Hello,
I am trying to modify densenet for my use case by taking the output of a denseblock and concat. it with another while retaining the information. See the code below:
class DenseNet(nn.Module):
def __init__(self, growthRate, depth, reduction, bottleneck):
super(DenseNet, self).__init__()
nDenseBlocks = (depth - 4) // 3
if bottleneck:
nDenseBlocks //= 2
nChannels = 2 * growthRate
self.conv1 = nn.Conv2d(3, nChannels, kernel_size=3, padding=1,
bias=False)
self.dense1 = self._make_dense(nChannels, growthRate, nDenseBlocks, bottleneck)
nChannels += nDenseBlocks * growthRate
nOutChannels = int(math.floor(nChannels * reduction))
self.trans1 = Transition(nChannels, nOutChannels)
nChannels = nOutChannels
self.dense2 = self._make_dense(nChannels, growthRate, nDenseBlocks, bottleneck)
nChannels += nDenseBlocks * growthRate
nOutChannels = int(math.floor(nChannels * reduction))
self.trans2 = Transition(516, nOutChannels)
nChannels = nOutChannels
self.dense3 = self._make_dense(nChannels, growthRate, nDenseBlocks, bottleneck)
nChannels += nDenseBlocks * growthRate
self.trans3 = Transition(nChannels, nOutChannels)
def _make_dense(self, nChannels, growthRate, nDenseBlocks, bottleneck):
layers = []
for i in range(int(nDenseBlocks)):
if bottleneck:
layers.append(Bottleneck(nChannels, growthRate))
else:
layers.append(SingleLayer(nChannels, growthRate))
nChannels += growthRate
return nn.Sequential(*layers)
def forward(self, x):
x_shape = x.shape[2:]
out = self.conv1(x)
out_a = self.dense1(out) # A
# out_a = self.dropout(out_a)
print(f'Shape of out_a is {out_a.shape}')
out_b = self.trans1(out_a.clone())
print(f'Shape of out_b is {out_b.shape}')
out_c = self.dense2(out_b)
print(f'Shape of out_c is {out_c.shape}')
out_d = torch.cat([out_c.clone(), out_a], 1)
out_e = self.trans2(out_d.clone())
out_f = self.dense3(out_e)
out_g = torch.cat([out_f.clone(), out_d], 1)
return out_g
However, when I tried to print my model summary, feeding it image size 224 x 224 of channel size 3 ==> 3, 224, 224
from torchsummary import summary
summary(model, (3, 224, 224))
I get the following error
Shape of out_a is torch.Size([2, 216, 224, 224])
Shape of out_b is torch.Size([2, 108, 112, 112])
Shape of out_c is torch.Size([2, 300, 112, 112])
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-9-d3385e62fb19> in <module>
1 from torchsummary import summary
2
----> 3 summary(model, (3, 224, 224), 16)
~/opt/anaconda3/envs/ge-mac/lib/python3.6/site-packages/torchsummary/torchsummary.py in summary(model, input_size, batch_size, device)
70 # make a forward pass
71 # print(x.shape)
---> 72 model(*x)
73
74 # remove these hooks
~/opt/anaconda3/envs/ge-mac/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
530 result = self._slow_forward(*input, **kwargs)
531 else:
--> 532 result = self.forward(*input, **kwargs)
533 for hook in self._forward_hooks.values():
534 hook_result = hook(self, input, result)
<ipython-input-5-aa8ff12b625d> in forward(self, x)
88 out_c = self.dense2(out_b)
89 print(f'Shape of out_c is {out_c.shape}')
---> 90 out_d = torch.cat([out_c.clone(), out_a], 1)
91
92 # out_c_ = F.interpolate(out_c.clone(), (224,224), mode='bilinear', align_corners=True)
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 112 and 224 in dimension 2 at /Users/distiller/project/conda/conda-bld/pytorch_1579022036889/work/aten/src/TH/generic/THTensor.cpp:612```
After printing out the output shape, I could see there is a size mismatch,
Shape of out_a is torch.Size([2, 216, 224, 224])
Shape of out_b is torch.Size([2, 108, 112, 112])
Shape of out_c is torch.Size([2, 300, 112, 112])```
I thought of upsampling the out_c i.e using F.interpolate(out_x, desired_shape, mode='bilinear', align_corners=True)
but would this retain my information without altering it? if not, any other way to match the size for subsequent concatenation?
Many thanks for the help