app.py

import gradio as gr
import numpy as np
import torch
from transformers import pipeline
from transformers.pipelines import PIPELINE_REGISTRY, FillMaskPipeline
from transformers import AutoModelForMaskedLM

ex_str1 = "A crustless sandwich made from two slices of baked bread. The sandwich includes first and second matching " \
          "crustless bread pieces. The bread pieces have the same general outer shape defined by an outer periphery " \
          "with central portions surrounded by an outer peripheral area, the bread pieces being at least partially " \
          "crimped together at the outer peripheral area."

ex_str2 = "The present disclosure provides a DNA-targeting RNA that comprises a targeting sequence and, together with" \
          " a modifying polypeptide, provides for site-specific modification of a target DNA and/or a polypeptide" \
          " associated with the target DNA. "

ex_str3 = "The graphite plane is composed of a two-dimensional hexagonal lattice of carbon atoms and the plate has a " \
          "length and a width parallel to the graphite plane and a thickness orthogonal to the graphite plane with at " \
          "least one of the length, width, and thickness values being 100 nanometers or smaller. "

examples = [[ex_str1, 1.2, 1],
           [ex_str2, 1.5, 10],
           [ex_str3, 1.4, 5]]


def add_mask(text, size=1):
    split_text = text.split()

    # If the user supplies a mask, don't add more
    if '[MASK]' in split_text:
        return text
    idx = np.random.randint(len(split_text), size=size)
    masked_strings = []
    for i in idx:
        masked_strings.append(split_text[i])
        split_text[i] = '[MASK]'
    masked_output = ' '.join(split_text)
    return masked_output, masked_strings


class TempScalePipe(FillMaskPipeline):
    def _sanitize_parameters(self, top_k=None, targets=None, temp=None):
        postprocess_params = {}

        if targets is not None:
            target_ids = self.get_target_ids(targets, top_k)
            postprocess_params["target_ids"] = target_ids

        if top_k is not None:
            postprocess_params["top_k"] = top_k

        if temp is not None:
            postprocess_params["temp"] = temp
        return {}, {}, postprocess_params


    def __call__(self, inputs, *args, **kwargs):
        """
        Fill the masked token in the text(s) given as inputs.

        Args:
            args (`str` or `List[str]`):
                One or several texts (or one list of prompts) with masked tokens.
            targets (`str` or `List[str]`, *optional*):
                When passed, the model will limit the scores to the passed targets instead of looking up in the whole
                vocab. If the provided targets are not in the model vocab, they will be tokenized and the first
                resulting token will be used (with a warning, and that might be slower).
            top_k (`int`, *optional*):
                When passed, overrides the number of predictions to return.

        Return:
            A list or a list of list of `dict`: Each result comes as list of dictionaries with the following keys:

            - **sequence** (`str`) -- The corresponding input with the mask token prediction.
            - **score** (`float`) -- The corresponding probability.
            - **token** (`int`) -- The predicted token id (to replace the masked one).
            - **token** (`str`) -- The predicted token (to replace the masked one).
        """
        outputs = super().__call__(inputs, **kwargs)
        if isinstance(inputs, list) and len(inputs) == 1:
            return outputs[0]
        return outputs

    def postprocess(self, model_outputs, top_k=10, target_ids=None, temp=1):
        # Cap top_k if there are targets
        if target_ids is not None and target_ids.shape[0] < top_k:
            top_k = target_ids.shape[0]
        input_ids = model_outputs["input_ids"][0]
        outputs = model_outputs["logits"]

        masked_index = torch.nonzero(input_ids == self.tokenizer.mask_token_id, as_tuple=False).squeeze(-1)
        # Fill mask pipeline supports only one ${mask_token} per sample

        logits = outputs[0, masked_index, :] / temp
        probs = logits.softmax(dim=-1)
        sampling = False
        if sampling:
            predictions = torch.multinomial(probs, num_samples=3)
            values = probs[0, predictions]
        if target_ids is not None:
            probs = probs[..., target_ids]
        if not sampling:
            values, predictions = probs.topk(top_k)

        result = []
        single_mask = values.shape[0] == 1
        for i, (_values, _predictions) in enumerate(zip(values.tolist(), predictions.tolist())):
            row = []
            for v, p in zip(_values, _predictions):
                # Copy is important since we're going to modify this array in place
                tokens = input_ids.numpy().copy()
                if target_ids is not None:
                    p = target_ids[p].tolist()

                tokens[masked_index[i]] = p
                # Filter padding out:
                tokens = tokens[np.where(tokens != self.tokenizer.pad_token_id)]
                # Originally we skip special tokens to give readable output.
                # For multi masks though, the other [MASK] would be removed otherwise
                # making the output look odd, so we add them back
                sequence = self.tokenizer.decode(tokens, skip_special_tokens=single_mask)
                proposition = {"score": v, "token": p, "token_str": self.tokenizer.decode([p]), "sequence": sequence}
                row.append(proposition)
            result.append(row)
        if single_mask:
            return result[0]
        return result


PIPELINE_REGISTRY.register_pipeline(
    "temp-scale",
    pipeline_class=TempScalePipe,
    pt_model=AutoModelForMaskedLM,
)
scrambler = pipeline("temp-scale", model="anferico/bert-for-patents")


def sample_output(out, sampling):
    score_to_str = {out[k]: k for k in out.keys()}
    score_list = list(score_to_str.keys())
    if sampling == 'multi':
        idx = np.argmax(np.random.multinomial(1, score_list, 1))
    else:
        idx = np.random.randint(0, len(score_list))
    score = score_list[idx]
    return score_to_str[score]


def unmask(text, temp, rounds):
    sampling = 'multi'
    for _ in range(rounds):
        tp = add_mask(text, size=1)
        masked_text, masked = tp[0], tp[1]
        split_text = masked_text.split()
        res = scrambler(masked_text, temp=temp, top_k=15)
        mask_pos = [i for i, t in enumerate(split_text) if 'MASK' in t][0]
        out = {item["token_str"]: item["score"] for item in res}
        new_token = sample_output(out, sampling)
        unsuccessful_iters = 0
        while new_token == masked[0]:
            if unsuccessful_iters > 5:
                break
            print('skipped', new_token)
            new_token = sample_output(out, sampling='uniform')
            unsuccessful_iters += 1
        if new_token == masked[0]:
            split_text[mask_pos] = new_token
        else:
            split_text[mask_pos] = '*' + new_token + '*'
        text = ' '.join(split_text)

    text = list(text)
    text[0] = text[0].upper()
    return ''.join(text)


textbox = gr.Textbox(label="Example prompts", lines=5)
output_textbox = gr.Textbox(placeholder="Output will appear here", lines=4)
temp_slider = gr.Slider(1.0, 3.0, value=1.0, label='Creativity')
edit_slider = gr.Slider(1, 20, step=5, value=1.0, label='Number of edits')

demo = gr.Interface(
    fn=unmask,
    inputs=[textbox, temp_slider, edit_slider],
    outputs=[output_textbox],
    examples=examples,
    allow_flagging='never'
)

demo.launch()