where to get index_to_name.json

[ppengkkang]

Could you let me know where to get index_to_name.json

[guidel96]

Hi,
I'm not sure you need it. You can use the model.config.id2label[<index_of_the_class_you_want>] to map the label/class index with its actual name.
For instance, you do the inference:

with torch.no_grad():
    logits_output = predicted_class_id = model(
        # take the input, which contains 'input_ids', 'attention_mask', and 'token_type_ids'
        input_ids = inputs['input_ids'].to(device),             
        token_type_ids = inputs['token_type_ids'].to(device)
        ).logits #we're only interested in the logits, so keep only those

And then you either just extract the predicted class with:

predicted_class_id = logits_output.argmax().item()
model.config.id2label[predicted_class_id]   # map the id of the prediction to the corresponding label word

Or you can also output probability of the sentence to be classified in each class:

probabilities = torch.nn.functional.softmax(logits_output, dim=1)       # Computing probabilities for each class

proba_per_class = list(zip(
    probabilities.flatten().numpy(),        # is a 2D tensor that we flatten to a 1D tensor, then convert to numpy
    model.config.id2label.values(),         # e.g. 'GHG_emissions' # model.config.id2label is a dict with value (class name, 'GHG_emissions')
    model.config.id2label.keys()            # and key (class number, e.g. 1)
    ))

print(proba_per_class)