How does the loss function for grade task work (CNN-only)?

[omniaalwazzan]

Hi, Richard. Your code has inspired me greatly:)

I'm not sure how you trained the vgg_19_bn to classify the grade; as I can see in your customised CNN, the output nueoron is one (with a probability ranging from -3 to 3); correct me if I'm wrong. However, in the CSV file, the grade has three labels: [0,1,2]. So, as far as I understand, the cnn model's output shape will be = [batch size,C), which is a one probability in this code.

For instance in case the batch size =2:

model=vgg()
output = model(image)
output_size = torch.Size([2, 1])
output_values = [[0.09075058],[0.10227629]]

Though, this code snippet confuses me

python train_cv.py  --exp_name grad_15 --task grad --mode path --model_name path --niter 0 --niter_decay 50 --batch_size 8 --lr 0.0005 --reg_type none --lambda_reg 0 --act LSM **--label_dim 3** --gpu_ids 0

If the label dim = 3, how does the loss function work?

loss nll = F.nll loss(pred, grade) In this case, pred and grade have different shapes.

Is there anything I'm missing or don't understand?

Thanks in advance, Omnia

[Richarizardd]

Hi @omniaalwazzan

Output neuron is 1 for the Cox Loss, but 3 for the GBMLGG grading task. You can see here where the last layer gets defined, which will vary based on the task.

[omniaalwazzan]

Oh my bad 😥

I’ve read the network.py from start to end and didn’t notice this🤦🏻‍♀️

Thanks a lot for the prompt response.

[Richarizardd]

No worries! Happy to help :)