W&B로 ML 실험 관리와 추적

data:
    train_path: '/path/to/my/train.csv'
    valid_path: '/path/to/my/valid.csv'

model:
        encoder:
            dimension: 256
            dropout: 0.2
        decoder:
            dimension: 512
            dropout: 0.1

trainer:
    learning_rate: 1e-3
    train_batch_size: 16
    val_batch_size: 16

dataset:
    name: cifar10
    path: /home/datasets/cifar10

import hydra
from omegaconf import DictConfig

@hydra.main(config_path="config.yaml")
def my_app(cfg: DictConfig) -> None:
        print(cfg.pretty())

if __name__ == "__main__":
        my_app()

# To access the dataset key from the above example

cfg.dataset
# OR
cfg['dataset']

# DI
├── config
│   ├── config.yaml
│   ├── dataset
|                   ├── cifar10.yaml
│         ├── imagenet.yaml
└── train.py

defaults:
    - dataset: cifar10

$ python3 app.py dataset.cifar10.path=/downloads/datasets/cifar10

defaults:
  - paths: colab
  - trainer: default_trainer

  # disable logging by Hydra
  - override hydra/job_logging: none
  - override hydra/hydra_logging: none
  - _self_

model:
  model_name: roberta-base
  class_name: RobertaForSequenceClassification

run:
  exp_num: "00" # serial number of experiment
  name: baseline # a short 1-2 word description of the experiment
  experiment_name: ${run.exp_num}_${model.model_name}_${run.name}
  comment: "baseline"
  n_folds: 5
  fold: 0
  debug: False
  disable_warnings: True

wandb:
  enabled: True
  log_artifacts: True
  project: USPPPM-Kaggle
  notes: ${run.comment}
  name: fold_${run.fold}

├── config.yaml
├── paths
│   ├── colab.yaml
│   ├── jarvislabs.yaml
│   ├── kaggle.yaml
│   └── local.yaml
└── trainer
    ├── deberta.yaml
    ├── deberta_v3_base.yaml
    ├── deberta_v3_large.yaml
    ├── default_trainer.yaml
    ├── roberta_base.yaml
    └── roberta_large.yaml

#!/bin/bash

cd /home/US-Patent-Matching-Kaggle/src

for fold in 0 1 2 3 4
do
    python3 train.py \
    paths="jarvislabs" \
    trainer="deberta_v3_base" \
    run.debug=False \
    run.fold=$fold \
    run.exp_num="2" \
    model.model_name="microsoft/deberta-v3-base" \
    model.class_name="AutoModel" \
    run.name="baseline" \
    run.comment="folds grouped by anchors + add contexts + deberta v3 base with no modifications"
done

# Code taken from: https://docs.wandb.ai/guides/integrations/keras
import wandb
from wandb.keras import WandbCallback
wandb.init(config={"hyper": "parameter"})
...  # code to set up your model in Keras
# 🧙 magic
model.fit(X_train, y_train,  validation_data=(X_test, y_test),
          callbacks=[WandbCallback()])

wandb.alert(
    title=f"{experiment_name}", 
    text=f"FOLD {fold} --> MAE: {mae}"
)