Add Weights and Baises Integration

[ayulockin]

This PR adds support for Weights and Biases Metric, model checkpointing and evaluation logging.

Usage

I have added the functionality to sr.py, sample.py and infer.py files. To enable logging to W&B pass -enable_wandb.

Super Resolution

• Train: python sr.py -p train -c config/sr_sr3_16_128.json -enable_wandb -log_wandb_ckpt
  Training and validation metrics, generated sr image and model checkpoint are logged as shown.

  Screen.Recording.2022-01-13.at.12.38.56.AM.mov

• Evaluate: python sr.py -p val -c config/sr_sr3_16_128.json -enable_wandb -log_eval
  Below you can see the use of W&B Tables to log the result of evaluation.

  Screen.Recording.2022-01-13.at.12.57.01.AM.mov

Unconditional Image Generation

• Train: python sample.py -p train -c config/sample_sr3_128.json -enable_wandb -log_wandb_ckpt
  The resulting W&B run page will look similar to super-resolution one.

• Evaluation: python sample.py -p train -c config/sample_sr3_128.json -enable_wandb
  Since the task is image generation, the generated images will be shown as a W&B image panel as shown.

  

Inference

To run inference python infer.py -c config/sr_sr3_16_128.json -enable_wandb -log_infer. This will give a table similar to evaluation but without psnr and ssim scores.

Notes

• If -enable_wandb is not passed, the scripts will run without any issue (normally).
• The major code related to W&B is in the core/wandb_logger.py file.
• The screenshots and recordings are based on the results that I got by using -debug argument. However, I have tested the implementation on full training.

I would love to know what you think of this and hope this will add value to your useful repository.

[Janspiry]

This PR adds support for Weights and Biases Metric, model checkpointing and evaluation logging.

Usage

I have added the functionality to sr.py, sample.py and infer.py files. To enable logging to W&B pass -enable_wandb.

Super Resolution

• Train: python sr.py -p train -c config/sr_sr3_16_128.json -enable_wandb -log_wandb_ckpt
  Training and validation metrics, generated sr image and model checkpoint are logged as shown.

  Screen.Recording.2022-01-13.at.12.38.56.AM.mov
  

• Evaluate: python sr.py -p val -c config/sr_sr3_16_128.json -enable_wandb -log_eval
  Below you can see the use of W&B Tables to log the result of evaluation.

  Screen.Recording.2022-01-13.at.12.57.01.AM.mov
  

Unconditional Image Generation

• Train: python sample.py -p train -c config/sample_sr3_128.json -enable_wandb -log_wandb_ckpt
  The resulting W&B run page will look similar to super-resolution one.
• Evaluation: python sample.py -p train -c config/sample_sr3_128.json -enable_wandb
  Since the task is image generation, the generated images will be shown as a W&B image panel as shown.
  

Inference

To run inference python infer.py -c config/sr_sr3_16_128.json -enable_wandb -log_infer. This will give a table similar to evaluation but without psnr and ssim scores.

Notes

• If -enable_wandb is not passed, the scripts will run without any issue (normally).
• The major code related to W&B is in the core/wandb_logger.py file.
• The screenshots and recordings are based on the results that I got by using -debug argument. However, I have tested the implementation on full training.

I would love to know what you think of this and hope this will add value to your useful repository.

Thanks so much for this great work, which is so awesome! and I will try this amazing tool and check it carefully soon.
Could you edit the README.md to add Weights and Biases part or add the link to some files like descriptions as mentioned above.

[ayulockin]

I will updated the README. Thank you for your feedback. 🙏

[kada0720]

@Janspiry
I have a question, why are all the images generated using 'Unconditional Image Generation' pure noise？