config_general¶
api_parser¶
The API parser controls the experiment by making objects and running trainers and testers. (The API parser is called by the Runner, which is the entry point to the program.)
By default, BaseAPIParser is used. If you use a custom trainer, the code will try to use API<name_of_your_trainer>, and if that doesn't exist, it will fall back to BaseAPIParser. However, if you explicitly set the api_parser option, then the one you specify will be used.
Default yaml:
api_parser: null
Example command line modification:
--api_parser {YourCustomAPIParser: {}}
trainer¶
The trainer trains your model.
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
Example command line modification:
# Swap in a different trainer, but keep the input parameters the same
--trainer~SWAP~1 {CascadedEmbeddings: null}
num_epochs_train¶
The maximum number of epochs to train for.
Default yaml:
num_epochs_train: 1000
Example command line modification:
--num_epochs_train 100
save_interval¶
Models will be evaluated and saved every save_interval epochs.
Default yaml:
save_interval: 2
Example command line modification:
--save_interval 10
patience¶
Training will end if the validation accuracy stops improving after patience+1 epochs.
Default yaml:
save_interval: 2
Example command line modification:
# Don't use patience at all
--patience null
check_untrained_accuracy¶
If True, then the tester will compute accuracy for the initial trunk (epoch -1) and initial trunk + embedder (epoch 0). Otherwise, these will be skipped.
Default yaml:
check_untrained_accuracy: True
Example command line modification:
--check_untrained_accuracy False
skip_eval_if_already_done¶
If True, then the tester will skip evaluation if a split/epoch has already been logged in the log files. If False, then the tester will evaluate a split/epoch regardless of whether it has already been done in the past. Previous logs will be preserved, hence the logs will contain duplicate results, and the most recent version for any split/epoch will be considered the "official" value for that split/epoch.
Default yaml:
skip_eval_if_already_done: True
Example command line modification:
--skip_eval_if_already_done False
skip_ensemble_eval_if_already_done¶
The same as skip_eval_if_already_done, but for ensembles.
Default yaml:
skip_ensemble_eval_if_already_done: True
Example command line modification:
--skip_ensemble_eval_if_already_done False
log_data_to_tensorboard¶
Set to False if you don't want to log data to tensorboard. You might want to do this if your disk I/O is slow.
Default yaml:
log_data_to_tensorboard: True
Example command line modification:
--log_data_to_tensorboard False
save_figures_on_tensorboard¶
Use matplotlib to plot things on tensorboard. (Most data doesn't require matplotlib.)
Default yaml:
save_figures_on_tensorboard: False
Example command line modification:
--save_figures_on_tensorboard True
save_lists_in_db¶
In record-keeper, non-scalar values are saved in the database as json-lists. This setting is False by default, because these lists can sometimes be quite large, causing the database file size to grow quickly.
Default yaml:
save_lists_in_db: False
Example command line modification:
--save_lists_in_db True
override_required_compatible_factories¶
Each APIParser comes with predefined compatible factories, which are used by default, regardless of what is specified in the factories config option. This allows you to specify a trainer without having to specify all the required factories. However, if you have your own custom factory that you know is compatible, and want to use that instead, you should set this flag to True.
Default yaml:
override_required_compatible_factories: False
Example command line modification:
--override_required_compatible_factories True