- Step 1 in inspecting their code: let's see what's their data format before feeding to the model
left_data: just images (batch x color_channel x 32 x 32)
right_data: (batch x num_expert x num_classes), they are one-hot vectors
label: just class labels, (batch)

- Okay, we are gonna try with extracted feature.
```
python3 main.py --case 1 --expertise 1 --path /scratch/tri/noisy_labels/data/cifar10/raw/ --device 1
```

- AHHH, now we have to tune lr to make this method work. What a waste of time! 
Next step: try several learning rates with an easy permuted confusion matrix.

