[Big Update] Implement separate callbacks class

[kaylode]

Helps Trainer more generalized and easy to be modified/intergrated

[kaylode]

Hi @lannguyen0910 , welcome to the project. Can you help me test this branch? See whether the callbacks are simple to use and customize then give me some feedbacks/suggestion.

[kaylode]

There is a CALLBACK_REGISTRY in which contains all callbacks function that are registered in pipeline.yaml file. All new callback classes must be defined with the same method names as in theseus/base/callbacks/base_callbacks.py. Then these methods will be automatically called by the trainer while training.

[lannguyen0910]

@kaylode I have tested on the classification + segmentation on this branch.
In classification:

• the visualizer on tensorboard works fine.
• no warnings or errors triggered.
• the logger also works fine.
• The loading, training is fast, and good.
• The model works ok, although it reported more error cases afterward, you can check the SupervisedTrainer again to see whether it has any potential error, for me it looks fine.
• Besides, it also works good without the visualizer callback.

In segmentation:

• Apart from the callbacks, i think we have an error with segmentation_models.pytorch dependency on create_model method, i suggest using git+https://github.com/qubvel/segmentation_models.pytorch instead.
• I think we have an error in _load_mask func in CSVDataset, it reports RuntimeError: Class values must be smaller than num_classes. in one-hot encoding the masks. So i try the previous implement to see if it pass-through the error.

I think my attempt is wrong because the sanity check plot this.

Overall, the callbacks works good, no major problems, i will take a deep look in the code again to see if there is minor bugs or improvements, i will be testing on WandB too.

[kaylode]

@lannguyen0910

Apart from the callbacks, i think we have an error with segmentation_models.pytorch dependency on create_model method, i suggest using git+https://github.com/qubvel/segmentation_models.pytorch instead.

Yes, I've been experiencing this as well. Basically it is due to the version of timm models. I think your suggestion is fine, do you know how can we include that in setup.py ?

I think we have an error in _load_mask func in CSVDataset, it reports RuntimeError: Class values must be smaller than num_classes. in one-hot encoding the masks. So i try the previous implement to see if it pass-through the error.

Yes, this happened because we must define a different _load_mask() function for the specific dataset (CarvanaCar in this case). The default _load_mask() in the current settings works for masks in which each pixel present a class.

Thank you very much, I'm gonna merge this branch and if there are any more bugs we will fix that later.

p/s: if you want to comment on any parts of the template that you see inconvenient or unnecessary, or want to contribute improvement, just create a new branch and make a PR. :>

[kaylode]

Solves #19