"Beginner Example" Notebook shows no performance improvement after many epochs about semi-supervised-learning

[5v3D3]

If this issue is still open, from what I can tell, it is because the train function of the Trainer class is only calling train_step (which isn't calling paramupdate hook anymore). So you could add backward and step calls after train_step or just use the train function inherited from algorithmbase via algorithm.train() instead (this one calls all necessary hooks, I needed to manually send my model to gpu though).

[hijihyo]

If this issue is still open, from what I can tell, it is because the train function of the Trainer class is only calling train_step (which isn't calling paramupdate hook anymore). So you could add backward and step calls after train_step or just use the train function inherited from algorithmbase via algorithm.train() instead (this one calls all necessary hooks, I needed to manually send my model to gpu though).

I met the same problem, but 5V3D3's solution did help. When I used algorithm.train() instead of trainer.fit() the performance was improving. I used the following code:

algorithm.model = algorithm.model.cuda()
algorithm.train()

[Hhhhhhao]

Can provide more information? Which model are you using? What's the learning rate?

[SamShowalter]

Hi, the beginner notebook is exactly the same as is provided except for the parameters that I noted above. So the model is the vit_tiny_patch and the learning rate 5e-4 as provided. The learning rate looks like it could be low to me, but I figured that the provided example was tuned to train well.

[ppsmk388]

I meet the same problem

[hldqiuye123]

Hi, the beginner notebook is exactly the same as is provided except for the parameters that I noted above. So the model is the vit_tiny_patch and the learning rate 5e-4 as provided. The learning rate looks like it could be low to me, but I figured that the provided example was tuned to train well.

Have you solved it? "I have encountered the same problem, but after increasing the learning rate, there is still no change. The model consistently outputs the same results.".

[wzc32]

I've met the same!!
Do other notebooks run all right?

[SamShowalter]

Any update? Could someone who has it working provide their config file as well as the output performance to serve as a sanity check?

[sachinmotwani20]

algorithm.model = algorithm.model.cuda()
algorithm.train()

The above code works for me. Not sure how to perform the evaluation in this context. Tried the following.
algorithm.eval()

AttributeError: 'FixMatch' object has no attribute 'eval'

Any suggestions?

[Hhhhhhao]

Fixed in PR #135

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[github-actions]

This issue was closed because it has been stalled for 5 days with no activity.