def output_shape_for(self, input1_shape, input2_shape):
"""
Ignore:
>>> self = DenseUNetUp(128, 256)
>>> input1_shape = [4, 128, 24, 24]
>>> input2_shape = [4, 256, 8, 8]
>>> output_shape = self.output_shape_for(input1_shape, input2_shape)
>>> inputs1 = torch.autograd.Variable(torch.rand(input1_shape))
>>> inputs2 = torch.autograd.Variable(torch.rand(input2_shape))
>>> assert self.forward(inputs1, inputs2).shape == output_shape
>>> print('output_shape = {!r}'.format(output_shape))
output_shape = (4, 64, 24, 24)
"""
output2_shape = OutputShapeFor(self.up)(input2_shape)
output2_shape = OutputShapeFor(self.pad)(output2_shape, input1_shape)
# Taking the easy way out and padding the upsampled layer instead of
# cropping the down layer
# output1_shape = OutputShapeFor(self.pad)(input1_shape, output2_shape)
# cat_shape = OutputShapeFor(torch.cat)([output1_shape, output2_shape], 1)
cat_shape = OutputShapeFor(torch.cat)([input1_shape, output2_shape], 1)
conv_shape = OutputShapeFor(self.conv)(cat_shape)
output_shape = OutputShapeFor(self.bottleneck)(conv_shape)
return output_shape
def activation_shapes(self, input_shape):
norm_shape = OutputShapeFor(self._modules['norm'])(input_shape)
noli_shape = OutputShapeFor(self._modules['noli'])(norm_shape)
conv_shape = OutputShapeFor(self._modules['conv'])(noli_shape)
pool_shape = OutputShapeFor(self._modules['pool'])(conv_shape)
activations = [
norm_shape
]
if not self._modules['noli'].inplace:
activations.append(np.prod(noli_shape))
activations += [conv_shape, pool_shape]
return activations
def output_shape_for(self, input_shape):
residual = input_shape
out = OutputShapeFor(self.convbnrelu1)(input_shape)
out = OutputShapeFor(self.convbn2)(out)
if self.downsample is not None:
residual = OutputShapeFor(self.downsample)(input_shape)
if residual[:-3] != out[:-3]:
print('disagree:')
print('out = {!r}'.format(out))
print('residual = {!r}'.format(residual))
out = OutputShapeFor(self.relu)(out)
return out
def activation_shapes(self, input1_shape, input2_shape):
up2_shape = OutputShapeFor(self.up)(input2_shape)
pad2_shape = OutputShapeFor(self.pad)(up2_shape, input1_shape)
cat_shape = OutputShapeFor(torch.cat)([input1_shape, pad2_shape], 1)
conv_shape = OutputShapeFor(self.conv)(cat_shape)
output_shape = OutputShapeFor(self.bottleneck)(conv_shape)
activations = [up2_shape]
activations += self.pad.activation_shapes(up2_shape, input1_shape)
activations += [cat_shape]
activations += self.conv.activation_shapes(cat_shape)
activations += [output_shape]
return activations
def __init__(self, input_shape):
"""
"""
super(C3D, self).__init__()
# nonlinearity = partial(nn.ReLU)
# kernels are specified in D, H, W
feats = [64, 128, 256, 512, 512]
conv_blocks = nn.Sequential(OrderedDict([
('block1', Conv3DBlock(in_channels=3, out_channels=feats[0], n_conv=1,
pool_kernel=(1, 2, 2), pool_stride=(1, 2, 2))),
('block2', Conv3DBlock(in_channels=feats[0], out_channels=feats[1], n_conv=1)),
('block3', Conv3DBlock(in_channels=feats[1], out_channels=feats[2], n_conv=2)),
('block4', Conv3DBlock(in_channels=feats[2], out_channels=feats[3], n_conv=2)),
('block5', Conv3DBlock(in_channels=feats[3], out_channels=feats[4], n_conv=2)),
]))
output_shape = OutputShapeFor(conv_blocks)(input_shape)
print('output_shape = {!r}'.format(output_shape))
import numpy as np
self.input_shape = input_shape
self.conv_blocks = conv_blocks
self.n_conv_output = int(np.prod(output_shape[1:]))
self.block6 = FCBlock(self.n_conv_output, 4096)
self.block7 = FCBlock(4096, 4096)
self.softmax = nn.Softmax(dim=1)
def activation_shapes(self, input_shape):
shape = input_shape
activations = []
for i in range(self.num_layers):
module = self._modules['denselayer%d' % (i + 1)]
activations += module.activation_shapes(shape)
shape = OutputShapeFor(module)(shape)
return activations
def output_shape_for(self, input1_shape, input2_shape):
"""
Example:
>>> self = UNetUp(256, 128)
>>> input1_shape = [4, 128, 24, 24]
>>> input2_shape = [4, 256, 8, 8]
>>> output_shape = self.output_shape_for(input1_shape, input2_shape)
>>> output_shape
(4, 128, 12, 12)
>>> inputs1 = torch.autograd.Variable(torch.rand(input1_shape))
>>> inputs2 = torch.autograd.Variable(torch.rand(input2_shape))
>>> assert self.forward(inputs1, inputs2).shape == output_shape
"""
output2_shape = OutputShapeFor(self.up)(input2_shape)
output1_shape = OutputShapeFor(self.pad)(input1_shape, output2_shape)
