def check_backward(self, x_data, W_data, b_data, y_grad):
xp = cuda.get_array_module(x_data)
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
self.assertFalse(x_data.flags.c_contiguous)
self.assertFalse(W_data.flags.c_contiguous)
self.assertFalse(y_grad.flags.c_contiguous)
if b_data is not None:
b = xp.empty((len(b_data) * 2, ), dtype=self.b.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data, )
with chainer.using_config('use_cudnn', self.use_cudnn):
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
gradient_check.check_backward(
convolution_2d.Convolution2DFunction(
self.stride, self.pad, self.cover_all), args, y_grad,
**self.check_backward_options)
def check_forward(self, x_data, W_data, b_data):
args1 = (x_data, W_data)
args2 = (x_data, W_data)
if b_data is not None:
args1 = args1 + (b_data, )
b_data = sum(numpy.split(b_data, W_data.shape[1]))
args2 = args2 + (b_data, )
f1 = depthwise_convolution_2d.DepthwiseConvolution2D(
self.stride, self.pad)
y1 = f1(*args1)
arys = numpy.split(y1.data, self.W.shape[1], axis=1)
y1 = sum(arys)
f2 = convolution_2d.Convolution2DFunction(self.stride, self.pad)
y2 = f2.apply(args2)[0].data
testing.assert_allclose(y1, y2, **self.check_forward_options)
def check_forward(self, x_data, W_data, b_data):
# If all the filters are the same,
# the operation is equivalent to convolution_2d
for i in moves.range(W_data.shape[1]):
for j in moves.range(W_data.shape[2]):
W_data[:, i, j, ...] = W_data[:, 0, 0, ...]
args1 = (x_data, W_data)
args2 = (x_data, W_data[:, 0, 0, ...])
if b_data is not None:
for i in moves.range(b_data.shape[1]):
for j in moves.range(b_data.shape[2]):
b_data[:, i, j, ] = b_data[:, 0, 0, ]
args1 = args1 + (b_data,)
b_data2 = b_data[:, 0, 0, ]
args2 = args2 + (b_data2,)
f1 = local_convolution_2d.LocalConvolution2DFunction(self.stride)
y1 = f1.apply(args1)[0].data
f2 = convolution_2d.Convolution2DFunction(self.stride, 0)
y2 = f2.apply(args2)[0].data
testing.assert_allclose(y1, y2, **self.check_forward_options)
def check_backward(self, x_data, W_data, b_data, y_grad):
xp = cuda.get_array_module(x_data)
# cuDNN < v3 does not support deterministic algorithms.
# In that case, Chainer should raise errors.
# As this behavior is tested by TestConvolution2DCudnnCall,
# we simply skip the test here.
should_raise_error = ((xp is not numpy) and self.use_cudnn
and self.cudnn_deterministic
and cuda.cudnn_enabled
and cuda.cudnn.cudnn.getVersion() < 3000)
if should_raise_error:
return
if not self.c_contiguous:
x_data = xp.asfortranarray(x_data)
W_data = xp.asfortranarray(W_data)
y_grad = xp.asfortranarray(y_grad)
self.assertFalse(x_data.flags.c_contiguous)
self.assertFalse(W_data.flags.c_contiguous)
self.assertFalse(y_grad.flags.c_contiguous)
if b_data is not None:
b = xp.empty((len(b_data) * 2, ), dtype=self.b.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data, )
with chainer.using_config('use_cudnn', self.use_cudnn):
with chainer.using_config('cudnn_deterministic',
self.cudnn_deterministic):
gradient_check.check_backward(
convolution_2d.Convolution2DFunction(
self.stride, self.pad, self.cover_all), args, y_grad,
**self.check_backward_options)
def check_backward(self, x_data, W_data, b_data, y_grad):
if not self.c_contiguous:
x_data = _asfortranarray(x_data)
W_data = _asfortranarray(W_data)
y_grad = _asfortranarray(y_grad)
self.assertFalse(x_data.flags.c_contiguous)
self.assertFalse(W_data.flags.c_contiguous)
self.assertFalse(y_grad.flags.c_contiguous)
if b_data is not None:
xp = cuda.get_array_module(b_data)
b = xp.empty((len(b_data) * 2, ), dtype=self.b.dtype)
b[::2] = b_data
b_data = b[::2]
self.assertFalse(b_data.flags.c_contiguous)
args = (x_data, W_data)
if b_data is not None:
args = args + (b_data, )
gradient_check.check_backward(convolution_2d.Convolution2DFunction(
self.stride, self.pad, self.use_cudnn),
args,
y_grad,
eps=1e-2)
