"name": "eeg",
"processing": []
}],
"split": {
"train": train_split,
"val": validation_split,
"test": None,
"parameters_init": train_split,
},
"hypnogram_filename": f"{scorer}.mm"
},
"save_folder": EXPERIMENT_OUTPUT_FOLDER,
}
folder = log_experiment(**experiment_to_log,
num_workers=num_workers,
generate_memmaps=False)
checkpoint = {"directory": folder, "net_to_load": "training/best_net"}
trainer_eval = {
"type": "base",
"args": {
"epochs": 0,
"patience": 0,
"num_workers": 0,
"optimizer": {
"type": "adam",
"args": {
"lr": 1e-3,
"amsgrad": True
}
},
for dataset in datasets:
train_split += datasets_split[dataset]["train"]
validation_split += datasets_split[dataset]["validation"]
test_split += datasets_split[dataset]["test"]
experiment_to_log = {
"memmap_description": memmaps_description[target_dataset],
"dataset_settings": settings[target_dataset],
"trainer_parameters": trainer_parameters,
"normalization_parameters": model_normalization,
"net_parameters": model_description,
"dataset_parameters": {
"temporal_context": TEMPORAL_CONTEXT,
"transform_parameters": [{
"name": "eeg",
"processing": []
}],
"split": {
"train": train_split,
"val": validation_split,
"test": test_split,
"parameters_init": train_split,
},
},
"save_folder": EXPERIMENT_OUTPUT_FOLDER,
}
folder = log_experiment(**experiment_to_log,
num_workers=num_workers,
generate_memmaps=False)
def run_experiment(settings,
memmaps_description,
temporal_context,
trainer,
normalization,
model,
split,
transform,
save_folder,
hypnogram_filename_train='hypno.mm',
hypnogram_filename_test='hypno.mm',
fold_to_run=None,
force=False,
error_tolerant=False,
checkpoint=None,
num_workers=0,
SEED=2019,
validate_with_ablation_modalities=None,
max_training_records=None):
if os.path.exists(save_folder) and force:
shutil.rmtree(save_folder)
description_hash = memmap_hash(memmaps_description)
print(settings["memmap_directory"])
h5_to_memmaps(
records=[
settings["h5_directory"] + record
for record in os.listdir(settings["h5_directory"])
],
memmap_description=memmaps_description,
memmap_directory=settings["memmap_directory"],
error_tolerant=error_tolerant,
)
dataset_dir = settings["memmap_directory"] + description_hash + "/"
available_dreem_records = [
dataset_dir + record + "/" for record in os.listdir(dataset_dir)
if ".json" not in record
]
# build the folds
rd.seed(SEED)
rd.shuffle(available_dreem_records)
assert split["type"] in ["loov", "kfolds"]
if split["type"] == "kfolds":
N_FOLDS = split["args"]["n_folds"]
if "subjects" not in split["args"]: # assumer record-wise split
folds = split_list(available_dreem_records, N_FOLDS)
else: # assume multiple record per subject and subject-wise split
subjects = []
for subject in split["args"]["subjects"]:
for record in subject["records"]:
if record in os.listdir(dataset_dir):
subjects += [subject]
break
subject_per_folds = split_list(subjects, N_FOLDS)
folds = []
for subjects in subject_per_folds:
record_in_fold = []
for subject in subjects:
for record in subject["records"]:
record_in_fold += [dataset_dir + record + "/"]
folds += [record_in_fold]
elif split["type"] == "loov":
# LOOV training
folds = [[record] for record in available_dreem_records]
else:
raise ValueError
if fold_to_run is None:
fold_to_run = [j for j, _ in enumerate(folds)]
outfolders = []
for i, fold in enumerate(folds):
if i in fold_to_run:
other_folds = [fold for k, fold in enumerate(folds) if k != i]
rd.seed(SEED + i)
rd.shuffle(other_folds)
n_val = max(1, int(len(other_folds) * 0.2))
train_folds, val_folds = other_folds[n_val:], other_folds[:n_val]
train_records = [
record for train_fold in train_folds for record in train_fold
]
val_records = [
record for val_fold in val_folds for record in val_fold
]
if max_training_records is not None:
rd.shuffle(train_records)
rd.shuffle(val_records)
record_to_use_for_training = min(
int(max_training_records * 0.7), max_training_records - 1)
record_to_use_for_validation = max_training_records - record_to_use_for_training
train_records = train_records[:record_to_use_for_training]
val_records = val_records[:record_to_use_for_validation]
experiment_description = {
"memmap_description": memmaps_description,
"dataset_settings": settings,
"trainer_parameters": trainer,
"normalization_parameters": normalization,
"net_parameters": model,
"dataset_parameters": {
"split": {
"train": train_records,
"val": val_records,
"test": fold
},
"temporal_context": temporal_context,
"transform_parameters": transform,
"hypnogram_filename": hypnogram_filename_train,
"hypnogram_filename_test": hypnogram_filename_test
},
"save_folder": f"{save_folder}",
}
outfolders += [
log_experiment(**experiment_description,
num_workers=num_workers,
generate_memmaps=False,
checkpoint=checkpoint,
validate_with_ablation_modalities=
validate_with_ablation_modalities)
]
return outfolders