cat_shape = OutputShapeFor(torch.cat)([output1_shape, output2_shape], 1)
output_shape = OutputShapeFor(self.conv)(cat_shape)
return output_shape
def io_shapes(conn, self, input_shape):
output_shapes = ub.odict()
input_shapes = ub.odict()
# prev = None
for node in conn.topsort:
in_names = conn.input_nodes[node]
if in_names is None:
in_shapes = [input_shape]
else:
in_shapes = list(ub.take(output_shapes, in_names))
input_shapes[node] = in_shapes
out_shapes = OutputShapeFor(getattr(self, node))(*in_shapes)
output_shapes[node] = out_shapes
conn.output_shapes = output_shapes
conn.input_shapes = input_shapes
def resnet_prepool_output_shape(self, input_shape):
"""
self = SiameseLP(input_shape=input_shape)
input_shape = (1, 3, 224, 224)
self.resnet_prepool_output_shape(input_shape)
self = SiameseLP(input_shape=input_shape)
input_shape = (1, 3, 416, 416)
self.resnet_prepool_output_shape(input_shape)
"""
# Figure out how big the output will be and redo the average pool layer
# to account for it
branch = self.branch
shape = input_shape
shape = OutputShapeFor(branch.conv1)(shape)
shape = OutputShapeFor(branch.bn1)(shape)
shape = OutputShapeFor(branch.relu)(shape)
shape = OutputShapeFor(branch.maxpool)(shape)
shape = OutputShapeFor(branch.layer1)(shape)
shape = OutputShapeFor(branch.layer2)(shape)
shape = OutputShapeFor(branch.layer3)(shape)
shape = OutputShapeFor(branch.layer4)(shape)
prepool_shape = shape
return prepool_shape
def activation_shapes(self, input_shape):
norm1_shape = OutputShapeFor(self._modules['norm.1'])(input_shape)
noli1_shape = OutputShapeFor(self._modules['noli.1'])(norm1_shape)
conv1_shape = OutputShapeFor(self._modules['conv.1'])(noli1_shape)
norm2_shape = OutputShapeFor(self._modules['norm.2'])(conv1_shape)
noli2_shape = OutputShapeFor(self._modules['noli.2'])(norm2_shape)
conv2_shape = OutputShapeFor(self._modules['conv.2'])(noli2_shape)
activations = [
norm1_shape,
]
if not self._modules['noli.1'].inplace:
activations.append(np.prod(noli1_shape))
activations += [
conv1_shape,
norm2_shape,
]
if not self._modules['noli.2'].inplace:
activations.append(np.prod(noli2_shape))
activations += [
conv2_shape,
]
return activations
def output_shape_for(self, input_shape):
return OutputShapeFor(self.dcb_unit)(input_shape)
def output_shape_for(self, input_shape1, input_shape2):
shape1 = OutputShapeFor(self.branch)(input_shape1)
shape2 = OutputShapeFor(self.branch)(input_shape2)
assert shape1 == shape2
output_shape = (shape1[0], 1)
return output_shape
def output_shape_for(self, input_shape):
return OutputShapeFor(self.sequence)(input_shape)
def output_shape_for(self, input_shape, math=math):
shape = OutputShapeFor(self.conv1[0])(input_shape)
shape = OutputShapeFor(self.conv2[0])(shape)
output_shape = shape
return output_shape
def raw_output_shape_for(self, input_shape):
# output shape without fancy prepad mirrors and post crops
shape = conv1 = OutputShapeFor(self.conv1)(input_shape)
shape = OutputShapeFor(self.maxpool1)(shape)
shape = conv2 = OutputShapeFor(self.conv2)(shape)
shape = OutputShapeFor(self.maxpool2)(shape)
shape = conv3 = OutputShapeFor(self.conv3)(shape)
shape = OutputShapeFor(self.maxpool3)(shape)
shape = conv4 = OutputShapeFor(self.conv4)(shape)
shape = OutputShapeFor(self.maxpool4)(shape)
shape = OutputShapeFor(self.center)(shape)
shape = OutputShapeFor(self.up_concat4)(conv4, shape)
shape = OutputShapeFor(self.up_concat3)(conv3, shape)
shape = OutputShapeFor(self.up_concat2)(conv2, shape)
shape = OutputShapeFor(self.up_concat1)(conv1, shape)
shape = OutputShapeFor(self.final)(shape)
output_shape = shape
return output_shape
def output_shape_for(self, input_shape):
# N1, C1, W1, H1 = input_shape
# output_shape = (N1, self.n_classes, W1, H1)
shape = input_shape
shape = conv1 = OutputShapeFor(self.conv1)(shape)
shape = OutputShapeFor(self.maxpool1)(shape)
shape = conv2 = OutputShapeFor(self.conv2)(shape)
shape = OutputShapeFor(self.maxpool2)(shape)
shape = conv3 = OutputShapeFor(self.conv3)(shape)
shape = OutputShapeFor(self.maxpool3)(shape)
shape = conv4 = OutputShapeFor(self.conv4)(shape)
shape = OutputShapeFor(self.maxpool4)(shape)
shape = OutputShapeFor(self.center)(shape)
shape = OutputShapeFor(self.up_concat4)(conv4, shape)
shape = OutputShapeFor(self.up_concat3)(conv3, shape)
shape = OutputShapeFor(self.up_concat2)(conv2, shape)
shape = OutputShapeFor(self.up_concat1)(conv1, shape)
shape = OutputShapeFor(self.final)(shape)
output_shape = shape
return output_shape