how to print the training set accuracy in the json file

[ghty11468]

In the training process, there is only the validation set accuracy in the json file, how to print the training set accuracy in the json file

[Ezra-Yu]

set the model.head.cal_acc to True, refer to https://github.com/open-mmlab/mmclassification/blob/46af7d3ed2538dea3133c06250dd3683fd792e38/mmcls/models/heads/cls_head.py#L22

The following is an example:

model = dict(
    type='ImageClassifier',
    backbone=dict(
        type='ResNeSt',
        depth=50,
        num_stages=4,
        out_indices=(3, ),
        style='pytorch'),
    neck=dict(type='GlobalAveragePooling'),
    head=dict(
        type='LinearClsHead',
        num_classes=1000,
        in_channels=2048,
        loss=dict(
            type='LabelSmoothLoss',
            label_smooth_val=0.1,
            num_classes=1000,
            reduction='mean',
            loss_weight=1.0),
        topk=(1, 5),
        cal_acc=True))

[ghty11468]

Thanks for your answer, but this method seems to print the accuracy for each iteration. Is there any other method that prints the accuracy of each epoch?

[Ezra-Yu]

Maybe you can let your train dataset be the same as val dataset.

[ghty11468]

This requires two training sessions, and the two sessions do not correspond to each other. Is there any way to print both training dataset and validation dataset accuracy of epoch by training once.To get the graph in the paper. Looking forward to your reply

[Ezra-Yu]

set cal_acc=True and then write a hook to collect all the iters and ACC for each iter [(iter, ACC), (iter, ACC) ....], then calculate the overall accuracy of the training epoch.

Or just inheriting and rewriting the LoggerHook.

[ghty11468]

This method is difficult for myself. I will look for other methods afterwards. Thank you very much for your reply!