Custom dataset

[ja16005]

Hi.
I was trying out few shot on custom dataset, My custom data has 17 categories and 6 images per category. created a class similar to MiniImageNet and used in it validation, i.e valset. But it is throwing me an error in line 114 as
loss = F.cross_entropy(logits, label)
File "/home/search_env/lib/python3.6/site-packages/torch/nn/functional.py", line 2056, in cross_entropy
return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)
File "/home/search_env/lib/python3.6/site-packages/torch/nn/functional.py", line 1869, in nll_loss
.format(input.size(0), target.size(0)))
ValueError: Expected input batch_size (5) to match target batch_size (75).

Am I missing something? Kindly guide

[yinboc]

I am not sure what the exact problem is, replacing mini with custom dataset should work. It seems the batch_size of logits and label did not match, could you check their shape?

[ja16005]

yes, the shapes are different ..
Shape of logits torch.Size([5, 5])
Shape of label torch.Size([75])

Are there any constraints for the custom dataset for the code to run? Should I tweak any parameters accordingly?
I used the following parameters
shot=5
query=15
train way=30
test way=5
I have used mini during training and custom dataset for validation which has 17 classes and 6 images per class.

[ja16005]

I referred to the paper and modified query=3 and shot=3 as per my custom training support. Now the code is working fine.

Thanks for the quick response! :)

[yinboc]

I remember there should not be constraints for custom dataset. You may check the intermediate shapes to see what the exact problem is. The shape of logits should be [n_way * n_query, n_way], and shape of label should be [n_way * n_query].