pretrained_model setting in hparams.yaml has no effect

[djangodesmet]

🐛 Bug

When calling the load_from_checkpoint function to load a model from a checkpoint, the hparams.yml file located in the parent folder does not get taken into account. For example, the pretrained_model setting in hparams.yml has no effect.

To Reproduce

contents of hparams.yml:

activations: Tanh
batch_size: 4
class_identifier: regression_metric
dropout: 0.1
encoder_learning_rate: 1.0e-06
encoder_model: XLM-RoBERTa
final_activation: null
hidden_sizes:
  - 3072
  - 1024
keep_embeddings_frozen: true
layer: mix
layer_norm: false
layer_transformation: sparsemax
layerwise_decay: 0.95
learning_rate: 1.5e-05
loss: mse
nr_frozen_epochs: 0.3
optimizer: AdamW
pool: avg
pretrained_model: /home/jovyan/nllb-gpu/models/xlm-roberta-large
train_data:
  - data/1720-da.csv
validation_data:
  - data/wmt-ende-newstest2021.csv
  - data/wmt-enru-newstest2021.csv
  - data/wmt-zhen-newstest2021.csv

comet = load_from_checkpoint(os.path.join(model_path, "checkpoint", "model.ckpt"))`

The roberta model is located in models/xlm-roberta-large as indicated by pretrained_model but an error is thrown because it still expects the roberta model to be in xlm-roberta-large in root. This gives the following error message:

EnvironmentError(
OSError: Can't load tokenizer for 'xlm-roberta-large'. If you were trying to load it from 'https://huggingface.co/models', make sure you don't have a local directory with the same name. Otherwise, make sure 'xlm-roberta-large' is the correct path to a directory containing all relevant files for a XLMRobertaTokenizerFast tokenizer.

Expected behaviour

I would expect that the pretrained_model parameter is used to determine the location of the model.

This could be achieved by adding the hparams_file as an argument to the model_class.load_from_checkpoint function in models/__init__.py

def load_from_checkpoint(checkpoint_path: str) -> CometModel:
    """Loads models from a checkpoint path.

    Args:
        checkpoint_path (str): Path to a model checkpoint.

    Return:
        COMET model.
    """
    checkpoint_path = Path(checkpoint_path)

    if not checkpoint_path.is_file():
        raise Exception(f"Invalid checkpoint path: {checkpoint_path}")

    parent_folder = checkpoint_path.parents[1]  # .parent.parent
    hparams_file = parent_folder / "hparams.yaml"

    if hparams_file.is_file():
        with open(hparams_file) as yaml_file:
            hparams = yaml.load(yaml_file.read(), Loader=yaml.FullLoader)
        model_class = str2model[hparams["class_identifier"]]
        model = model_class.load_from_checkpoint(
            checkpoint_path, load_pretrained_weights=False, hparams_file=hparams_file
        )
        return model
    else:
        raise Exception(f"hparams.yaml file is missing from {parent_folder}!")

Environment

OS: Linux
Packaging: pip
Version 2.0.1

[ricardorei]

Hi @djangodesmet I added a bit more flexibility to load_from_checkpoint with my last push. To solve your case I added an argument reload_hparams.