File size: 7,925 Bytes
819c0e1
046b1e3
141a5cd
 
 
 
89c0f3c
141a5cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66cdd61
141a5cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66cdd61
141a5cd
 
 
 
 
 
 
 
 
 
fe84005
205ce56
fe84005
036b110
 
fe84005
 
 
 
23771bd
141a5cd
 
 
89fdef5
141a5cd
 
 
 
 
 
 
 
 
 
 
 
 
 
89c0f3c
141a5cd
 
 
 
 
 
 
 
 
 
046b1e3
141a5cd
 
 
 
 
 
 
 
 
 
 
faf19c1
f239270
141a5cd
 
 
 
 
 
 
 
 
 
faf19c1
f239270
141a5cd
 
 
 
 
 
 
 
faf19c1
f239270
141a5cd
046b1e3
141a5cd
 
 
 
 
 
 
 
faf19c1
f239270
141a5cd
8e1a142
141a5cd
4d4e16c
 
 
 
 
faf19c1
f239270
141a5cd
046b1e3
ed48478
3b03cc7
ed48478
 
 
 
 
 
0d50329
 
 
 
 
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
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import gradio as gr

from vid2persona import init
from vid2persona.pipeline import vlm
from vid2persona.pipeline import llm

init.init_model("HuggingFaceH4/zephyr-7b-beta")
init.auth_gcp()
init.get_env_vars()
prompt_tpl_path = "vid2persona/prompts"

async def extract_traits(video_path):
    traits = await vlm.get_traits(
        init.gcp_project_id, 
        init.gcp_project_location, 
        video_path,
        prompt_tpl_path
    )
    if 'characters' in traits:
        traits = traits['characters'][0]

    return [
        traits, [], 
        gr.Textbox("", interactive=True),
        gr.Button(interactive=True),
        gr.Button(interactive=True),
        gr.Button(interactive=True)
    ]

async def conversation(
    message: str, messages: list, traits: dict,
    model_id: str, max_input_token_length: int, 
    max_new_tokens: int, temperature: float, 
    top_p: float, top_k: float, repetition_penalty: float, 
):
    messages = messages + [[message, ""]]
    yield [messages, message, gr.Button(interactive=False), gr.Button(interactive=False)]

    async for partial_response in llm.chat(
        message, messages, traits,
        prompt_tpl_path, model_id, 
        max_input_token_length, max_new_tokens,
        temperature, top_p, top_k, 
        repetition_penalty, hf_token=init.hf_access_token
    ):
        last_message = messages[-1]
        last_message[1] = last_message[1] + partial_response
        messages[-1] = last_message
        yield [messages, "", gr.Button(interactive=False), gr.Button(interactive=False)]

    yield [messages, "", gr.Button(interactive=True), gr.Button(interactive=True)]

async def regen_conversation(
    messages: list, traits: dict,
    model_id: str, max_input_token_length: int, 
    max_new_tokens: int, temperature: float, 
    top_p: float, top_k: float, repetition_penalty: float, 
):
    if len(messages) > 0:
        message = messages[-1][0]
        messages = messages[:-1]
        messages = messages + [[message, ""]]
        yield [messages, "", gr.Button(interactive=False), gr.Button(interactive=False)]

        async for partial_response in llm.chat(
            message, messages, traits,
            prompt_tpl_path, model_id, 
            max_input_token_length, max_new_tokens,
            temperature, top_p, top_k, 
            repetition_penalty, hf_token=init.hf_access_token
        ):
            last_message = messages[-1]
            last_message[1] = last_message[1] + partial_response
            messages[-1] = last_message
            yield [messages, "", gr.Button(interactive=False), gr.Button(interactive=False)]

        yield [messages, "", gr.Button(interactive=True), gr.Button(interactive=True)]

with gr.Blocks(css="styles.css", theme=gr.themes.Soft()) as demo:
    gr.Markdown("Vid2Persona", elem_classes=["md-center", "h1-font"])
    
    gr.Markdown("This project breathes life into video characters by using AI to describe their personality and then chat with you as them. "
                "[Gemini 1.0 Pro Vision model on Vertex AI](https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/overview) is used "
                "to grasp traits of video characters, then [HuggingFaceH4/zephyr-7b-beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta) model "
                "is used to make conversation with them.",)

    gr.Markdown("This space is modified to be working on Hugging Face [ZeroGPU](https://huggingface.co/zero-gpu-explorers). If you wish to run "
                "the same application on your own machine, please check out the [project repository](https://github.com/deep-diver/Vid2Persona). "
                "You can interact with other LLMs to make conversation besides HuggingFaceH4/zephyr-7b-beta by running them locally, or by "
                "connecting them through remotely hosted within Text Generation Inference framework as [Hugging Face PRO](https://huggingface.co/blog/inference-pro) user.")

