Skip to content

P
pytensor
提交 e44ce417 authored 作者: affanv14's avatar affanv14 提交者： Mohammed Affan
浏览文件
操作

Improve test for abstractconv

上级 95450282
隐藏空白字符变更
内嵌 并排
正在显示 包含 123 行增加 和 111 行删除
theano/tensor/nnet/tests/test_abstract_conv.py
浏览文件 @ e44ce417
......@@ -3,6 +3,7 @@ import unittest
import numpy as np
from nose.plugins.skip import SkipTest
from nose.tools import assert_raises, assert_true
import theano
from theano import tensor
from theano.configparser import change_flags
......@@ -1701,126 +1702,137 @@ class TestConv2dGrads(unittest.TestCase):
class Grouped_conv_noOptim(unittest.TestCase):
conv2d = staticmethod(theano.tensor.nnet.abstract_conv.AbstractConv2d)
conv2d_gradw = staticmethod(theano.tensor.nnet.abstract_conv.AbstractConv2d_gradWeights)
conv2d_gradi = staticmethod(theano.tensor.nnet.abstract_conv.AbstractConv2d_gradInputs)
mode = theano.Mode(optimizer=None)
flip_filter = False
def setUp(self):
self.num_groups = 3
self.num_groups = [3, 2, 4, 4]
self.border_mode = 'valid'
self.subsample = (1, 1)
self.img_shape = (5, 6, 5, 5)
self.kern_shape = (6, 2, 3, 3)
self.top_shape = (5, 6, 3, 3)
self.img_shape = [(5, 6, 5, 5), (4, 4, 7, 5), (3, 8, 5, 3), (2, 4, 7, 7)]
self.kern_shape = [(6, 2, 3, 3), (6, 2, 5, 3), (4, 2, 3, 3), (4, 1, 3, 5)]
self.top_shape = [(5, 6, 3, 3), (4, 6, 3, 3), (3, 4, 3, 1), (2, 4, 5, 3)]
self.filter_dilation = (1, 1)
self.mode = theano.Mode(optimizer=None)
self.ref_mode = 'FAST_RUN'
def test_fwd(self):
img = np.random.random(self.img_shape).astype('float32')
kern = np.random.random(self.kern_shape).astype('float32')
split_imgs = np.split(img, self.num_groups, axis=1)
split_kern = np.split(kern, self.num_groups, axis=0)
img_var = theano.tensor.tensor4('img')
kern_var = theano.tensor.tensor4('kern')
grouped_abstractconv_func = theano.tensor.nnet.abstract_conv.AbstractConv2d(imshp=None,
kshp=None,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation,
num_groups=self.num_groups)
grouped_conv_output = grouped_abstractconv_func(img_var, kern_var)
grouped_func = theano.function([img_var, kern_var], grouped_conv_output, mode=self.mode)
grouped_output = grouped_func(img, kern)
normal_conv_output = conv2d_corr(img_var,
kern_var,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation)
normal_func = theano.function([img_var, kern_var], normal_conv_output,
mode='FAST_RUN')
normal_concat_output = [normal_func(img_arr, kern_arr)
for img_arr, kern_arr in zip(split_imgs, split_kern)]
normal_concat_output = np.concatenate(normal_concat_output, axis=1)
utt.assert_allclose(grouped_output, normal_concat_output)
utt.verify_grad(grouped_abstractconv_func,
[img, kern],
mode=self.mode)
for imshp, kshp, groups in zip(self.img_shape, self.kern_shape, self.num_groups):
img = np.random.random(imshp).astype('float32')
kern = np.random.random(kshp).astype('float32')
split_imgs = np.split(img, groups, axis=1)
split_kern = np.split(kern, groups, axis=0)
img_sym = theano.tensor.tensor4('img')
kern_sym = theano.tensor.tensor4('kern')
grouped_abstractconv_func = self.conv2d(border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation,
num_groups=groups)
if self.flip_filter:
grouped_conv_output = grouped_abstractconv_func(img_sym, kern_sym[:, :, ::-1, ::-1])
else:
grouped_conv_output = grouped_abstractconv_func(img_sym, kern_sym)
grouped_func = theano.function([img_sym, kern_sym], grouped_conv_output, mode=self.mode)
grouped_output = grouped_func(img, kern)
normal_conv_output = conv2d_corr(img_sym,
kern_sym,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation)
normal_func = theano.function([img_sym, kern_sym], normal_conv_output,
mode=self.ref_mode)
normal_concat_output = [normal_func(img_arr, kern_arr)
for img_arr, kern_arr in zip(split_imgs, split_kern)]
normal_concat_output = np.concatenate(normal_concat_output, axis=1)
utt.assert_allclose(grouped_output, normal_concat_output)
utt.verify_grad(grouped_abstractconv_func,
[img, kern],
mode=self.mode)
def test_gradweights(self):
img = np.random.random(self.img_shape).astype('float32')
top = np.random.random(self.top_shape).astype('float32')
split_imgs = np.split(img, self.num_groups, axis=1)
split_top = np.split(top, self.num_groups, axis=1)
img_var = theano.tensor.tensor4('img')
top_var = theano.tensor.tensor4('top')
grouped_abstractconvgrad_func = theano.tensor.nnet.abstract_conv.AbstractConv2d_gradWeights(imshp=None,
kshp=None,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation,
num_groups=self.num_groups)
grouped_conv_output = grouped_abstractconvgrad_func(img_var, top_var, self.kern_shape[-2:])
grouped_func = theano.function([img_var, top_var], grouped_conv_output, mode=self.mode)
grouped_output = grouped_func(img, top)
normal_conv_output = conv2d_corr_gw(img_var,
top_var,
self.kern_shape,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation)
normal_func = theano.function([img_var, top_var], normal_conv_output,
