def conv_out_size(self, h, w, c=None):
"""Helper function ot return the output size for a given input shape,
without actually performing a forward pass through the model."""
for child in self.block.children():
if isinstance(child, torch.nn.Conv2d):
h, w = conv2d_output_shape(h, w, child.kernel_size,
child.stride, child.padding)
c = child.out_channels
elif isinstance(child, torch.nn.MaxPool2d):
h, w = conv2d_output_shape(h, w, child.kernel_size,
child.stride, child.padding)
elif isinstance(child, ResidualBlock):
h, w, c = child.conv_out_size(h, w, c)
return h, w, c
def conv_out_size(self, h, w, c=None):
for child in self.conv.children():
try:
h, w = conv2d_output_shape(h, w, child.kernel_size,
child.stride, child.padding)
except AttributeError:
pass # Not a conv or maxpool layer.
try:
c = child.out_channels
except AttributeError:
pass # Not a conv layer.
return h * w * c
def conv_out_size(self, h, w, c=None):
"""Helper function ot return the output size for a given input shape,
without actually performing a forward pass through the model."""
for child in self.block.children():
try:
h, w = conv2d_output_shape(h, w, child.kernel_size,
child.stride, child.padding)
except AttributeError:
pass # Not a conv or maxpool layer.
try:
c = child.out_channels
except AttributeError:
pass # Not a conv layer.
return h, w, c
def __init__(self,
h,
w,
in_channels=3,
out_channels=16,
kernel=3,
stride=1,
maxstride=2):
super().__init__()
p = int((kernel - 1) // 2) # Same padding
base_conv = torch.nn.Sequential(
torch.nn.Conv2d(in_channels,
out_channels,
kernel_size=kernel,
stride=stride,
padding=p),
torch.nn.MaxPool2d(kernel, maxstride, p))
self.output_size = conv2d_output_shape(
h, w, kernel_size=kernel, stride=maxstride,
padding=p) # Output after maxpool
self.conv = torch.nn.Sequential(base_conv,
IMPALAResidualBlock(out_channels),
IMPALAResidualBlock(out_channels))