Is it possible to use a from_pretrained() method to respawn model and tokenizer from my own s3 bucket?

Beginners
1

I saved a DistilBertModel and a tokenizer with the help of save_pretrained() method. Now, when I load them locally using from_pretrained(’/path_to_distilbert_model’) everything works fine and as intended.
I need these models to be loaded from my own s3 bucket.
But when I try to load them using TFDistilBertModel.from_pretrained(‘s3://my_bucket/path_to_distilbert_model’) there is an error, stating that model configuration file cannot be found or that ‘vocab.txt’ cannot be found if I use DistilBertTokenizer.from_pretrained().
Is it possible to load a model from my own s3 bucket and what is the suggested way to do that? Thanks

2

Looking at the source code, I do not think it is possible to load models from external URLs other than the HuggingFace bucket.

That being said, why don’t you upload your model to the HuggingFace model hub?

3

Wanted to see if you’ve found any solution to this problem?

I’m having the same problem where I need to read a pretrained model via from_pretrained(“s3://bucket_name/model_folder”), which throws the same errors.
I think the reason is, based on the source code, os.path.isdir(“s3://bucket_name/model_folder”) returns False.

4

Any solution so far?

On S3 there is no such concept as a “folder” link. That could be a reason that providing a folder path is not working.

5

I am not sure if this is still an issue, but I came across this at stackoverflow when looking for storing my own fine-tuned BERT model artifacts somewhere to use during the inference.

It seems helpful, and I am assuming adding
AutoTokenizer.from_pretrained(tokenizer.name, config=tokenizer_config.json)
would solve the tokenizer artifacts, but I feel like other artifacts(vocab.txt etc.) might be problematic to add with the code provided.

So, the simple solution appeared to me just creating a model repo in Huggingface Hub(might be obvious to the experienced eye, but was new to me) and calling your own pre-trained or fine-tuned models from there like
tokenizer = AutoTokenizer.from_pretrained("username/repo_name")
model = AutoModel.from_pretrained("username/repo_name").

I think one can just use pt_model.push_to_hub("my-awesome-model") documented here or using git commit/push as usual to do that.

Of course, this assumes that storing the artifacts in S3 is no must.