Skip to content

Latest commit

 

History

History
163 lines (122 loc) · 5.73 KB

README.md

File metadata and controls

163 lines (122 loc) · 5.73 KB
Raw

NEW Release 20230312!!!

update dropout Dropout Reduces Underfitting

paper

git

Usage

CHECK PLZ run_train_gpu.sh and ./models/dense_model/

1. make drop_scheduler.py

2. make dropout scheduling function your each model classes

Example

# 1) update dropout function
def update_dropout(self, drop_rate):
    self.drop_rate = drop_rate
    for module in self.modules():
        if isinstance(module, nn.Dropout):
            module.p = drop_rate

# 2) we have to count each step (for calculate schedulering global steps)
def on_train_start(self):
    from models.dense_model.drop_scheduler import drop_scheduler

    self.drop_scheduler = {}
    if self.args.dropout_p > 0.0:
        self.drop_scheduler["do"] = drop_scheduler(
            self.args.dropout_p,
            self.args.max_epochs,
            self.trainer.num_training_batches,
            self.args.cutoff_epoch,
            self.args.drop_mode,
            self.args.drop_schedule,
        )
        print(
            "on_train_start :: Min DO = %.7f, Max DO = %.7f"
            % (min(self.drop_scheduler["do"]), max(self.drop_scheduler["do"]))
        )

# 3) Finally, you can scheduling dropout prob in your training_step
if "do" in self.drop_scheduler:
        dropout_p = self.drop_scheduler["do"][self.trainer.global_step]
        self.update_dropout(dropout_p)
        self.log("dropout_p", dropout_p, sync_dist=(self.device != "cpu"))

3. we have to input dropout schedule setting in training_arg

4. replace run script

--dropout_p=0.1 \
--cutoff_epoch=1 \
--drop_mode=standard \
--drop_schedule=constant

If you want to use, normal style dropout, input dropout_p and drop_mode=standard (default) and drop_schedule=constant (default)

you can check your dropout scheduling process in wandb

image

pytorch-lightning-template

very simple but, write down is boring
boring boiling code rolling ⚡
If you need some function or someting, plz comment issues (plz write eng or ko). I reply and implement ASAP!!

WanDB

https://docs.wandb.ai/v/ko/quickstart

Training Detail

  • Using DDP, Not DP or CPU
    Maybe want to using DP or CPU, Change some argument or python Script
    See more detail: PyTorch-Lightning Dev Guide
  • Optimizer: AdamW
  • Monitoring Tool: WanDB

Pytorch-Lightning Life Cycle

Training

  1. train.py(main) -> argparse
    • using simple_parsing library looks like HFArgumentParser
    • Trainer Argument placed with pl.Trainer.add_argparse_args (automatic define argparse)
  2. def] WandbLogger, set seed(os, random, np, torch, torch.cuda)
  3. def] CustomDataModule (LightningDataModule)
    • You Not have to using LightningDataModule. but, if you implement that in 'LightningModule', source code is looked mess
    • DataModule important prepare_data and setup
      • prepare_data is only run on cpu and not multi processing (Warning, if you using distributed learning, this place's variable is not share)
        • I recommand, It just using data download or datasets save
      • setup is run on gpu or cpu and distributed. using map or dataload or something!
        • setup can have stage(fit (train), test, predict)
    • DataModule can have each stage's dataloader
      • using default or someting
    • Dataset can define this section or making each python script and just import & using!
  4. def] CustomNet (LightningModule)
    • each step and step_end or epoch, epoch_end
    • i think using just training_step, validation_step, validation_epoch_end is simple and best
      • training_step -> forward -> configure_optimizers
      • when count in each validation step (each batch step validation) -> validation_epoch_end (all batch result gather) -> log (on wandb)
  5. wandb logger additional setting
  6. checkpoint setting
    • monitor name is same on your each step's log name
  7. learning_rate monitor setting
  8. ddp strategy modify
    • if your dataset is so big to ddp, timeout parameter change like that
    • huggingface is so hard to make it. but lightning is feel free
  9. make trainer to your arg
  10. training run and model save!

Training Script Usage

  1. cd your project root(./pytorch-lightning-template)
# Don't Script RUN in your scripts FOLDER!!!!! CHK PLZ!!!!!!!
bash scripts/run_train_~~~.sh

Inference

  1. inference.py(main) -> argparse
  2. set seed
  3. model load (second param is your model init param)
  4. simply torch inference & END!

Inference Script Usage

  1. cd your project root(./pytorch-lightning-template)
# Don't Script RUN in your scripts FOLDER!!!!! CHK PLZ!!!!!!!
bash scripts/run_inference~~~.sh

(Optional) Install DeepSpeed

  1. run pip_install_deepspeed.sh
bash pip_install_deepspeed.sh