Spaces:
Runtime error
Runtime error
| import jax | |
| import jax.numpy as jnp | |
| from transformers import FlaxBigBirdForQuestionAnswering, BigBirdTokenizerFast | |
| import gradio as gr | |
| FLAX_MODEL_ID = "vasudevgupta/flax-bigbird-natural-questions" | |
| if __name__ == "__main__": | |
| model = FlaxBigBirdForQuestionAnswering.from_pretrained(FLAX_MODEL_ID, block_size=64, num_random_blocks=3) | |
| tokenizer = BigBirdTokenizerFast.from_pretrained(FLAX_MODEL_ID) | |
| def forward(*args, **kwargs): | |
| return model(*args, **kwargs) | |
| def get_answer(question, context): | |
| encoding = tokenizer(question, context, return_tensors="jax", max_length=512, padding="max_length", truncation=True) | |
| start_scores, end_scores = forward(**encoding).to_tuple() | |
| # Let's take the most likely token using `argmax` and retrieve the answer | |
| all_tokens = tokenizer.convert_ids_to_tokens(encoding["input_ids"][0].tolist()) | |
| answer_tokens = all_tokens[jnp.argmax(start_scores): jnp.argmax(end_scores)+1] | |
| answer = tokenizer.decode(tokenizer.convert_tokens_to_ids(answer_tokens)) | |
| return answer | |
| default_context = "Models like BERT, RoBERTa have a token limit of 512. But BigBird supports up to 4096 tokens! How does it do that? How can transformers be applied to longer sequences? In Abhishek Thakur's next Talks, I will discuss BigBird!! Attend this Friday, 9:30 PM IST Live link: https://www.youtube.com/watch?v=G22vNvHmHQ0.\nBigBird is a transformer based model which can process long sequences (upto 4096) very efficiently. RoBERTa variant of BigBird has shown outstanding results on long document question answering." | |
| question = gr.inputs.Textbox(lines=2, default="When is talk happening?", label="Question") | |
| context = gr.inputs.Textbox(lines=10, default=default_context, label="Context") | |
| title = "BigBird-RoBERTa" | |
| desc = "BigBird is a transformer based model which can process long sequences (upto 4096) very efficiently. RoBERTa variant of BigBird has shown outstanding results on long document question answering." | |
| gr.Interface(fn=get_answer, inputs=[question, context], outputs="text", title=title, description=desc).launch() | |