CNN + LSTM for Video Classification

V 0.2

1. view-> permute
2. Solve cross-border issues(def len(self): return len(self.data_lis)//self.seq_len)
3. There was a problem with the crop code before, now the data has been re-crop.
4. Find the reason of no drop in accuracy, Mainly because of data imbalance.
5. Fixed bug: losses.update(loss.item(), data.size(1)) & accuracies.update(acc, data.size(1)) # data.size(1) is batch-size, not data.size(0)
6. Add validation set code.
7. Remove the fully connected layer of Resnet

Note that: V0.2 used the thyroid data set of the medical examination department, and I manually duplicated the C classification to ensure that the data is balanced. What's more, the data of the validation set comes from the training set. So the next step is to focus on the production of the data set.

V 0.1

1. Implement seq_len in dataloader module of pytorch.
2. Implement classification prediction for consecutive video frames.

Environments

GPU: 2080 Ti

Cuda: 11.0

The more specific environment in environments.yaml

License

This project is licensed under the MIT License