def torchmoji_transfer()

in torchmoji/model_def.py [0:0]

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

def torchmoji_transfer(nb_classes, weight_path=None, extend_embedding=0,
                      embed_dropout_rate=0.1, final_dropout_rate=0.5):
    """ Loads the pretrained torchMoji model for finetuning/transfer learning.
        Does not load weights for the softmax layer.

        Note that if you are planning to use class average F1 for evaluation,
        nb_classes should be set to 2 instead of the actual number of classes
        in the dataset, since binary classification will be performed on each
        class individually.

        Note that for the 'new' method, weight_path should be left as None.

    # Arguments:
        nb_classes: Number of classes in the dataset.
        weight_path: Path to model weights to be loaded.
        extend_embedding: Number of tokens that have been added to the
            vocabulary on top of NB_TOKENS. If this number is larger than 0,
            the embedding layer's dimensions are adjusted accordingly, with the
            additional weights being set to random values.
        embed_dropout_rate: Dropout rate for the embedding layer.
        final_dropout_rate: Dropout rate for the final Softmax layer.

    # Returns:
        Model with the given parameters.
    """

    model = TorchMoji(nb_classes=nb_classes,
                     nb_tokens=NB_TOKENS + extend_embedding,
                     embed_dropout_rate=embed_dropout_rate,
                     final_dropout_rate=final_dropout_rate,
                     output_logits=True)
    if weight_path is not None:
        load_specific_weights(model, weight_path,
                              exclude_names=['output_layer'],
                              extend_embedding=extend_embedding)
    return model