Overfitting problem while implementing in Camelyon16

[pzSuen]

Hello,

I randomly select 90% and 10% of official train data as train-set and val-set. The test-set is official provided. But the gap between training and inference is big. Have you encountered this problem?

[Bingchao-Zhao]

Hello, @pzSuen I'm having some problems reproducing transmil. As mentioned in the article, the feature of patch is obtained from resnet50, but the output of resnet50 backbone is 2048, while the input of _fc1 of transmil is 1024. how can I get the right input vector?

[pzSuen]

As in the paper, you can find these related content from the paper "Data-efficient and weakly supervised computational pathology on whole slide images".

[Bingchao-Zhao]

@pzSuen Thanks for the reminder, you have solved a puzzle that has been bothering me for a long time.

[marvinyan080]

@pzSuen Have you been able to reproduce the results reported in the paper?

[Bingchao-Zhao]

I was able to get a performance close to the reported after extracting features using the network provided by CLAM.

@.***

From: marvinyan080 Date: 2022-08-09 13:28 To: szc19990412/TransMIL CC: Bingchao Zhao; Comment Subject: Re: [szc19990412/TransMIL] Overfitting problem while implementing in Camelyon16 (Issue #13) @pzSuen Have you been able to reproduce the results reported in the paper? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

[szc19990412]

Hi. Since the CAMELYON16 dataset is a public dataset for a competition, and all the results are obtained in the public test set, so we only reported the results with the highest folds. Indeed, the mean and variance should be used as a more reasonable criterion, and we will pay more attention to this issue in the future.