File size: 8,128 Bytes
598d165
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
import torch
from transformers import TextStreamer

import os
import sys
sys.path.insert(0, os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "Evaluation"))
from llava.constants import IMAGE_TOKEN_INDEX
from llava.conversation import conv_templates, SeparatorStyle
from llava.mm_utils import get_model_name_from_path, KeywordsStoppingCriteria, tokenizer_image_token
from llava.model.builder import load_pretrained_model
from llava.utils import disable_torch_init
import shutil

#   <a href="https://github.com/SNUMPR/vlm-rlaif.git" style="margin-right: 20px; text-decoration: none; display: flex; align-items: center;">
#     <img src="https://z1.ax1x.com/2023/11/07/pil4sqH.png" alt="VLM-RLAIF" style="max-width: 120px; height: auto;">
#   </a>

cur_dir = os.path.dirname(os.path.abspath(__file__))
title_markdown = ("""
<div style="display: flex; justify-content: center; align-items: center; text-align: center;">
  <img src="/dataset/dcahn/yura/vlm-rlaif/asset/Model.png" alt="VLM-RLAIF" style="max-width: 120px; height: auto;">
  <img src="file:/dataset/dcahn/yura/vlm-rlaif/asset/Model.png" alt="VLM-RLAIF" style="max-width: 120px; height: auto;">
  <div>
    <h1 >VLM-RLAIF: Tuning Large Multimodal Models for Videos using Reinforcement Learning from AI Feedback (ACL 2024 Oral) </h1>
    <h5 style="margin: 0;">If you like our project, please give us a star โœจ on Github for the latest update.</h5>
  </div>
</div>


<div align="center">
    <div style="display:flex; gap: 0.25rem;" align="center">
        <a href='https://github.com/SNUMPR/vlm-rlaif'><img src='https://img.shields.io/badge/Github-Code-blue'></a>
        <a href="https://arxiv.org/abs/2402.03746"><img src="https://img.shields.io/badge/Paper-arxiv-green"></a>
    </div>
</div>
""")
        # <a href='https://github.com/PKU-YuanGroup/Video-LLaVA/stargazers'><img src='https://img.shields.io/github/stars/PKU-YuanGroup/Video-LLaVA.svg?style=social'></a> # arXiv ๋ฒ„ํŠผ ์˜†์— ์ถ”๊ฐ€?

block_css = """
#buttons button {
    min-width: min(120px,100%);
}
"""

tos_markdown = ("""
### Terms of use
By using this service, users are required to agree to the following terms:
The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. The service may collect user dialogue data for future research.
Please click the "Flag" button if you get any inappropriate answer! We will collect those to keep improving our moderator.
For an optimal experience, please use desktop computers for this demo, as mobile devices may compromise its quality.
""")

learn_more_markdown = ("""
### License
The service is a research preview intended for non-commercial use only, subject to the model [License](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) of LLaMA, [Terms of Use](https://openai.com/policies/terms-of-use) of the data generated by OpenAI, and [Privacy Practices](https://chrome.google.com/webstore/detail/sharegpt-share-your-chatg/daiacboceoaocpibfodeljbdfacokfjb) of ShareGPT. Please contact us if you find any potential violation.
""")


class Chat:
    def __init__(self, model_path, conv_mode, model_base=None, load_8bit=False, load_4bit=False, device='cuda', cache_dir=None):
        # model_base = '/dataset/yura/vlm-rlaif/pretrained/final_models/Video_LLaVA_SFT'
        # model_base='/dataset/yura/vlm-rlaif/pretrained/llava-v1.5-7b-lora_w_lora_16_sftv2_short1632_and_then_long_rank32_alpha32_lr1e4_allmodels/SFT_merged'
        # model_path = '/dataset/yura/vlm-rlaif/pretrained/LLaVA_Video-RL-Fact-RLHF-7b_SFTv2_RM_13b_v1_40k-v1.5-336-lora-padding/checkpoint-180/adapter_model/lora_policy'

        disable_torch_init()
        model_name = get_model_name_from_path(model_path)
        # self.tokenizer, self.model, image_processor, context_len = load_pretrained_model(model_path, model_base, model_name,
        #                                                                            load_8bit, load_4bit,
        #                                                                            device=device, cache_dir=cache_dir)
        is_rlhf_checkpoint = 'rlhf' in model_path.lower()
        print("MODEL_PATH", model_path)
        print("RLHF Checkpoint: ", is_rlhf_checkpoint)
        if not model_base or model_base == "none": model_base = None
        if is_rlhf_checkpoint:
            model_name = model_path            
            print("Config?", os.path.exists(os.path.join(model_path, "config.json")))
            if not os.path.exists(os.path.join(model_path, "config.json")):
                print("Copying")
                shutil.copy(os.path.join(model_base, "config.json"), os.path.join(model_path, "config.json")) # Copy SFT model's config -> to RLHF folder
                print("Listed", os.listdir(model_path))
                print("Copying done")
            # return(model_name)
        # return
        # self.tokenizer, self.model, image_processor, context_len = load_pretrained_model(model_path, model_base, model_name, load_8bit, load_4bit, device=device)
        self.tokenizer, self.model, image_processor, context_len = load_pretrained_model(model_path, model_base, model_name, False, False, device=device)

        

        self.image_processor = image_processor
        # self.image_processor = processor['image']
        # self.video_processor = processor['video']
        self.conv_mode = conv_mode
        self.conv = conv_templates[conv_mode].copy()
        self.device = self.model.device
        print(self.model)

    def get_prompt(self, qs, state):
        state.append_message(state.roles[0], qs)
        state.append_message(state.roles[1], None)
        return state
    
    def _get_latest_prompt(self, state):
        new_state = state.copy()
        new_state.messages = state.messages[-2:]
        return new_state

    @torch.inference_mode()
    # def generate(self, images_tensor: list, prompt: str, first_run: bool, state):
    def generate(self, images_tensor: torch.Tensor, prompt: str, first_run: bool, state):
        tokenizer, model, image_processor = self.tokenizer, self.model, self.image_processor

        state = self.get_prompt(prompt, state)
        # prompt = state.get_prompt()
        latest_state = self._get_latest_prompt(state)
        prompt = latest_state.get_prompt()
    
        # print('\n\n\n')
        # print(prompt)

        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(self.device)

        temperature = 0.2

        max_new_tokens = 1024

        stop_str = self.conv.sep if self.conv.sep_style != SeparatorStyle.TWO else self.conv.sep2
        keywords = [stop_str]
        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
        streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
        print(prompt, input_ids.shape, images_tensor.shape)
        # print(images_tensor)
        with torch.inference_mode():
            output_ids = model.generate(
                input_ids,
                images=images_tensor,
                do_sample=True,
                temperature=temperature,
                max_new_tokens=max_new_tokens,
                streamer=streamer,
                use_cache=True,
                stopping_criteria=[stopping_criteria])

        input_token_len = input_ids.shape[1]
        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
        if n_diff_input_output > 0:
            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
        outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
        outputs = outputs.strip()
        outputs = outputs.replace("QA_GT_caption_based_noisy", "")
        if outputs.endswith(stop_str):
            outputs = outputs[:-len(stop_str)]
        outputs = outputs.strip()

        print('response', outputs)
        return outputs, state