Yaml Syntax

Config files in this library are yaml files, so they follow standard yaml syntax. But you can also use all of the command line syntax within your config files. For example, here are two config files in the config_general category:

default.yaml

trainer:
  MetricLossOnly:
    iterations_per_epoch: 100
    dataloader_num_workers: 2
    batch_size: 32
    freeze_trunk_batchnorm: True
    label_hierarchy_level: 0
    loss_weights: null
    set_min_label_to_zero: True

num_epochs_train: 1000
save_interval: 2
patience: 9

check_untrained_accuracy: True
skip_eval_if_already_done: True
skip_ensemble_eval_if_already_done: True
save_figures_on_tensorboard: False
save_lists_in_db: False
override_required_compatible_factories: False

with_daml.yaml

trainer~SWAP~1:
  DeepAdversarialMetricLearning:


trainer~APPLY~2:
  g_alone_epochs: 0
  metric_alone_epochs: 0
  g_triplets_per_anchor: 100
  loss_weights:
    metric_loss: 1
    synth_loss: 0.1
    g_adv_loss: 0.1
    g_hard_loss: 0.1
    g_reg_loss: 0.1

The with_daml config file contains the special flags ~SWAP~ and ~APPLY~. This particular config file won't work by itself. However, it can be loaded in conjunction with default at the command line:

--config_general [default, with_daml]

This loads default.yaml, followed by with_daml.yaml. Now the special ~SWAP~ and ~APPLY~ flags will have an effect. Specifically, MetricLossOnly will get swapped out for DeepAdversarialMetricLearning, and then the parameters for DeepAdversarialMetricLearning will be applied to the trainer dictionary. The final config file ends up looking like this:

trainer:
  DeepAdversarialMetricLearning:
    iterations_per_epoch: 100
    dataloader_num_workers: 2
    batch_size: 32
    freeze_trunk_batchnorm: True
    label_hierarchy_level: 0
    loss_weights:
      metric_loss: 1
      synth_loss: 0.1
      g_adv_loss: 0.1
      g_hard_loss: 0.1
      g_reg_loss: 0.1
    set_min_label_to_zero: True
    g_alone_epochs: 0
    metric_alone_epochs: 0
    g_triplets_per_anchor: 100

num_epochs_train: 1000
save_interval: 2
patience: 9

check_untrained_accuracy: True
skip_eval_if_already_done: True
skip_ensemble_eval_if_already_done: True
save_figures_on_tensorboard: False
save_lists_in_db: False
override_required_compatible_factories: False