mode='FAST_RUN')
normal_concat_output = [normal_func(img_arr, top_arr)
for img_arr, top_arr in zip(split_imgs, split_top)]
normal_concat_output = np.concatenate(normal_concat_output, axis=0)
utt.assert_allclose(grouped_output, normal_concat_output)
def abstract_conv_gradweight(inputs_val, output_val):
return grouped_abstractconvgrad_func(inputs_val, output_val, self.kern_shape[-2:])
for imshp, kshp, tshp, groups in zip(self.img_shape, self.kern_shape, self.top_shape, self.num_groups):
img = np.random.random(imshp).astype('float32')
top = np.random.random(tshp).astype('float32')
split_imgs = np.split(img, groups, axis=1)
split_top = np.split(top, groups, axis=1)
img_sym = theano.tensor.tensor4('img')
top_sym = theano.tensor.tensor4('top')
grouped_abstractconvgrad_func = self.conv2d_gradw(border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation,
num_groups=groups)
grouped_conv_output = grouped_abstractconvgrad_func(img_sym, top_sym, kshp[-2:])
if self.flip_filter:
grouped_conv_output = grouped_conv_output[:, :, ::-1, ::-1]
grouped_func = theano.function([img_sym, top_sym], grouped_conv_output, mode=self.mode)
grouped_output = grouped_func(img, top)
normal_conv_output = conv2d_corr_gw(img_sym,
top_sym,
kshp,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation)
normal_func = theano.function([img_sym, top_sym], normal_conv_output,
mode=self.ref_mode)
normal_concat_output = [normal_func(img_arr, top_arr)
for img_arr, top_arr in zip(split_imgs, split_top)]
normal_concat_output = np.concatenate(normal_concat_output, axis=0)
utt.assert_allclose(grouped_output, normal_concat_output)
def abstract_conv_gradweight(inputs_val, output_val):
return grouped_abstractconvgrad_func(inputs_val, output_val, kshp[-2:])
utt.verify_grad(abstract_conv_gradweight,
[img, top],
mode=self.mode, eps=1)
utt.verify_grad(abstract_conv_gradweight,
[img, top],
mode=self.mode, eps=1)
def test_gradinputs(self):
kern = np.random.random(self.kern_shape).astype('float32')
top = np.random.random(self.top_shape).astype('float32')
split_kerns = np.split(kern, self.num_groups, axis=0)
split_top = np.split(top, self.num_groups, axis=1)
kern_var = theano.tensor.tensor4('kern')
top_var = theano.tensor.tensor4('top')
grouped_abstractconvgrad_func = theano.tensor.nnet.abstract_conv.AbstractConv2d_gradInputs(imshp=None,
kshp=None,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation,
num_groups=self.num_groups)
grouped_conv_output = grouped_abstractconvgrad_func(kern_var, top_var, self.img_shape[-2:])
grouped_func = theano.function([kern_var, top_var], grouped_conv_output, mode=self.mode)
grouped_output = grouped_func(kern, top)
normal_conv_output = conv2d_corr_gi(kern_var,
top_var,
self.img_shape,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation)
normal_func = theano.function([kern_var, top_var], normal_conv_output,
mode='FAST_RUN')
normal_concat_output = [normal_func(kern_arr, top_arr)
for kern_arr, top_arr in zip(split_kerns, split_top)]
normal_concat_output = np.concatenate(normal_concat_output, axis=1)
utt.assert_allclose(grouped_output, normal_concat_output)
def abstract_conv_gradinputs(filters_val, output_val):
return grouped_abstractconvgrad_func(filters_val, output_val, self.img_shape[2:])
for imshp, kshp, tshp, groups in zip(self.img_shape, self.kern_shape, self.top_shape, self.num_groups):
kern = np.random.random(kshp).astype('float32')
top = np.random.random(tshp).astype('float32')
split_kerns = np.split(kern, groups, axis=0)
split_top = np.split(top, groups, axis=1)
kern_sym = theano.tensor.tensor4('kern')
top_sym = theano.tensor.tensor4('top')
grouped_abstractconvgrad_func = self.conv2d_gradi(border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation,
num_groups=groups)
if self.flip_filter:
grouped_conv_output = grouped_abstractconvgrad_func(kern_sym[:, :, ::-1, ::-1], top_sym, imshp[-2:])
else:
grouped_conv_output = grouped_abstractconvgrad_func(kern_sym, top_sym, imshp[-2:])
grouped_func = theano.function([kern_sym, top_sym], grouped_conv_output, mode=self.mode)
grouped_output = grouped_func(kern, top)
normal_conv_output = conv2d_corr_gi(kern_sym,
top_sym,
imshp,
border_mode=self.border_mode,
subsample=self.subsample,
filter_dilation=self.filter_dilation)
normal_func = theano.function([kern_sym, top_sym], normal_conv_output,
mode=self.ref_mode)
normal_concat_output = [normal_func(kern_arr, top_arr)
for kern_arr, top_arr in zip(split_kerns, split_top)]
normal_concat_output = np.concatenate(normal_concat_output, axis=1)
utt.assert_allclose(grouped_output, normal_concat_output)
def abstract_conv_gradinputs(filters_val, output_val):
return grouped_abstractconvgrad_func(filters_val, output_val, imshp[2:])
utt.verify_grad(abstract_conv_gradinputs,
[kern, top],
mode=self.mode, eps=1)
utt.verify_grad(abstract_conv_gradinputs,
[kern, top],
mode=self.mode, eps=1)
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