I built a custom transformer for a SequenceClassification task. Everything works fine but when I want to load my model using pipeline, I get wrong results.
model_ckpt = 'roberta-base'
tokenizer = AutoTokenizer.from_pretrained(model_ckpt)
labels =['health','business', 'politics', 'entertainment', 'sports', 'tech']
index2cat = {idx:cat for idx, cat in enumerate(labels)}
cat2index = {cat: idx for idx, cat in enumerate(labels)}
Here is my custom SequenceClassification model
Trainer and training args
Using pipeline
Calling pipeline with the task, model and tokenizer gives the correct results but with the model ID on hub or local directory, I get wrong results. See sample below
I have been on this for few days now and still cannot have a breakthrough at solving it. Thank you