def test_iconv(self, channels=64, seed=2019, h=8):
torch.random.manual_seed(seed)
weight_obj = iConv2d(channels, channels)
w = h
input_activation = torch.randn(1, channels, h, w)
_ = weight_obj(input_activation)
weight = weight_obj.conv.weight
weight_numpy = weight.detach().cpu().permute((2, 3, 0, 1)).numpy()
# compute 2d fft
# print(weight_numpy.shape)
kernel_fft = np.fft.fft2(weight_numpy, [h, w], axes=[0, 1], norm=None)
padded_numpy = np.pad(weight_numpy,
((0, h - 3), (0, w - 3), (0, 0), (0, 0)),
mode='constant')
kernel_fft2 = np.fft.fft2(padded_numpy, axes=[0, 1])
#print("original",(kernel_fft-kernel_fft2))
# then take svds
svds = np.linalg.svd(kernel_fft, compute_uv=False)
# finally log det is sum(log(singular values))
true_logdet = np.sum(np.log(svds))
#print(np.min(svds))
relative_error = torch.norm(
true_logdet - weight_obj.logdet()) / np.linalg.norm(true_logdet)
#print('relative error is: ', relative_error)
self.assertLess(relative_error, 4e-3)
def test_fftconv(self):
w = h = 3
channels = 5
torch.random.manual_seed(2019)
input_activation = torch.randn(1, channels, h, w)
layer = iConv2d(channels, channels)
fft_output = fft_conv3x3(input_activation,
layer.conv.weight).data.numpy()
conv_output = F.conv2d(input_activation, layer.conv.weight,
padding=1).data.numpy()
rel_error = np.linalg.norm(fft_output -
conv_output) / np.linalg.norm(fft_output)
self.assertLess(rel_error, 1e-6)
def test_ifftconv_pytorch(self):
w = h = 8
channels = 128
torch.random.manual_seed(2019)
x = torch.randn(1, channels, h, w)
layer = iConv2d(channels, channels)
conv_output = layer(x) - layer.conv.bias[None, :, None, None]
ifft_output = inverse_fft_conv3x3_pytorch(conv_output,
layer.conv.weight)
#print(ifft_output)
#print(x)
rel_error = (ifft_output - x).norm() / x.norm()
#print(rel_error)
self.assertLess(rel_error, 1e-4)
def test_clipping_pytorch(self):
w = h = 16
channels = 16
torch.random.manual_seed(2019)
x = torch.randn(1, channels, h, w).cuda()
layer = iConv2d(channels, channels).cuda()
#print(layer.conv.weight.data)
conv_output = layer(x) - layer.conv.bias[None, :, None, None]
clipped_np = Clip_OperatorNorm(layer.conv.weight.data, (h, w), (3, 10))
clipped_pt = Clip_OperatorNorm_PT(layer.conv.weight.data, (h, w),
(3, 10))
rel_err = (clipped_np - clipped_pt).norm() / clipped_np.norm()
#layer.conv.weight.data = Clip_OperatorNorm(layer.conv.weight.data,(h,w),(1,None))
#layer.conv.weight.data = Clip_OperatorNorm(layer.conv.weight.data,(h,w),(1,None))
#print(ifft_output)
#print(x)
# rel_error = (ifft_output-x).norm()/x.norm()
# #print(rel_error)
self.assertLess(rel_err, 1e-6)