def test_DepthConcat(self):
outputSize = torch.IntTensor((5, 6, 7, 8))
input = torch.randn(2, 3, 12, 12)
gradOutput = torch.randn(2, int(outputSize.sum()), 12, 12)
concat = nn.DepthConcat(1)
concat.add(nn.SpatialConvolution(3, outputSize[0], 1, 1, 1, 1)) # > 2, 5, 12, 12
concat.add(nn.SpatialConvolution(3, outputSize[1], 3, 3, 1, 1)) # > 2, 6, 10, 10
concat.add(nn.SpatialConvolution(3, outputSize[2], 4, 4, 1, 1)) # > 2, 7, 9, 9
concat.add(nn.SpatialConvolution(3, outputSize[3], 5, 5, 1, 1)) # > 2, 8, 8, 8
concat.zeroGradParameters()
# forward/backward
outputConcat = concat.forward(input)
gradInputConcat = concat.backward(input, gradOutput)
# the spatial dims are the largest, the nFilters is the sum
output = torch.Tensor(2, int(outputSize.sum()), 12, 12).zero_() # zero for padding
narrows = ((slice(None), slice(0, 5), slice(None), slice(None)),
(slice(None), slice(5, 11), slice(1, 11), slice(1, 11)),
(slice(None), slice(11, 18), slice(1, 10), slice(1, 10)),
(slice(None), slice(18, 26), slice(2, 10), slice(2, 10)))
gradInput = input.clone().zero_()
for i in range(4):
conv = concat.get(i)
gradWeight = conv.gradWeight.clone()
conv.zeroGradParameters()
output[narrows[i]].copy_(conv.forward(input))
gradInput.add_(conv.backward(input, gradOutput[narrows[i]]))
self.assertEqual(gradWeight, conv.gradWeight)
self.assertEqual(output, outputConcat)
self.assertEqual(gradInput, gradInputConcat)
# Check that these don't raise errors
concat.__repr__()
str(concat)
def build_conv_block(dim, padding_type, use_instance_norm):
conv_block = nn.Sequential()
p = 0
if padding_type == 'reflect':
conv_block.add(nn.SpatialReflectionPadding(1,1,1,1))
elif padding_type == 'replicate':
conv_block.add(nn.SpatialReplicationPadding(1,1,1,1))
elif padding_type == 'zero':
p = 1
conv_block.add(nn.SpationConvolution(dim, dim, 3, 3, 1, 1, p, p))
if use_instance_norm == 1:
conv_block.add(nn.InstanceNormalization(dim))
else:
conv_block.add(nn.SpatialBatchNormalization(dim))
conv_block.add(F.ReLU(True))
if padding_type == 'reflect':
conv_block.add(nn.SpatialReflectionPadding(1, 1, 1, 1))
elif padding_type == 'replicate':
conv_block.add(nn.SpatialReplicationPadding(1, 1, 1, 1))
conv_block.add(nn.SpatialConvolution(dim, dim, 3, 3, 1, 1, p, p))
if use_instance_norm == 1:
conv_block.add(nn.InstanceNormalization(dim))
else:
conv_block.add(nn.SpatialBatchNormalization(dim))
return conv_block
def FCNN(): num_classes = 2 n_layers_enc = 32 n_layers_ctx = 128 n_input = 5 prob_drop = 0.25 layers = [] # Encoder pool = nn2.SpatialMaxPooling(2,2,2,2) layers.append(nn2.SpatialConvolution(n_input, n_layers_enc, 3, 3, 1, 1, 1, 1)) layers.append(nn.ELU()) layers.append(nn2.SpatialConvolution(n_layers_enc, n_layers_enc, 3, 3, 1, 1, 1, 1)) layers.append(nn.ELU()) layers.append(pool) # Context Module layers.append(nn2.SpatialDilatedConvolution(n_layers_enc, n_layers_ctx, 3, 3, 1, 1, 1, 1, 1, 1)) layers.append(nn.ELU()) layers.append(nn2.SpatialDropout(prob_drop)) layers.append(nn2.SpatialDilatedConvolution(n_layers_ctx, n_layers_ctx, 3, 