import gradio as gr from transformers import GPT2Tokenizer, AutoModelForCausalLM import numpy as np MODEL_NAME = "gpt2" if __name__ == "__main__": # Define your model and your tokenizer tokenizer = GPT2Tokenizer.from_pretrained(MODEL_NAME) model = AutoModelForCausalLM.from_pretrained(MODEL_NAME) if tokenizer.pad_token_id is None: tokenizer.pad_token_id = tokenizer.eos_token_id model.config.pad_token_id = model.config.eos_token_id # Define your color-coding labels; if prob > x, then label = y; Sorted in descending probability order! probs_to_label = [ (0.1, "p >= 10%"), (0.01, "p >= 1%"), (1e-20, "p < 1%"), ] label_to_color = { "p >= 10%": "green", "p >= 1%": "yellow", "p < 1%": "red" } def get_tokens_and_labels(prompt): """ Given the prompt (text), return a list of tuples (decoded_token, label) """ inputs = tokenizer([prompt], return_tensors="pt") outputs = model.generate( **inputs, max_new_tokens=50, return_dict_in_generate=True, output_scores=True, do_sample=True ) # Important, don't forget to set `normalize_logits=True` to obtain normalized probabilities (i.e. sum(p) = 1) transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, normalize_logits=True) transition_proba = np.exp(transition_scores) # We only have scores for the generated tokens, so pop out the prompt tokens input_length = 1 if model.config.is_encoder_decoder else inputs.input_ids.shape[1] generated_tokens = outputs.sequences[:, input_length:] # initialize the highlighted output with the prompt, which will have no color label highlighted_out = [(tokenizer.decode(token), None) for token in inputs.input_ids] # get the (decoded_token, label) pairs for the generated tokens for token, proba in zip(generated_tokens[0], transition_proba[0]): this_label = None assert 0. <= proba <= 1.0 for min_proba, label in probs_to_label: if proba >= min_proba: this_label = label break highlighted_out.append((tokenizer.decode(token), this_label)) return highlighted_out demo = gr.Blocks() with demo: gr.Markdown( """ # Foo Bar """ ) prompt = gr.Textbox(label="Prompt", lines=3, value="Today is") highlighted_text = gr.HighlightedText( label="Highlighted generation", combine_adjacent=True, show_legend=True, ).style(color_map=label_to_color), button = gr.Button(f"Generate with {MODEL_NAME}") button.click(get_tokens_and_labels, inputs=prompt, outputs=highlighted_text) if __name__ == "__main__": demo.launch()