Skip to content

P
pytensor
提交 30c18aa4 authored 作者: Tegan Maharaj's avatar Tegan Maharaj
浏览文件
操作

fixed bugs

上级 4eda607b
显示空白字符变更
内嵌 并排
正在显示 包含 37 行增加 和 15 行删除
theano/tensor/nnet/tests/test_abstract_conv.py
浏览文件 @ 30c18aa4
...@@ -1605,10 +1605,9 @@ class TestConv2dGrads(unittest.TestCase): ...@@ -1605,10 +1605,9 @@ class TestConv2dGrads(unittest.TestCase):
self.output_grad = theano.tensor.tensor4() self.output_grad = theano.tensor.tensor4()
self.output_grad_wrt = theano.tensor.tensor4() self.output_grad_wrt = theano.tensor.tensor4()
self.filters = theano.tensor.tensor4()
self.x = theano.tensor.tensor4('x', theano.config.floatX) # inputs self.x = theano.tensor.tensor4('x', theano.config.floatX) # inputs
self.w = theano.tensor.tensor4('w', theano.config.floatX) # weights self.w = theano.tensor.tensor4('w', theano.config.floatX) # filter weights
def test_conv2d_grad_wrt_inputs(self): def test_conv2d_grad_wrt_inputs(self):
"""Compares calculated abstract grads wrt inputs with the fwd grads """Compares calculated abstract grads wrt inputs with the fwd grads
...@@ -1616,32 +1615,48 @@ class TestConv2dGrads(unittest.TestCase): ...@@ -1616,32 +1615,48 @@ class TestConv2dGrads(unittest.TestCase):
the outputs of T.nnet.conv forward grads to make sure the the outputs of T.nnet.conv forward grads to make sure the
results are the same. results are the same.
""" """
for (in_shape, fltr_shape) in zip(self.inputs_shapes, self.filters_shapes): for (in_shape, fltr_shape) in zip(self.inputs_shapes, self.filters_shapes):
for bm in self.border_modes: for bm in self.border_modes:
for ss in self.subsamples: for ss in self.subsamples:
for ff in self.filter_flip: for ff in self.filter_flip:
if self.filter_flip: if self.filter_flip:
fltr_shape = fltr_shape[::1] # conv2d doesn't seem to have filter_flip fltr_shape = fltr_shape[::1] # conv2d doesn't seem to have filter_flip
# get random values of the right shapes
input_val = self.random_stream.random_sample(in_shape).astype(theano.config.floatX)
filter_val = self.random_stream.random_sample(fltr_shape).astype(theano.config.floatX)
out_grad_shape = theano.tensor.nnet.abstract_conv.get_conv_output_shape(image_shape=in_shape,
kernel_shape=fltr_shape,
border_mode=bm,
subsample=ss)
out_grad_val = self.random_stream.random_sample(out_grad_shape).astype(theano.config.floatX)
# old conv
conv_out = theano.tensor.nnet.conv.conv2d(self.x, conv_out = theano.tensor.nnet.conv.conv2d(self.x,
filters=self.filters, filters=self.w,
border_mode=bm, border_mode=bm,
subsample=ss, subsample=ss,
image_shape=in_shape, image_shape=in_shape,
filter_shape=fltr_shape filter_shape=fltr_shape
) )
# grad of old conv
conv_grad = theano.grad(conv_out.sum(), wrt=[self.x], known_grads={conv_out: self.output_grad}) conv_grad = theano.grad(conv_out.sum(), wrt=[self.x], known_grads={conv_out: self.output_grad})
f_prime = theano.function([self.x, self.output_grad, self.filters], conv_grad) f_old = theano.function([self.x, self.w, self.output_grad], conv_grad)
conv_wrt_i_out = theano.tensor.nnet.conv.abstract_conv.conv2d_grad_wrt_inputs(self.output_grad_wrt, # new conv + grad (wrt i)
filters=self.filters, conv_wrt_i_out = theano.tensor.nnet.abstract_conv.conv2d_grad_wrt_inputs(output_grad=self.output_grad_wrt,
filters=self.w,
border_mode=bm, border_mode=bm,
subsample=ss, subsample=ss,
input_shape=in_shape, input_shape=in_shape,
filter_shape=fltr_shape, filter_shape=fltr_shape,
filter_flip=ff filter_flip=ff
) )
f = theano.function([self.x, self.output_grad_wrt, self.filters], conv_wrt_i_out) f_new = theano.function([self.w, self.output_grad_wrt], conv_wrt_i_out)
utt.assert_allclose(f, f_prime)
# check that they're equal
utt.assert_allclose(f_new(filter_val, out_grad_val), f_old(input_val, filter_val, out_grad_val))
def test_conv2d_grad_wrt_weights(self): def test_conv2d_grad_wrt_weights(self):
"""Compares calculated abstract grads wrt weights with the fwd grads """Compares calculated abstract grads wrt weights with the fwd grads
...@@ -1656,23 +1671,30 @@ class TestConv2dGrads(unittest.TestCase): ...@@ -1656,23 +1671,30 @@ class TestConv2dGrads(unittest.TestCase):
for ff in self.filter_flip: for ff in self.filter_flip:
if self.filter_flip: if self.filter_flip:
fltr_shape = fltr_shape[::1] # conv2d doesn't seem to have filter_flip fltr_shape = fltr_shape[::1] # conv2d doesn't seem to have filter_flip
conv_out = theano.tensor.nnet.conv.conv2d(self.w, input_val = self.random_stream.random_sample(in_shape).astype(theano.config.floatX)
filters=self.filters, filter_val = self.random_stream.random_sample(fltr_shape).astype(theano.config.floatX)
out_grad_shape = theano.tensor.nnet.abstract_conv.get_conv_output_shape(image_shape=in_shape,
kernel_shape=fltr_shape,
border_mode=bm,
subsample=ss)
out_grad_val = self.random_stream.random_sample(out_grad_shape).astype(theano.config.floatX)
conv_out = theano.tensor.nnet.conv.conv2d(self.x,
filters=self.w,
border_mode=bm, border_mode=bm,
subsample=ss, subsample=ss,
image_shape=in_shape, image_shape=in_shape,
filter_shape=fltr_shape filter_shape=fltr_shape
) )
conv_grad = theano.grad(conv_out.sum(), wrt=[self.w], known_grads={conv_out: self.output_grad}) conv_grad = theano.grad(conv_out.sum(), wrt=[self.w], known_grads={conv_out: self.output_grad})
f_prime = theano.function([self.w, self.output_grad, self.filters], conv_grad) f_old = theano.function([self.x, self.w, self.output_grad], conv_grad)
conv_wrt_w_out = theano.tensor.nnet.conv.abstract_conv.conv2d_grad_wrt_weights(self.output_grad_wrt, conv_wrt_w_out = theano.tensor.nnet.abstract_conv.conv2d_grad_wrt_weights(self.x,
filters=self.filters, output_grad=self.output_grad_wrt,
border_mode=bm, border_mode=bm,
subsample=ss, subsample=ss,
input_shape=in_shape, input_shape=in_shape,
filter_shape=fltr_shape, filter_shape=fltr_shape,
filter_flip=ff filter_flip=ff
) )
f = theano.function([self.w, self.output_grad_wrt, self.filters], conv_wrt_w_out) f_new = theano.function([self.x, self.output_grad_wrt], conv_wrt_w_out)
utt.assert_allclose(f, f_prime) utt.assert_allclose(f_new(input_val, out_grad_val), f_old(input_val, filter_val, out_grad_val))
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