Training

With sadaco installed, we can now start training on prepared datasets. Here, we provide example pipeline on ICBHI dataset.

First, import sadaco library, and build a trainer instance that will help us train a deep model.

In order to build a trainer, we have to first load a trainer configuration written in yaml format.

YAML template example

Master Config

# TODO: We are planning to modify this structure with inheritance feature enabled, as in the detectron2(https://github.com/facebookresearch/detectron2/blob/main/configs/COCO-Detection/fast_rcnn_R_50_FPN_1x.yaml)
prefix : <PREFIX>
project_name : <PROJECT NAME> #This will be used as a wandb project name
use_wandb : !!bool True
data_configs :
  file : <DATA_CONFIG_FILE_PATH>
  split : [train, test]
model_configs:
  file : <MODEL_CONFIG_FILE_PATH>
  resume : False # If True, model will resume to the .pth file specified in MODEL_CONFIG
  resume_optimizer : False # If True, optimizer will also resume. Else, only model resumes.

output_dir : checkpoints/

train: # Training Pipeline Configuration
  method : basic # TODO: Currently not handled by the trainer. 
  target_metric : F1-Score # Target metric that will be used to determine the best model.
  max_epochs : !!int 650
  save_interval : !!int 1 # Model saving interval. Only saves the last if -1.
  update_interval : !!int 6 # Gradient Accumulation interval.
  criterion:
      name: CELoss
      loss_mixup : !!bool True
      params : 
          mode : onehot
          reduction : mean
  optimizer:
      name: Adam
      params:
      lr : !!float 3e-6
      weight_decay : !!float 5e-7
      betas : !!python/tuple [0.95, 0.999]
  lr_scheduler:
      name : CosineAnnealingWarmUpRestarts
      params :
      T_0: !!int 40
      T_mult : !!int 1
      eta_max : !!float 5e-4
      T_up : !!int 10
      gamma: !!float 1.


data:
  train_dataloader:
      sampler : 
          name : BalancedBatchSampler
          params:
              n_classes : 4
              n_samples : 10
      params : 
          shuffle : True
          batch_size : 128
          num_workers : 8
          pin_memory : True
          persistent_workers : True
          drop_last : False
  val_dataloader:
      params : 
          shuffle : False
          batch_size : 16
          num_workers : 8
          pin_memory : True
          persistent_workers : False
          drop_last : False

import sadaco
from sadaco.utils import config_parser

master_config_path = '<MASTER_CONFIG_PATH>'
my_configs = config_parser.parse_config_obj(yml_path=master_config_path)

config_parser is designed to return configs in python object, so that we can access sub-configs like foo.bar. If you want to get an dictionary way of configs, simply call my_configs.__dict__ or, change config_parser.parse_config_obj to config_parser.parse_config_dict.

After loading the configuration, we can now create a trainer instance with our loaded configs as follows.

trainer = ICBHI_Basic_Trainer(my_configs)

Once the trainer is initialized, all we have to do is then to call train method to proceed training.

trainer.train()

or, if you want to conduct a k-fold cross validation, simply call

trainer.train_kfold(FOLD_NUM)

You can see the full example of the training pipeline in the main() routine from sadaco/pipelines/ICBHI