iobt-vistec / octave Goto Github PK

Proposal

Need to add a logging method that supports image logging in training with Lightning.

Discussion

Logging is needed. Normally we can just use Lightning's built-in dashboard as usual. However, for Image logging for attention visualization. We need to convert the tensor into a PIL image. In which, I'm not sure about performance impact.

I'm currently considering the usage of visdom logger. But I'm not sure we can easily integrate it with lightning though.

Thank you for your extraordinary contributions. I have a few questions. There are too many "train_xxx.py" files here. If I only want to perform retinal vessel segmentation tasks using my own dataset, which "train" code should I use? Also, my experimental server does not have resource scheduling tools like "slurm". How should I make a modification to the train script to use LocalExecutory?

Please let me know if you need any further assistance with this issue.

Proposal

For the sake of readability. Henceforth I declared the prohibition of any Pytorch tensor manipulation method such as

• unsqueeze()
• squeeze()
• stack()
• sum()
• mean()
• min()
• etc.

Any of the existing Pytorch tensor manipulation methods that existed previously are subject to be changed.

Discussion

Why?

Because einops and einsum existed. That's all you gonna need.

Thank you for your wonderful work. But I got some questions for you. Could you please show us in detail how to train this model? For example, which folder should I put the dataset in? Or which file should I run to start training?
Thank you again, I am looking forward to your response!