Metric learning Experment

Overview

Baseline Model Training

In baseline model training, we use Source domain dataset ( ).

1. Weight Initialization: Cross-Entropy Loss

• Edit exp/sample_experiment/baseline_softmax/params.py

      params = {
      'n_epochs': 50,
      'n_class': 19,  # TODO
      'size': [224, 224],
      'batch_size': 64,
      'lr': 'default',
      'early_stopping': 5,
  
      'train_ds': '/path/to/train_ds',  # TODO
      'test_ds': '/path/to/test_ds',  # TODO
      'save_every_n_epoch': 1
  }

• Start training:

  python3 train_softmax.py <path/to/params.py>

• During training, to visualize loss and accuracy:

  tensorboard --logdir <path/to/params.py parent>

• After training complete:

  • Model Checkpoint directory is same as path/to/params.py

2. Baseline Feature Extractor: Triplet Loss

• Edit exp/sample_experiment/baseline_triplet/params.py

  params = {
      'n_epochs': 100,
      'n_class': 19,  # TODO
      'n_class_per_batch': 19,  # TODO
      'n_per_class': 10,  # TODO
      'size': [224, 224],
      'margin': 0.7,  # TODO
      'lr': 'default',
      'early_stopping': 20,
  
      'pretrained_weight': '/path/to/baseline_softmax/model',  # TODO
      'train_ds': '/path/to/train_ds',  # TODO
      'save_every_n_epoch': 1
  }

  ⚠️ Caution: pretrained_weight should left model in the last in order to read checkpoint properly

• During training, to visualize triplet loss, hardest negative distance (HND) and hardest positive distance (HPD):

  tensorboard --logdir <path/to/params.py parent>

• Start training:

  python3 train_softmax2triplet.py <path/to/params.py>

FSL Update Training

In FSL update training, we use Target domain dataset ( ).

• Edit exp/sample_experiment/fewshot-triplet/params.py

  params = {
      'n_epochs': 100,
      'n_class': 23,  # TODO
      'n_class_per_batch': 23,  # TODO
      'n_per_class': 10,  # TODO
      'size': [224, 224],
      'margin': 0.7,  # TODO
      'lr': 'default',
      'early_stopping': 20,
  
      'pretrained_weight': '/path/to/baseline_triplet/model',  # TODO
      'train_ds': '/path/to/train_ds',  # TODO
      'test_ds': '/path/to/train_ds',  # For export embeddings, could be same as train_ds
      'save_every_n_epoch': 1
  }

• During training, to visualize triplet loss, hardest negative distance (HND) and hardest positive distance (HPD):

  tensorboard --logdir <path/to/params.py parent>

• Start training:

  python3 train_triplet_fine_tune.py <path/to/params.py>

tags: FSL, Triplet loss