3, 1, 1, 2, 2, 2, 2)) layers.append(nn.ELU()) layers.append(nn2.SpatialDropout(prob_drop)) layers.append(nn2.SpatialDilatedConvolution(n_layers_ctx, n_layers_ctx, 3, 3, 1, 1, 4, 4, 4, 4)) layers.append(nn.ELU()) layers.append(nn2.SpatialDropout(prob_drop)) layers.append(nn2.SpatialDilatedConvolution(n_layers_ctx, n_layers_ctx, 3, 3, 1, 1, 8, 8, 8, 8)) layers.append(nn.ELU()) layers.append(nn2.SpatialDropout(prob_drop)) layers.append(nn2.SpatialDilatedConvolution(n_layers_ctx, n_layers_ctx, 3, 3, 1, 1, 16, 16, 16, 16)) layers.append(nn.ELU()) layers.append(nn2.SpatialDropout(prob_drop)) layers.append(nn2.SpatialDilatedConvolution(n_layers_ctx, n_layers_ctx, 3, 3, 1, 1, 32, 32, 32, 32)) layers.append(nn.ELU()) layers.append(nn2.SpatialDropout(prob_drop)) layers.append(nn2.SpatialDilatedConvolution(n_layers_ctx, n_layers_ctx, 3, 3, 1, 1, 64, 64, 64, 64)) layers.append(nn.ELU()) layers.append(nn2.SpatialDropout(prob_drop)) layers.append(nn2.SpatialDilatedConvolution(n_layers_ctx, n_layers_enc, 1, 1)) layers.append(nn.ELU()) # Nao havia no paper # Decoder layers.append(nn2.SpatialMaxUnpooling(pool)) layers.append(nn2.SpatialConvolution(n_layers_enc, n_layers_enc, 3, 3, 1, 1, 1, 1)) layers.append(nn.ELU()) layers.append(nn2.SpatialConvolution(n_layers_enc, num_classes, 3, 3, 1, 1, 1, 1)) layers.append(nn.ELU()) layers.append(nn.SoftMax()) # Nao havia no paper return nn.Sequential(*layers)
def build_model(opt):
model = torch.nn.Sequential()
prev_dim = 3
arch = opt.arch.split(',')
model = torch.nn.Sequential()
arch_len = len(arch)
for i in range(arch_len):
v = arch[i]
first_char = v[0]
needs_relu = True
needs_bn = True
next_dim = None
layer = None
if first_char == 'c':
f = int(v[1])
p = (f - 1) / 2
s = int(v[3])
next_dim = int(v[5])
if opt.padding_type == 'reflect':
model.add(nn.SpatialReflectionPadding(p, p, p, p))
p = 0
elif opt.padding_type == 'replicate':
model.add(nn.SpatialReplicationPadding(p, p, p, p))
p = 0
elif padding_type == 'none':
p = 0
layer = nn.SpatialConvolution(prev_dim, next_dim, f, f, s, s, p, p)
elif first_char == 'f':
f = int(v[1])
p = (f - 1) / 2
s = int(v[3])
a = s - 1
next_dim = int(v[5])
layer = nn.SpatialFullConvolution(prev_dim, next_dim, f, f, s, s, p, p, a, a)
elif first_char == 'd':
next_dim = int(v[1])
layer = nn.SpatialConvolution(prev_dim, next_dim, 3, 3, 2, 2, 1, 1)
elif first_char == 'U':
next_dim = prev_dim
scale = int(v[1])
layer = nn.SpatialFullConvolution(prev_dim, next_dim, 3, 3, 2, 2, 1, 1, 1, 1)
needs_bn = False
needs_relu = True
model.add(layer)
if i == arch_len - 1:
needs_bn = False
needs_relu = False
if needs_bn == True:
if opt.use_instance_norm == 1:
model.add(nn.InstanceNormalization(next_dim))
else:
model.add(nn.SpatialBatchNormalization(next_dim))
if needs_relu == True:
model.add(nn.ReLU(true))
prev_dim = next_dim
model.add(bb.Tanh())
model.add(nn.MulConstant(opt.tanh_constant))
model.add(nn.TotalVariation(opt.tv_strength))
def test_convolution(self):
self._test_single_layer(
nn.SpatialConvolution(3, 64, 7, 7, 2, 2, 3, 3),
decimal=6
)