def create_layers(self, shape, crop_size=8, conv_args=None, fc_args=None):
'''Creates layers
conv_args are in format (dim_h, f_size, stride, pad, batch_norm, dropout, nonlinearity, pool)
fc_args are in format (dim_h, batch_norm, dropout, nonlinearity)
Args:
shape: Shape of input.
crop_size: Size of crops
conv_args: List of tuple of convolutional arguments.
fc_args: List of tuple of fully-connected arguments.
'''
self.crop_size = crop_size
dim_x, dim_y, dim_in = shape
if dim_x != dim_y:
raise ValueError('x and y dimensions must be the same to use Folded encoders.')
self.final_size = 2 * (dim_x // self.crop_size) - 1
self.unfold = Unfold(dim_x, self.crop_size)
self.refold = Fold(dim_x, self.crop_size)
shape = (self.crop_size, self.crop_size, dim_in)
self.conv_layers, self.conv_shape = self.create_conv_layers(shape, conv_args)
dim_x, dim_y, dim_out = self.conv_shape
dim_r = dim_x * dim_y * dim_out
self.reshape = View(-1, dim_r)
self.fc_layers, _ = self.create_linear_layers(dim_r, fc_args)
def create_layers(self,
shape,
conv_before_args=None,
res_args=None,
conv_after_args=None,
fc_args=None):
'''Creates layers
Args:
shape: Shape of the input.
conv_before_args: Arguments for convolutional layers before residuals.
res_args: Residual args.
conv_after_args: Arguments for convolutional layers after residuals.
fc_args: Fully-connected arguments.
'''
dim_x, dim_y, dim_in = shape
shape = (dim_x, dim_y, dim_in)
self.conv_before_layers, self.conv_before_shape = self.create_conv_layers(
shape, conv_before_args)
self.res_layers, self.res_shape = self.create_res_layers(
self.conv_before_shape, res_args)
self.conv_after_layers, self.conv_after_shape = self.create_conv_layers(
self.res_shape, conv_after_args)
dim_x, dim_y, dim_out = self.conv_after_shape
dim_r = dim_x * dim_y * dim_out
self.reshape = View(-1, dim_r)
self.fc_layers, _ = self.create_linear_layers(dim_r, fc_args)
def create_layers(self, shape, conv_args=None, fc_args=None):
'''Creates layers
conv_args are in format (dim_h, f_size, stride, pad, batch_norm, dropout, nonlinearity, pool)
fc_args are in format (dim_h, batch_norm, dropout, nonlinearity)
Args:
shape: Shape of input.
conv_args: List of tuple of convolutional arguments.
fc_args: List of tuple of fully-connected arguments.
'''
self.conv_layers, self.conv_shape = self.create_conv_layers(shape, conv_args)
dim_x, dim_y, dim_out = self.conv_shape
dim_r = dim_x * dim_y * dim_out
self.reshape = View(-1, dim_r)
self.fc_layers, _ = self.create_linear_layers(dim_r, fc_args)
def create_layers(self,
shape,
crop_size=8,
conv_before_args=None,
res_args=None,
conv_after_args=None,
fc_args=None):
'''Creates layers
Args:
shape: Shape of the input.
crop_size: Size of the crops.
conv_before_args: Arguments for convolutional layers before residuals.
res_args: Residual args.
conv_after_args: Arguments for convolutional layers after residuals.
fc_args: Fully-connected arguments.
'''
self.crop_size = crop_size
dim_x, dim_y, dim_in = shape
self.final_size = 2 * (dim_x // self.crop_size) - 1
self.unfold = Unfold(dim_x, self.crop_size)
self.refold = Fold(dim_x, self.crop_size)
shape = (self.crop_size, self.crop_size, dim_in)
self.conv_before_layers, self.conv_before_shape = self.create_conv_layers(
shape, conv_before_args)
self.res_layers, self.res_shape = self.create_res_layers(
self.conv_before_shape, res_args)
self.conv_after_layers, self.conv_after_shape = self.create_conv_layers(
self.res_shape, conv_after_args)
self.conv_after_shape = self.res_shape
dim_x, dim_y, dim_out = self.conv_after_shape
dim_r = dim_x * dim_y * dim_out
self.reshape = View(-1, dim_r)
self.fc_layers, _ = self.create_linear_layers(dim_r, fc_args)
def handle_layer(self, block, shape, layer, layer_type):
'''Handles the layer arguments and adds layer to the block.
Args:
block: nn.Sequential to add modules to.
shape: Shape of the input.
layer: Layer arguments.
layer_type: Type of layer.
Returns:
tuple: Output shape.
'''
args = layer.pop('args', None)
if layer_type == 'linear':
if len(shape) == 3:
dim_x, dim_y, dim_out = shape
shape = (dim_x * dim_y * dim_out, )
block.add_module('flatten', View(-1, shape[0]))
bn = layer.get('bn', False)
bias = layer.pop('bias', None)
init = layer.pop('init', None)
init_args = layer.pop('init_args', {})
shape = self.add_linear_layer(block,
shape,
args=args,
bn=bn,
bias=bias,
init=init,
init_args=init_args)
elif layer_type == 'conv':
if len(shape) == 1:
shape = (1, 1, shape[0])
block.add_module('expand', Expand2d())
bn = layer.get('bn', False)
bias = layer.pop('bias', None)
init = layer.pop('init', None)
init_args = layer.pop('init_args', {})
shape = self.add_conv_layer(block,
shape,
args=args,
bn=bn,
bias=bias,
init=init,
init_args=init_args)
elif layer_type == 'tconv':
if len(shape) == 1:
raise ValueError('Transpose conv needs 4d input')
bn = layer.get('bn', False)
bias = layer.pop('bias', True)
shape = self.add_tconv_layer(block,
shape,
args=args,
bn=bn,
bias=bias)
elif layer_type == 'flatten':
if len(shape) == 3:
dim_x, dim_y, dim_out = shape
shape = (dim_x * dim_y * dim_out, )
block.add_module(layer_type, View(-1, shape[0]))
elif layer_type == 'reshape':
if args is None:
raise ValueError('reshape needs args')
new_shape = args
dim_new = 1
dim_out = 1
for s in new_shape:
dim_new *= s
for s in shape:
dim_out *= s
if dim_new != dim_out:
raise ValueError(
'New shape {} not compatible with old shape {}.'.format(
new_shape, shape))
block.add_module(layer_type, View((-1, ) + new_shape))
shape = new_shape[::-1]
elif layer_type is None:
pass
else:
raise NotImplementedError('Layer {} not supported. Use {}'.format(
layer_type, self._supported_types))
return shape