HFValidationError when loading the model

Hi there! I am trying to load a model I have stored at Google Drive for inferencing:

# Load SetFit model
tuned_model = SetFitModel.from_pretrained("/content/drive/My Drive/models/tuned-model")
# Run inference
tuned_model(["i didnt feel humiliated", "i feel romantic too", "im grabbing a minute to post i feel greedy wrong"])

But I get the following error:

HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/content/drive/My Drive/models /tuned-model'. Use repo_type argument if needed.

it works however when I load it from the same script in which I have saved it, by:

# Save trained model to disk
trainer.model.save_pretrained("/content/drive/My Drive/models/tuned-model")

What can be the problem? Can’t just I save/load from pretrained to Google Drive?

Many thanks in advance for your support and terrific work.

Issue Analytics

State:
Created a year ago
Comments:10 (5 by maintainers)

Top GitHub Comments

1reaction

Mouhanedg56commented, Oct 19, 2022

I saved a trained model on local path. I can’t see anything wrong when loading the model using from_pretrained with correct path.

This error appears when you try to load a model from a nonexistent local path which have more than 1 backslash \ with local_files_only=True.

if we pass a nonexistent path:
- if path is in the form ‘repo_name’ or ‘namespace/repo_name’: ModelHubMixin.from_pretrained will throw FileNotFoundError.
- else: ModelHubMixin.from_pretrained will throw a HFValidationError because the path does not exist locally and is not in the expected hub form.
if we pass an existing local folder with no body: SentenceTransformer will throw an OSError: /path/to/your/model does not appear to have a file named config.json
if we pass an existing folder with a body and head: no issue in this case
if we pass an existing folder with a body with no head: SetFitModel._from_pretrained throws HFValidationError not catched by try/except. Since the expected behaviour in SetFit _from_pretrained classmethod is to initialise classification head with random weights when the MODEL_HEAD_NAME is not found, this can be considered as a bug.

@pdhall99 's suggestion looks fixing the issue and follows the same logic in the ModelHubMixin.from_pretrained:

if folder exists:
- if model head file exists => model_head_file = os.path.join(model_id, MODEL_HEAD_NAME) else None => initialise classification head with random weights
else: look for the model head in the hub using hf_hub_download.

1reaction

pdhall99commented, Oct 18, 2022

I get the same error. When I try to load a locally saved model:

from setfit import SetFitModel

model = SetFitModel.from_pretrained("/path/to/model-directory", local_files_only=True)

I get

HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/path/to/model-directory'. Use `repo_type` argument if needed.

I think this could be solved by changing these lines from

        if os.path.isdir(model_id) and MODEL_HEAD_NAME in os.listdir(model_id):
            model_head_file = os.path.join(model_id, MODEL_HEAD_NAME)

to something like

        if os.path.isdir(model_id):
            if MODEL_HEAD_NAME in os.listdir(model_id):
                model_head_file = os.path.join(model_id, MODEL_HEAD_NAME)
            else:
                model_head_file = None