no loss for training with small batch size,

[yikuanli]

When I train the Deepsurv model with small batch size (64), there is no loss and always early stop

However, if I train with larger batch size (1024), it almost always training properly

May I ask what might be the possible reason ? I am not familiar with the progress bar

[havakv]

That is very strange. I would always think a loss should be produced. Can you post a full example that reproduce the issue?

[yikuanli]

That is very strange. I would always think a loss should be produced. Can you post a full example that reproduce the issue?

thanks for the quick response, I have identified the issue, which is caused by having zero positive event when batch size is small. Therefore, dividing the number of positive events can explode the loss. I have added a small number to the denominator to avoid this in my local branch

[juancq]

@yikuanli can you share your solution here?

[Minxiangliu]

Hi @yikuanli , I also want to know.