How to avoid overfitting

[Fly2flies]

Hello,
ZheDong, thanks for you sharing such a good work. I want to reproduce it in Pytorch,but I'm sorry that I encountered the overfitting problem.
To get the results quickly, I randomly choose 10,000 samples as traindata and 1,000 as valdata, 1,000 as testdata separately. Finally I got about 100% recall@5 on the training set while only half of it on the val data.
And I'm a fresh man to ImageTextEmbedding,could you share some solutions to that. I guess there are relevant reasons:

1. Data normalization. I don't compute the mean and var of train_data explicitly, and just divide it by 255, subtract 0.5, and thendivide it by 0.5

2. L2 regularization. I just use the 1e-5 regularization intensity

3. The complexity of classifier. After generator, I add a classifier with a softmax layer directly. Whether more fully connection layers can slow down the fitting of the training set

Finally, I want to ask how to mine the hard triplet online in Pytorch efficiently.

Thanks.

[layumi]

Thank you @EternallyTruth

• Have you tried to add dropout layer?
• You may consider to use larger learning rate.
• You may check my another repo about triplet loss in Pytorch (https://github.com/layumi/Person-reID-triplet-loss).