Modules

1. Data
   1. Obtain data from Kaggle A. Download
   2. Transform data A. Clean B. Put into appropriate task-specific format
   3. Split data A. Train B. Validate C. Test
2. Run 0. Initialization - Have defaults

   1. Train A. Prepare directories B. Build hooks C. Build model D. Build loss E. Build optimizer F. Load checkpoint G. Build scheduler H. Build dataloader/dataset I. Build summary writer J. Run train loop for epoch in range(last_epoch, config.train.num_epochs) # Train for dataloader in dataloaders: split = dataloader['split'] dataset_mode = dataloader['mode'] if dataset_mode != 'train': continue dataloader = dataloader['dataloader'] train_single_epoch(config, model, split, dataloader, hooks, optimizer, scheduler, epoch) # Validation score_dict = {} checkpoint_score = None for dataloader in dataloaders: split = dataloader['split'] dataset_mode = dataloader['mode'] if dataset_mode != 'validation': continue dataloader = dataloader['dataloader'] score = evaludate_single_epoch(config, model, split, dataloader, hooks, optimizer, scheduler, epoch) score_dict[split] = score if checkpoint_score is None: checkpoint_score = score # Update Learning Rates # Calling scheduler.step(**kwargs) if config.scheduler.name == 'ReduceLROnPlateau': scheduler.step(checkpoint_score) elif config.scheduler.name == 'CosineAnnealingLR': param_epoch = (epoch + 1) % config.scheduler.params.T_max scheduler.step(param_epoch + 1) elif config.scheduler.name != 'OneCycleLR' and config.scheduler.name != 'ReduceLROnPlateau': scheduler.step()

           # Checkpointing
           if checkpoint_score > best_checkpoint_score:
               best_checkpoint_score = checkpoint_score
               knlp.utils.save_checkpoint(config, model, optimizer, epoch, keep=None, name='best.score')
               knlp.utils.copy_last_n_checkpoints(config, 5, 'best.score.{:04d}.pth')
           if epoch % config.train.save_checkpoint_epoch == 0:
               knlp.utils.save_checkpoint(config, model, optimizer, epoch, keep=config.train.num_keep_checkpoint)
      

   2. Evaluate

   3. Inference

   4. SWA*

3. Submit

Configurations - Dataset - Name - Path - Splits - Transform - Model - Train - Evaluate - Optimizer - Scheduler - Hooks - Metric - Loss - Build Model - Post Forward - Write Result