protosleepnet.baselines.test_occlusion#

Channel occlusion robustness test for baseline 3ch models.

Tests how the original Phan models (SleepTransformer, SeqSleepNet) with 3 channels handle missing/occluded channels at inference time.

Evaluates on the test split with 5 scenarios (clean + 4 occlusion types). Supports SHHS (SleepTransformer) and MASS (SeqSleepNet) datasets.

Usage:

# SleepTransformer on SHHS python -m protosleepnet.baselines.test_occlusion –model_dir /path/to/sleeptransformer-phan-3ch –gpu_id 0

# SeqSleepNet on MASS python -m protosleepnet.baselines.test_occlusion –model_dir /path/to/seqsleepnet-phan-3ch –dataset mass –seq_len 20 –gpu_id 0

Attributes#

Wraps a model to apply channel occlusion (zeroing) at input level.

`build_dataset`(dataset_name, channels, pipeline, seq_len)	Build dataset, handling MASS multi-cohort case.
`compute_metrics`(all_proba, all_targets[, ignore_index])	Compute metrics from aggregated probabilities and targets.
`evaluate_subject`(model, inputs, L, device)	Sliding-window voting evaluation for a single subject.
`load_model`(model_dir, device)	Load model from config.json + model.pt.
`main`()

class protosleepnet.baselines.test_occlusion.ChannelOcclusionWrapper(model, mode=None, p=0.0, channels_to_occlude=None)#

Bases: torch.nn.Module

Wraps a model to apply channel occlusion (zeroing) at input level.

Modes:

None: no occlusion (passthrough) “random”: zero each channel independently with probability p per epoch

(at least 1 channel kept per sample per epoch)

“fixed”: zero specific channels for the entire night

protosleepnet.baselines.test_occlusion.build_dataset(dataset_name, channels, pipeline, seq_len)#: Build dataset, handling MASS multi-cohort case.

protosleepnet.baselines.test_occlusion.compute_metrics(all_proba, all_targets, ignore_index=-1)#: Compute metrics from aggregated probabilities and targets.

protosleepnet.baselines.test_occlusion.evaluate_subject(model, inputs, L, device)#: Sliding-window voting evaluation for a single subject.

protosleepnet.baselines.test_occlusion.load_model(model_dir, device)#: Load model from config.json + model.pt.

protosleepnet.baselines.test_occlusion.CLASS_NAMES = ['W', 'N1', 'N2', 'N3', 'REM']#