`inference_loop` crashes on tensor shape mismatch when sequences have variable lengths

[ZahraSobhaninia]

Description

During evaluation (evaluate_finetune.py), inference_loop crashes with a RuntimeError when sequences in the dataset have different lengths:
RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 512 but got size 440

Suggestion fix:

def pad_patient(patient, max_length):   # for me for instance max_length=512 
    pad_fields = {
        "concepts": 0,
        "segments": 0,
        "ages": 0.0,
        "abspos": 0.0,
    }
    for field, pad_value in pad_fields.items():
        values = getattr(patient, field)
        if len(values) < max_length:
            padded = values + [pad_value] * (max_length - len(values))
        else:
            padded = values[:max_length]
        setattr(patient, field, padded)
    return patient

test_data.patients = [pad_patient(p) for p in test_data.patients]

[kirilklein]

@Montgomeryyyy could you have a look at this?

[kirilklein]

@ZahraSobhaninia Do you specify collate_fn in your config?

[Montgomeryyyy]

The eval script should use the EHRTrainer with dynamic padding as default. The used trainer_args are also printed when running the script, eg. 'Trainer args: {'save_every_k_steps': inf, 'collate_fn': <function dynamic_padding at 0x140c88400>, 'test_batch_size': 128, 'effective_batch_size': 128, 'batch_size': 128}'. Can you check which args are being used during the evaluation?

[mikkelfo]

@ZahraSobhaninia Is this still an issue?