Spaces:
Build error
Build error
Download server.py from daohieu17ctt/WaterFALL: direct link, hf CLI and curl.
- Browser
- Download file 2.86 kB
-
https://huggingface.co/spaces/daohieu17ctt/WaterFALL/resolve/main/server.py
- Command line
-
hf download hf://spaces/daohieu17ctt/WaterFALL/server.py
-
curl -L -o server.py https://huggingface.co/spaces/daohieu17ctt/WaterFALL/resolve/main/server.py
2.86 kB
| from fastapi import FastAPI | |
| from pydantic import BaseModel | |
| import torch | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| import numpy as np | |
| from Watermark.WatermarkingFnFourier import WatermarkingFnFourier | |
| from Watermark.WatermarkingFnSquare import WatermarkingFnSquare | |
| from Watermark.WatermarkerBase import Watermarker | |
| from sentence_transformers import SentenceTransformer | |
| app = FastAPI() | |
| model_name_or_path = "meta-llama/Llama-2-13b-chat-hf" | |
| device = "cuda:0" | |
| k_p=1 | |
| watermarkingFnClass = WatermarkingFnFourier | |
| # Initialize tokenizer and model | |
| tokenizer = AutoTokenizer.from_pretrained(model_name_or_path) | |
| model = AutoModelForCausalLM.from_pretrained( | |
| model_name_or_path, | |
| torch_dtype=torch.bfloat16, | |
| load_in_4bit=True | |
| ).to(device) | |
| sts_model = SentenceTransformer('sentence-transformers/all-mpnet-base-v2', device=device) | |
| prompt = ( | |
| "Paraphrase the user provided text while preserving semantic similarity. " | |
| "Do not include any other sentences in the response, such as explanations of the paraphrasing. " | |
| "Do not summarize." | |
| ) | |
| pre_paraphrased = "Here is a paraphrased version of the text while preserving the semantic similarity:\n\n" | |
| # Define the request format | |
| class WatermarkRequest(BaseModel): | |
| text: str | |
| key: int | |
| strength: int | |
| # Paraphrasing endpoint | |
| def watermark(request: WatermarkRequest): | |
| T_o = request.text | |
| id = request.key | |
| kappa = request.strength | |
| try: | |
| id = int(id) | |
| kappa = int(kappa) | |
| except ValueError: | |
| return {'text': "Please enter valid numbers."} | |
| # init watermarker | |
| watermarker = Watermarker(model, tokenizer, id, kappa, k_p, watermarkingFnClass=watermarkingFnClass) | |
| # watermark | |
| paraphrasing_prompt = tokenizer.apply_chat_template( | |
| [ | |
| {"role":"system", "content":prompt}, | |
| {"role":"user", "content":T_o}, | |
| ], tokenize=False, add_generation_prompt = True) + f"{pre_paraphrased}\n\n" | |
| watermarked = watermarker.generate( | |
| paraphrasing_prompt, | |
| return_scores=True, | |
| max_new_tokens=int(len(paraphrasing_prompt) * 1.5), | |
| do_sample=False, temperature=None, top_p=None, | |
| num_beams=8, num_beam_groups=4, num_return_sequences=8, diversity_penalty = 0.5 | |
| ) | |
| sts_scores = sts_model.encode([T_o, *watermarked["text"]], convert_to_tensor=True) | |
| cos_sim = torch.nn.functional.cosine_similarity(sts_scores[0], sts_scores[1:], dim=1).cpu() | |
| selection_score = cos_sim + watermarked["q_score"] | |
| selection = torch.argmax(selection_score) | |
| watermarked = watermarked["text"][selection] | |
| res_o = watermarker.verify(T_o)[0] | |
| q_o = res_o[k_p-1] | |
| # Verify on watermarked text | |
| res_w = watermarker.verify(watermarked)[0] | |
| q_w = res_w[k_p-1] | |
| return {'text': watermarked, 'q_o': q_o, 'q_w': q_w} | |