The task at work is to make a CNN model to do some classification task on images. Also, I should be able to view the feature maps after making a classification on an image i.e the images obtained after applying convolution or pooling operations. Below is how I defined my CNN class:
class ConvNet(nn.Module):
def __init__(self, input_channels, output_dim):
super().__init__()
# input 48x48
self.architecture = {
"conv1": self.convblock(input_channels, 128, (3,3)), # 46x46
"conv2" : self.convblock(128, 64, (3, 3), bnorm=True), # 44x44
"pool1" : self.poolblock((2,2)), # 22x22
"conv3" : self.convblock(64, 16, (3,3), stride=2), #10x10
"conv4" : self.convblock(16, 10, (3,3)), # 8x8
"pool2" : self.poolblock((2,2), bnorm=10), # 4x4
"feedforward" : nn.Sequential(
nn.Flatten(), # 4x4x10 = 160
nn.Linear(160, 128), # 128
nn.ReLU(inplace=True),
nn.Dropout(0.3),
nn.Linear(128, output_dim), # 3
nn.Softmax(dim=1)
)
}
self.maps = {}
def forward(self, x):
image = x
for name, layer in self.architecture.items():
out = layer(image)
self.maps[name] = out
image = out
return image
def convblock(self, inp, out, kernel, stride=1, bnorm=False):
if bnorm:
return nn.Sequential(
nn.Conv2d(inp, out, kernel, stride=stride),
nn.ReLU(inplace=True),
nn.BatchNorm2d(out)
)
else:
return nn.Sequential(
nn.Conv2d(inp, out, kernel, stride=stride),
nn.ReLU(inplace=True)
)
def poolblock(self, kernel, bnorm=None):
if bnorm is None:
return nn.MaxPool2d(kernel)
else:
return nn.Sequential(
nn.MaxPool2d(kernel),
nn.BatchNorm2d(bnorm)
)
def get_map(self, im, layer):
fig, ax = plt.subplots(1,2, figsize=(20,10), gridspec_kw={'width_ratios': [1,3]})
ax[0].set_xticks([])
ax[0].set_yticks([])
ax[0].imshow(im.reshape(im.shape[-2],im.shape[-1],1), cmap="gray") # Shows Input image
self(im)
map = self.maps[layer]
map=map.reshape(map.shape[1],1,map.shape[-2],map.shape[-1])
ax[1].set_xticks([])
ax[1].set_yticks([])
rows = max(int(map.shape[0]/8), 8)
ax[1].imshow(make_grid(map,nrow=rows).permute(1, 2, 0)) # Shows all the channels after an operation.
The idea is to store blocks of convolution and pooling layers in the self.architecture dictionary with names such as ‘conv1’, ‘conv2’, ‘pool1’ etc…
Then, in the forward method, I would run the input images through each block and also store the outputs of each block in the self.maps dictionary to retrieve later (self.get_map does this).
The problem is that the model’s parameters are not being set properly. Below is the code where I instantiated the model and optimizer:
model = ConvNet(1, 3)
adam_opt = torch.optim.Adam(model.parameters(), lr=0.01)
However I ran into the following error:
/usr/local/lib/python3.10/dist-packages/torch/optim/optimizer.py in __init__(self, params, defaults)
271 param_groups = list(params)
272 if len(param_groups) == 0:
--> 273 raise ValueError("optimizer got an empty parameter list")
274 if not isinstance(param_groups[0], dict):
275 param_groups = [{'params': param_groups}]
ValueError: optimizer got an empty parameter list
I printed out the parameter list and its empty. I don’t understand why this is happening.
Is my way of defining the architecture or forward method wrong and pytorch expects some specific behaviour when subclassing nn.Module? If so what is it and how can I change my class? Any extra information on how pytorch actually stores the class parameters is most welcomed.
Thank you.