app.py

import os
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline

tokenizer = AutoTokenizer.from_pretrained("milyiyo/paraphraser-german-mt5-small")
model = AutoModelForSeq2SeqLM.from_pretrained("milyiyo/paraphraser-german-mt5-small")


def generate_v1(inputs, count):
    """Generate text using a Beam Search strategy with repetition penalty."""
    model_outputs = model.generate(inputs["input_ids"], 
                                           early_stopping=True,
                                           length_penalty=1.0, 
                                           max_length=1024, 
                                           no_repeat_ngram_size=2, 
                                           num_beams=10, 
                                           repetition_penalty=3.5,
                                           num_return_sequences=count
                                           )
    sentences = []
    for output in model_outputs:
        sentences.append(tokenizer.decode(output, skip_special_tokens=True))
    return sentences


def generate_v2(inputs, count):
    """Generate text using a Beam Search strategy."""
    model_outputs = model.generate(inputs["input_ids"],
                                           early_stopping=True,
                                           length_penalty=2.0,
                                           max_length=1024,
                                           no_repeat_ngram_size=2,
                                           num_beams=5,
                                           temperature=1.5,
                                           num_return_sequences=count
                                          )
    sentences = []
    for output in model_outputs:
        sentences.append(tokenizer.decode(output, skip_special_tokens=True))
    return sentences


def generate_v3(inputs, count):
    """Generate text using a Diverse Beam Search strategy."""
    model_outputs = model.generate(inputs["input_ids"],
                                num_beams=5,
                                max_length=1024,
                                temperature=1.5,
                                num_beam_groups=5,
                                diversity_penalty=2.0,
                                no_repeat_ngram_size=2,
                                early_stopping=True,
                                length_penalty=2.0,
                                num_return_sequences=count)
    sentences = []
    for output in model_outputs:
        sentences.append(tokenizer.decode(output, skip_special_tokens=True))
    return sentences


def generate_v4(encoding, count):
    """Generate text using a Diverse Beam Search strategy."""
    print(encoding)
    input_ids, attention_masks = encoding["input_ids"], encoding["attention_mask"]
    print(input_ids)
    print(attention_masks)
    outputs = model.generate(input_ids=input_ids, 
                             attention_mask=attention_masks,
                             max_length=512,
                             do_sample=True,
                             top_k=120,
                             top_p=0.95,
                             early_stopping=True,
                             num_return_sequences=count)
    res = []
    for output in outputs:
        line = tokenizer.decode(output, skip_special_tokens=True, clean_up_tokenization_spaces=True)
        res.append(line)
    return res


def paraphrase(sentence: str, count: str):
    p_count = int(count)
    if p_count <= 0 or len(sentence.strip()) == 0:
        return {'result': []}
    sentence_input = sentence
    text = f"paraphrase: {sentence_input} </s>"
    # encoding = tokenizer.encode_plus(text, padding=True, return_tensors="pt")
    encoding = tokenizer(text, return_tensors="pt")
    # input_ids, attention_masks = encoding["input_ids"], encoding["attention_mask"]
    
    # outputs = model.generate(
    #     input_ids=input_ids, attention_mask=attention_masks,
    #     max_length=512,  # 256,
    #     do_sample=True,
    #     top_k=120,
    #     top_p=0.95,
    #     early_stopping=True,
    #     num_return_sequences=p_count
    # )
    # res = []
    # for output in outputs:
    #     line = tokenizer.decode(
    #         output, skip_special_tokens=True, clean_up_tokenization_spaces=True)
    #     res.append(line)

    # print(res)

    # 
    input_ids, attention_masks = encoding["input_ids"], encoding["attention_mask"]
    outputs = model.generate(input_ids=input_ids, 
                             attention_mask=attention_masks,
                             max_length=512,
                             do_sample=True,
                             top_k=120,
                             top_p=0.95,
                             early_stopping=True,
                             num_return_sequences=p_count)
    result_v4 = []
    for output in outputs:
        line = tokenizer.decode(output, skip_special_tokens=True, clean_up_tokenization_spaces=True)
        result_v4.append(line)
    # 
    
    return {
        'result': {
            # 'generate_v1':generate_v1(encoding, count),
            # 'generate_v2':generate_v2(encoding, count),
            # 'generate_v3':generate_v3(encoding, count),
            'generate_v4':result_v4
        }
    }
        
def paraphrase_dummy(sentence: str, count: str):
    return {'result': []}


iface = gr.Interface(fn=paraphrase,
                     inputs=[
                         gr.inputs.Textbox(lines=2, placeholder=None, label='Sentence'),
                         gr.inputs.Number(default=3, label='Paraphrases count'),
                     ],
                     outputs=[gr.outputs.JSON(label=None)])
iface.launch()