File size: 7,805 Bytes
61f3f56
 
 
 
 
 
 
 
 
 
d1b9dab
932c4bc
61f3f56
 
d1b9dab
61f3f56
 
 
 
 
 
 
932c4bc
 
 
61f3f56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6e0af0
61f3f56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch

from llava.constants import X_TOKEN_INDEX
from llava.conversation import conv_templates, SeparatorStyle
from llava.mm_utils import get_model_name_from_path, KeywordsStoppingCriteria, tokenizer_X_token
from llava.model.builder import load_pretrained_model
from llava.utils import disable_torch_init

title_markdown = ("""
<div style="display: flex; justify-content: center; align-items: center; text-align: center;">
  <a href="https://github.com/PKU-YuanGroup/Video-LLaVA" style="margin-right: 20px; text-decoration: none; display: flex; align-items: center;">
    <img src="https://z1.ax1x.com/2023/11/07/pil4sqH.png" alt="Video-LLaVA🚀" style="max-width: 120px; height: auto;">
  </a>
  <div>
    <h1 >Video-LLaVA: Video-LLaVA: Learning United Visual Representation by Alignment Before Projection</h1>
    <h5 style="margin: 0;">If you like our project, please give us a star ✨ 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/PKU-YuanGroup/Video-LLaVA'><img src='https://img.shields.io/badge/Github-Code-blue'></a>
        <a href="https://arxiv.org/pdf/2311.10122.pdf"><img src="https://img.shields.io/badge/Arxiv-2311.10122-red"></a>
        <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>
    </div>
</div>
""")

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'):
        disable_torch_init()
        model_name = get_model_name_from_path(model_path)
        self.tokenizer, self.model, processor, context_len = load_pretrained_model(model_path, model_base, model_name,
                                                                         load_8bit, load_4bit,
                                                                         device=device)
        self.image_processor = processor['image']
        self.video_processor = processor['video']
        self.conv_mode = conv_mode
        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

    @torch.inference_mode()
    def generate(self, images_tensor: list, 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()
        print('\n\n\n')
        print(prompt)

        if 'image' in images_tensor[1] and 'video' not in images_tensor[1]:
            input_ids = tokenizer_X_token(prompt, tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt').unsqueeze(0).to(self.device)
        elif 'image' not in images_tensor[1] and 'video' in images_tensor[1]:
            input_ids = tokenizer_X_token(prompt, tokenizer, X_TOKEN_INDEX['VIDEO'], return_tensors='pt').unsqueeze(0).to(self.device)
        elif 'image' in images_tensor[1] and 'video' in images_tensor[1]:
            # <video>\nxxxxxxx\n<image>
            '''
            tensor([[1, -200, 29871, 13, 3068, 366, 1074, 1716, 278, 1967, 322, 4863, 29973, 319, 1799, 9047, 13566, 29901]])
            tensor([[1, -201, 29871, 13]])
            '''
            print("split: ", prompt.split('\n<image>'))
            # print("\n", tokenizer_X_token('\n', tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt'))
            # print("?", tokenizer_X_token('?', tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt'))
            # print("image", tokenizer_X_token('image', tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt'))
            # print("image?", tokenizer_X_token('image?', tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt'))
            # print("USER: <image>\nWhat is unusual about this image?", tokenizer_X_token('USER: <image>\nWhat is unusual about this image?', tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt'))
            input_ids1 = tokenizer_X_token(prompt.split('\n<image>')[0], tokenizer, X_TOKEN_INDEX['VIDEO'], return_tensors='pt').unsqueeze(0).to(self.device)
            print('input_ids1', input_ids1)
            input_ids2 = tokenizer_X_token(prompt.split('\n<image>')[-1], tokenizer, X_TOKEN_INDEX['VIDEO'], return_tensors='pt').unsqueeze(0).to(self.device)
            print('input_ids2', input_ids2)
            input_ids3 = tokenizer_X_token('\n<image>', tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt').unsqueeze(0).to(self.device)
            print('input_ids3', input_ids3)

            input_ids = torch.cat([input_ids1, input_ids3[:, 1:], input_ids2[:, 1:]], dim=-1)
            print('input_ids', input_ids)
            print(*[tokenizer.decode(i) for i in input_ids2[0]])
        else:
            input_ids = tokenizer_X_token(prompt, tokenizer, X_TOKEN_INDEX['IMAGE'], return_tensors='pt').unsqueeze(0).to(self.device)
        
        temperature = 0.1
        
        max_new_tokens = 1024

        stop_str = conv_templates[self.conv_mode].copy().sep if conv_templates[self.conv_mode].copy().sep_style != SeparatorStyle.TWO else \
        conv_templates[self.conv_mode].copy().sep2
        keywords = [stop_str]
        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
        # streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
        # print(input_ids, images_tensor[0][0].shape)
        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()
        if outputs.endswith(stop_str):
            outputs = outputs[:-len(stop_str)]
        outputs = outputs.strip()

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