Accuracy of CAN Method about libfewshot

[wZuck]

Thank you for reporting this issue. You can modify line 327 to

acc = accuracy(cls_scores, query_targets.view(-1), topk=1)

We will update can.py later.

And CAN uses episodic-tasks in both the training and testing stage, so batch_size is useless in this method, you can set episode_size for training/testing/debugging.

If you are using the main branch of LibFewShot, you will get the episode_size = episode_size. If you are using the add_ddp branch of LibFewShot, you will only get the episode_size = episode_size / n_gpu.

[viviGeng]

Thank you for your reply.
I have another question about the shape of input data with different parameter setting.
For training CAN model using miniImageNet, set batch_size = 32 episode_size = 1 n_gpu = 1, then the shape of input data is [100,3,84,84]. If we change the setting as batch_size = 128 episode_size = 1 n_gpu = 1, then the shape of input data is [80,3,84,84].
It seems the first element of shape array is somehow relevant to batch_size. Could you give me some suggestion about how to set those parameters appropriately.

[wZuck]

Could you upload your config of 2 different settings? It's better if you can upload the config.yaml in the checkpoint's directory.

CAN-miniImageNet--ravi-Conv64F-5-1-XXXXXXXXXXXX
├── checkpoints
├── config.yaml        # <- this file
└── log_files

You may set shot/test_shot to different values accidentally. Beacause 100 = 5way * (5shot+15query) and 80 = 5way * (1shot+15query)

[viviGeng]

Thank you for the explanation.
The difference is caused by value of shot.