    with gr.Column(elem_classes=["group"]):
        with gr.Row():
            video = gr.Video(label="upload short video clip", max_length=180)
            traits = gr.Json(label="extracted traits")
        
        with gr.Row():
            trait_gen = gr.Button("generate  traits")

    with gr.Column(elem_classes=["group"]):
        chatbot = gr.Chatbot([], label="chatbot", elem_id="chatbot", elem_classes=["chatbot-no-label"])
        with gr.Row():
            clear = gr.Button("clear conversation", interactive=False)
            regen = gr.Button("regenerate the last", interactive=False)
            stop = gr.Button("stop", interactive=False) 
        user_input = gr.Textbox(placeholder="ask anything", interactive=False, elem_classes=["textbox-no-label", "textbox-no-top-bottom-borders"])

        with gr.Accordion("parameters' control pane", open=False):
            model_id = gr.Dropdown(choices=init.ALLOWED_LLM_FOR_HF_PRO_ACCOUNTS, value="HuggingFaceH4/zephyr-7b-beta", label="Model ID", visible=False)

            with gr.Row():
                max_input_token_length = gr.Slider(minimum=1024, maximum=4096, value=4096, label="max-input-tokens")
                max_new_tokens = gr.Slider(minimum=128, maximum=2048, value=256, label="max-new-tokens")

            with gr.Row():
                temperature = gr.Slider(minimum=0, maximum=2, step=0.1, value=0.6, label="temperature")
                top_p = gr.Slider(minimum=0, maximum=2, step=0.1, value=0.9, label="top-p")
                top_k = gr.Slider(minimum=0, maximum=2, step=0.1, value=50, label="top-k")
                repetition_penalty = gr.Slider(minimum=0, maximum=2, step=0.1, value=1.2, label="repetition-penalty")
    
    with gr.Row():
        gr.Markdown(
            "[![GitHub Repo](https://img.shields.io/badge/GitHub%20Repo-gray?style=for-the-badge&logo=github&link=https://github.com/deep-diver/Vid2Persona)](https://github.com/deep-diver/Vid2Persona) "
            "[![Chansung](https://img.shields.io/badge/Chansung-blue?style=for-the-badge&logo=twitter&link=https://twitter.com/algo_diver)](https://twitter.com/algo_diver) "
            "[![Sayak](https://img.shields.io/badge/Sayak-blue?style=for-the-badge&logo=twitter&link=https://twitter.com/RisingSayak)](https://twitter.com/RisingSayak )",
            elem_id="bottom-md"
        )

    trait_gen.click(
        extract_traits,
        [video],
        [traits, chatbot, user_input, clear, regen, stop],
        concurrency_limit=5,
    )

    conv = user_input.submit(
        conversation,
        [
            user_input, chatbot, traits,
            model_id, max_input_token_length, 
            max_new_tokens, temperature, 
            top_p, top_k, repetition_penalty,
        ],
        [chatbot, user_input, clear, regen],
        concurrency_limit=5,
    )

    clear.click(
        lambda: [
            gr.Chatbot([]),
            gr.Button(interactive=False),
            gr.Button(interactive=False),
        ],
        None, [chatbot, clear, regen],
        concurrency_limit=5,
    )

    conv_regen = regen.click(
        regen_conversation,
        [
            chatbot, traits,
            model_id, max_input_token_length, 
            max_new_tokens, temperature, 
            top_p, top_k, repetition_penalty, 
        ],
        [chatbot, user_input, clear, regen],
        concurrency_limit=5,
    )

    stop.click(
        lambda: [
            gr.Button(interactive=True),
            gr.Button(interactive=True),
            gr.Button(interactive=True),
        ], None, [clear, regen, stop], 
        cancels=[conv, conv_regen],
        concurrency_limit=5,
    )

    gr.Examples(
        [["assets/sample1.mp4"], ["assets/sample2.mp4"], ["assets/sample3.mp4"], ["assets/sample4.mp4"]],
        video,
        [traits, chatbot, user_input, clear, regen, stop],
        extract_traits,
        cache_examples=True
    )

demo.queue(
    max_size=256
).launch(
    debug=True
)