def load_model_from_params_file()

in lib/utils/checkpoints.py [0:0]

• 26 lines of code
• 12 McCabe index (conditional complexity)

def load_model_from_params_file(model):
    """
    case 1: CHECKPOINT.RESUME = False and TRAIN.PARAMS_FILE is not none:
        load params_file

    case 2: CHECKPOINT.RESUME = True and TRAIN.PARAMS_FILE is not none:
        case 2a: if checkpoint exist: use checkpoint
        case 2b: if checkpoint not exist: use params_file

    case 3: CHECKPOINT.RESUME = True and TRAIN.PARAMS_FILE is none:
        case 3a: if checkpoint exist: use checkpoint
        case 3b: if checkpoint not exist: set start_model_iter = 0
    """

    use_checkpoint = cfg.CHECKPOINT.RESUME and find_checkpoint()

    if cfg.TRAIN.PARAMS_FILE and not use_checkpoint:
        logger.info('Initializing from pre-trained file...')
        start_model_iter, prev_lr = initialize_params_from_file(
            model=model, weights_file=cfg.TRAIN.PARAMS_FILE,
            load_momentum=False,  # not load momentum if it is pretrained
        )
        logger.info(('Loaded: start_model_iter: {}; prev_lr: {:.8f}').format(
            start_model_iter, prev_lr))
        model.current_lr = prev_lr

        # correct start_model_iter if pretraining uses a different batch size
        # (mainly used for 1-node warmup)
        if cfg.TRAIN.RESUME_FROM_BATCH_SIZE > 0:
            start_model_iter = misc.resume_from(start_model_iter)

        # if we only want the weights
        if cfg.TRAIN.RESET_START_ITER:
            start_model_iter = 0

    elif use_checkpoint:
        logger.info('Initializing from checkpoints...')
        start_model_iter, prev_lr = initialize_params_from_file(
            model=model, weights_file=get_checkpoint_resume_file())
        logger.info(('Loaded: start_model_iter: {}; prev_lr: {:.8f}').format(
            start_model_iter, prev_lr))
        model.current_lr = prev_lr
    else:  # no checkpoint, no params_file
        # Do nothing and return 0
        start_model_iter = 0
        logger.info('No checkpoint found; training from scratch...')

    return start_model_iter