[09 - PyTorch Model Deployment] Hugging Face URL Error for DEFAULT weights

[Bruno-Messias]

Hello Everyone!

Today I was trying to deploy the food vision model to Hugging Face, but I come up to a URL Error at hugging face logs that I dont understood. The error was showing me that the download link of PyTorch is not working for the DEFAULT weights of EfficientNet_B2_Weights. I manage to make work changing to IMAGENET1K_V1 weights, like that:

weights = torchvision.models.EfficientNet_B2_Weights.IMAGENET1K_V1

But nothing like that happened when I was coding at Google Colab.

Does anyone have an idea of what it could be? Or come with the same error?

[mrdbourke]

Hey hey @Bruno-Messias,

I'm not what could be happening here.

Looks like from the PyTorch documentation that .DEFAULT should work fine?

See: https://pytorch.org/vision/stable/models/generated/torchvision.models.efficientnet_b2.html#torchvision.models.EfficientNet_B2_Weights

The model builder above accepts the following values as the weights parameter. 

EfficientNet_B2_Weights.DEFAULT is equivalent to EfficientNet_B2_Weights.IMAGENET1K_V1. You can also use strings, e.g. weights='DEFAULT' or weights='IMAGENET1K_V1'.

These weights are ported from the original paper. Also available as EfficientNet_B2_Weights.DEFAULT.

But this error is only happening to you on Hugging Face not Colab?

[Bruno-Messias]

It's only happening in Hugging Face. On Colab all works fine.

[mrdbourke]

🤔

That is strange to me

Potentially Hugging Face are using a different version of PyTorch?

Does the Gradio app still run?

[Bruno-Messias]

I think I cath the error! Thanks the insigth!

The requirements was something different from the course, I was using this as requirement:

torch==1.13.0
torchvision==0.14.0

But at the course there are created as:

torch==1.12.0
torchvision==0.13.0

I just changed these requirements and the weights to DEFAULT and works!