New projects default to train.loss_weight_exp: 1.0, silently overriding the documented default of 0.25 #181

Summary

Every project created by initialize_project() gets train.loss_weight_exp: 1.0, which silently overrides the framework default of 0.25. Because of Hydra composition order, the project template always wins, so in practice no project ever trains with 0.25 unless the user edits project_config.yaml by hand or passes a command-line override.

Where

• deepethogram/conf/train.yaml:55 → loss_weight_exp: 0.25
• deepethogram/conf/project/project_config_default.yaml:46 → loss_weight_exp: 1.0

configuration.make_config() documents the composition order as:

1. Defaults
2. Preset
3. Project configuration
4. Command line

The project configuration is composed after the defaults, so 1.0 overrides 0.25 for every project.

Why it looks unintended

deepethogram/conf/tune/feature_extractor.yaml sweeps this parameter over [0.0, 1.0] with current_best: 0.25, which suggests 0.25 is the intended value.

Effect

data/utils.make_loss_weight() computes, per class:

pos_weight = (num_neg / num_pos) ** weight_exp

The exponent interpolates between no correction (0) and full inverse-frequency weighting (1). For a behavior at 1% prevalence:

That is a ~30x difference in how strongly rare-class positives are upweighted in BCEWithLogitsLoss, which is a substantial and silent change in training behavior for anyone with rare behaviors.

Suggested fix

Change deepethogram/conf/project/project_config_default.yaml to loss_weight_exp: 0.25 so the template matches the framework default. (Existing projects keep whatever is already written into their project_config.yaml.)

Version

master @ 1541eec (v0.4.0)