,
local_abstractconv_gradinputs_gemm_alt) 'alternative',
'conv_dnn:default',
blas.GpuCorrMM)
self.optimizer_2d([tshp, kshp, imshp], 2, self.optimizer_2d([tshp, kshp, imshp], 2,
local_abstractconv_cudnn_alternative) 'alternative',
'conv_gemm:default',
dnn.GpuDnnConv)
imshp3d = [(2, 3, 5, 5, 5), (2, 2, 5, 7, 5), (2, 1, 3, 3, 3)] imshp3d = [(2, 3, 5, 5, 5), (2, 2, 5, 7, 5), (2, 1, 3, 3, 3)]
kshp3d = [(4, 3, 3, 3, 3), (3, 2, 3, 5, 3), (4, 1, 1, 1, 1)] kshp3d = [(4, 3, 3, 3, 3), (3, 2, 3, 5, 3), (4, 1, 1, 1, 1)]
...@@ -849,22 +830,36 @@ class Conv_opt_test(unittest.TestCase): ...@@ -849,22 +830,36 @@ class Conv_opt_test(unittest.TestCase):
for imshp, kshp, tshp in zip(imshp3d, kshp3d, tshp3d): for imshp, kshp, tshp in zip(imshp3d, kshp3d, tshp3d):
# forwards passes # forwards passes
self.optimizer_3d([imshp, kshp, tshp], 0, self.optimizer_3d([imshp, kshp, tshp], 0,
local_abstractconv3d_alt) 'alternative',
'conv_dnn:default:conv3d2d',
blas.GpuCorr3dMM_gradWeights)
self.optimizer_3d([imshp, kshp, tshp], 0, self.optimizer_3d([imshp, kshp, tshp], 0,
local_abstractconv3d2d) 'conv3d2d',
'default',
None)
self.optimizer_3d([imshp, kshp, tshp], 0, self.optimizer_3d([imshp, kshp, tshp], 0,
local_abstractconv3d_cudnn_alternative) 'alternative',
'conv_gemm:default:conv3d2d',
dnn.GpuDnnConvGradW)
# backward pass wrt weight # backward pass wrt weight
self.optimizer_3d([imshp, tshp, kshp], 1, self.optimizer_3d([imshp, tshp, kshp], 1,
local_abstractconv3d_gemm_gradweights_alt) 'alternative',
'conv_dnn:default',
blas.GpuCorr3dMM)
self.optimizer_3d([imshp, tshp, kshp], 1, self.optimizer_3d([imshp, tshp, kshp], 1,
local_abstractconv3d_cudnn_alternative) 'alternative',
'conv_gemm:default',
dnn.GpuDnnConv)
# backward pass wrt inputs # backward pass wrt inputs
self.optimizer_3d([tshp, kshp, imshp], 2, self.optimizer_3d([tshp, kshp, imshp], 2,
local_abstractconv3d_gradinputs_gemm_alt) 'alternative',
'conv_dnn:default',
blas.GpuCorr3dMM)
self.optimizer_3d([tshp, kshp, imshp], 2, self.optimizer_3d([tshp, kshp, imshp], 2,
local_abstractconv3d_cudnn_alternative) 'alternative',
'conv_gemm:default',
dnn.GpuDnnConv)
# conv2d forward pass with Non-default border_mode and filter_dilation # conv2d forward pass with Non-default border_mode and filter_dilation
imshp2d = [(2, 3, 5, 5), (4, 2, 5, 5)] imshp2d = [(2, 3, 5, 5), (4, 2, 5, 5)]
...@@ -872,12 +867,16 @@ class Conv_opt_test(unittest.TestCase): ...@@ -872,12 +867,16 @@ class Conv_opt_test(unittest.TestCase):
filter_dilation = [(1, 1), (2, 2)] filter_dilation = [(1, 1), (2, 2)]
for imshp, kshp, fdil in zip(imshp2d, kshp2d, filter_dilation): for imshp, kshp, fdil in zip(imshp2d, kshp2d, filter_dilation):
self.optimizer_2d([imshp, kshp], 0, self.optimizer_2d([imshp, kshp], 0,
local_abstractconv_gemm_alternative, 'alternative',
'conv_dnn:default',
blas.GpuCorrMM_gradInputs,
border_mode='full', border_mode='full',
filter_dilation=fdil) filter_dilation=fdil)
# works only for cudnn > 6.0 # works only for cudnn > 6.0
self.optimizer_2d([imshp, kshp], 0, self.optimizer_2d([imshp, kshp], 0,
local_abstractconv_cudnn_alternative, 'alternative',
'conv_gemm:default',
dnn.GpuDnnConvGradI,
border_mode='full', border_mode='full',
filter_dilation=fdil) filter_dilation=fdil)
# conv3d forward pass with Non-default border_mode and filter_dilation # conv3d forward pass with Non-default border_mode and filter_dilation
...@@ -886,11 +885,15 @@ class Conv_opt_test(unittest.TestCase): ...@@ -886,11 +885,15 @@ class Conv_opt_test(unittest.TestCase):
filter_dilation = [(1, 1, 1), (2, 2, 2)] filter_dilation = [(1, 1, 1), (2, 2, 2)]
for imshp, kshp, fdil in zip(imshp3d, kshp3d, filter_dilation): for imshp, kshp, fdil in zip(imshp3d, kshp3d, filter_dilation):
self.optimizer_3d([imshp, kshp], 0, self.optimizer_3d([imshp, kshp], 0,
local_abstractconv3d_alt, 'alternative',
'conv_dnn:default:conv3d2d',
blas.GpuCorr3dMM_gradInputs,
border_mode='full', border_mode='full',
filter_dilation=fdil) filter_dilation=fdil)
# works only for cudnn > 6.0 # works only for cudnn > 6.0
self.optimizer_3d([imshp, kshp], 0, self.optimizer_3d([imshp, kshp], 0,
local_abstractconv3d_cudnn_alternative, 'alternative',
'conv_gemm:default:conv3d2d',
dnn.GpuDnnConvGradI,
border_mode='full', border_mode='full',
filter_dilation=fdil) filter_dilation=fdil)
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