app.py

#!/usr/local/bin/python3
#-*- coding:utf-8 -*-

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import os

checkpoint = "gpt2-large"
# checkpoint = "/innev/open-ai/huggingface/models/gpt2-large"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# model = AutoModelForCausalLM.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, pad_token_id=tokenizer.eos_token_id)

# 简单生成
def sampleGen(text):
    # text = 'Who was Jim Henson ? Jim Henson was a'

    # 编码一段文本
    # 编码后为[8241, 373, 5395, 367, 19069, 5633, 5395, 367, 19069, 373, 257]
    indexed_tokens = tokenizer.encode(text)
    # 转换为pytorch tensor
    # tensor([[ 8241,   373,  5395,   367, 19069,  5633,  5395,   367, 19069,   373, 257]])
    # shape为 torch.Size([1, 11])
    tokens_tensor = torch.tensor([indexed_tokens])

    # 设置为evaluation模式，去取消激活dropout等模块。
    # 在huggingface/transformers框架中，默认就是eval模式
    model.eval()

    # 预测所有token
    with torch.no_grad():
        # 将输入tensor输入，就得到了模型的输出，非常简单
        # outputs是一个元组，所有huggingface/transformers模型的输出都是元组
        # 本初的元组有两个，第一个是预测得分（没经过softmax之前的，也叫作logits），
        # 第二个是past，里面的attention计算的key value值
        # 此时我们需要的是第一个值
        outputs = model(tokens_tensor)
        # predictions shape为 torch.Size([1, 11, 50257])，
        # 也就是11个词每个词的预测得分（没经过softmax之前的）
        # 也叫做logits
        predictions = outputs[0]

    # 我们需要预测下一个单词，所以是使用predictions第一个batch，最后一个词的logits去计算
    # predicted_index = 582，通过计算最大得分的索引得到的
    predicted_index = torch.argmax(predictions[0, -1, :]).item()
    # 反向解码为我们需要的文本
    predicted_text = tokenizer.decode(indexed_tokens + [predicted_index])
    # predicted_text = tokenizer.decode([predicted_index])
    # 解码后的文本：'Who was Jim Henson? Jim Henson was a man'
    # 成功预测出单词 'man'

    return predicted_text

# 关键词预测 生成文本
def loopGen(prompts):
    text = prompts
    total = 1
    while text[-1] != "." and total < 20:
        text = sampleGen(text)
        print("Index %s: %s" % (total, text))
        total = total + 1
    
    return text, total

# 贪心搜索 生成文本
def greedySearch(prompts):
    input_ids = tokenizer(prompts, return_tensors='pt').input_ids

    # generate the result with greedy search
    output = model.generate(input_ids, max_length=128)
    text = tokenizer.decode(output[0], skip_special_tokens=True)
    return text, 1

# 随机方法 生成文本
def randomSearch(prompts):
    input_ids = tokenizer(prompts, return_tensors='pt').input_ids

    # generate the result with random search
    torch.manual_seed(0.)
    output = model.generate(input_ids, do_sample=True, max_length=128, top_p=0.95, top_k=0)
    text = tokenizer.decode(output[0], skip_special_tokens=True)
    return text, 1

# 对比搜索 生成文本
def contrastiveSearch(prompts):
    input_ids = tokenizer(prompts, return_tensors='pt').input_ids

    # generate the result with contrastive search
    output = model.generate(input_ids, penalty_alpha=0.6, top_k=4, max_length=512)
    text = tokenizer.decode(output[0], skip_special_tokens=True)

    return text, 1

def predict(searchType, prompts='Who was Jim Henson ? Jim Henson was a'):
    if searchType == "贪心搜索":
        return greedySearch(prompts)
    elif searchType == "随机方法":
        return randomSearch(prompts)
    elif searchType == "对比搜索":
        return contrastiveSearch(prompts)
    else:
        return loopGen(prompts)


title = "GPT2 large"

searchMapping = ['关键词预测', '贪心搜索', '随机方法', '对比搜索']

description = """
本例为使用GPT2模型的简单推测语句DEMO，输入前面的句子，推测出后面的句子。

使用原始模型，未经过微调。只支持英文输入输出。
"""

examples = [
    [None, "DeepMind Company is", None],
    [None, "Who was Jim Henson ? Jim Henson was a", None],
    [None, "China is", None]
]

article = """
## 文章参考
- [在 Transformers 中使用对比搜索生成可媲美人类水平的文本 🤗](https://mp.weixin.qq.com/s/mydQLDlGUzFJuNBCIYc3CA)
"""

gr.Interface(
    fn=predict,
    inputs=[
        gr.Radio(label="搜索方法", choices=searchMapping, value="关键词预测"),
        gr.Text(label="输入前置语句"),
    ],
    outputs=[
        gr.Text(label="生成文本"),
        gr.Text(label="循环次数"),
    ],
    title=title,
    description=description,
    article=article,
    examples=examples,
).launch()