def mapped_transposed_conv_test(conv_layer,
weight,
input,
oh,
ow,
sample_map,
interp_weights=None,
cuda=False):
# Make sure the input requires a gradient
input.requires_grad = True
if cuda:
input = input.cuda()
sample_map = sample_map.cuda()
# Set the layer weights
bias = params.bias(weight.shape[1])
conv_layer.weight.data = weight if not cuda else weight.cuda()
conv_layer.bias.data = bias if not cuda else bias.cuda()
# Run a forward pass
output, forward_time = time_cuda(
conv_layer, [input, oh, ow, sample_map, interp_weights])
# Run a backward pass
_, backward_time = time_cuda(output.backward, [params.gradients(output)])
# Run a numerical gradient check
gradcheck_result = gradcheck(conv_layer,
(input, oh, ow, sample_map, interp_weights))
# Return the report
return output, forward_time, backward_time, gradcheck_result
def mapped_resample_test(resample_layer,
input,
sample_map,
interp_weights=None,
cuda=False):
# Make sure the input requires a gradient
input.requires_grad = True
if cuda:
input = input.cuda()
sample_map = sample_map.cuda()
if interp_weights is not None:
interp_weights = interp_weights.cuda()
# Run a forward pass
output, forward_time = time_cuda(resample_layer,
[input, sample_map, interp_weights])
# Run a backward pass
_, backward_time = time_cuda(output.backward, [params.gradients(output)])
# Run a numerical gradient check
gradcheck_result = gradcheck(resample_layer,
(input, sample_map, interp_weights))
# Return the report
return output, forward_time, backward_time, gradcheck_result
