[Exp] Implement Wandb Techniques

[yehyunsuh]

What

Automation tool for experimenting hyperparameters & Save images

Why

Time is money. Save time.

TODO

• [x] Implement Wandb Sweep
• [x] Implement Wandb Visualization

[yehyunsuh]

https://docs.wandb.ai/guides/sweeps https://www.youtube.com/watch?v=WZvG6hwxUEw

• Example Code

  
  command:
  - python3
  - main.py
  - "--pixel_loss" 
  - "--angle_loss" 
  - "--progressive_erosion"
  - "--progressive_weight" 
  - "--pretrained" 
  - "--wandb" 
  - "--wandb_sweep"
  - "--no_image_save" 
  - "--dilate"
  - "68"
  - "--dilation_decrease" 
  - "2" 
  - "--dilation_epoch" 
  - "10" 
  - "--epochs" 
  - "340" 
  - "--output_channel" 
  - "6"
  - "--angle_loss_weight" 
  - "1000" 
  - "--decoder_channel" 
  - "[512,256,128,64,32]"
  - "--wandb_name"
  - "Sweep_Label6_D68-2_to2_AW-P1000+A_every10_chx2"

method: bayes project: joint-replacement entity: yehyun-suh metric: name: Best Angle Difference goal: minimize parameters: learning_rate: distribution: uniform min: 5e-5 max: 5e-4 angle_loss_weight: distribution: int_uniform min: 750 max: 1250 batch_size: distribution: int_uniform min: 12 max: 34 decoder_channel: values:

• [256,128,64,32,16]
• [384,192,96,48,24]
• [512,256,128,64,32]

  - How to create an agent
  ```shell
  wandb sweep sweep.yaml

[yehyunsuh]

https://greeksharifa.github.io/references/2020/06/10/wandb-usage/#image