File size: 9,128 Bytes
2a6a910
3b85b9a
2a6a910
3b85b9a
a38ae12
2a6a910
3b85b9a
2a6a910
 
826e03f
2a6a910
 
 
8781a5b
a38ae12
4d55ef2
8781a5b
2a6a910
8781a5b
2a6a910
1cfb0d6
8781a5b
1cfb0d6
8781a5b
40ccb44
1cfb0d6
a38ae12
 
40ccb44
 
2a6a910
 
826e03f
8781a5b
 
1cfb0d6
 
 
 
 
 
 
 
8781a5b
 
2a6a910
 
8781a5b
2a6a910
 
 
1cfb0d6
 
 
 
8781a5b
 
2a6a910
 
 
826e03f
8781a5b
 
1cfb0d6
 
 
 
 
 
 
 
 
8781a5b
1cfb0d6
8781a5b
 
 
 
 
 
1cfb0d6
 
 
 
8781a5b
 
2a6a910
 
 
 
8781a5b
 
3b85b9a
 
 
2a6a910
 
 
1cfb0d6
 
 
 
 
 
 
 
2a6a910
 
41ebd25
2a6a910
8781a5b
2a6a910
8781a5b
2a6a910
 
1cfb0d6
 
 
 
 
 
 
 
 
 
3b85b9a
 
2a6a910
 
3b85b9a
2a6a910
41ebd25
3b85b9a
 
 
2a6a910
d08d04d
8781a5b
2a6a910
 
d08d04d
8781a5b
2a6a910
 
d08d04d
8781a5b
2a6a910
 
d08d04d
8781a5b
2a6a910
 
d08d04d
8781a5b
2a6a910
3b85b9a
 
8781a5b
2a6a910
40ccb44
 
 
 
 
2a6a910
0ffc0ad
09f41cc
 
 
 
 
 
74834a7
1a4f7ce
 
09f41cc
 
 
 
8781a5b
1cfb0d6
8781a5b
 
09f41cc
2a6a910
 
c91582d
 
2a6a910
 
 
1cfb0d6
2a6a910
 
c91582d
 
2a6a910
 
 
1cfb0d6
2a6a910
1cfb0d6
 
c91582d
2a6a910
 
 
1cfb0d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a6a910
 
 
 
40ccb44
2a6a910
3b85b9a
2a6a910
 
 
 
1cfb0d6
2a6a910
 
1cfb0d6
 
 
 
 
2a6a910
 
8781a5b
2a6a910
3b85b9a
2a6a910
 
 
1cfb0d6
2a6a910
 
1cfb0d6
 
 
 
 
2a6a910
71a369c
2a6a910
 
 
 
 
 
 
 
 
 
09f41cc
 
 
 
 
2a6a910
 
 
 
 
8781a5b
09f41cc
 
 
 
 
2a6a910
 
8781a5b
2a6a910
 
 
 
 
 
 
 
8781a5b
2a6a910
 
8781a5b
2a6a910
 
 
0b05515
2a6a910
 
 
 
3b85b9a
 
40ccb44
 
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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
#!/usr/bin/env python

from __future__ import annotations

import os
import string

import gradio as gr
import PIL.Image
import spaces
import torch
from transformers import AutoProcessor, Blip2ForConditionalGeneration

DESCRIPTION = "# [BLIP-2](https://github.com/salesforce/LAVIS/tree/main/projects/blip2)"

if not torch.cuda.is_available():
    DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

MODEL_ID_OPT_2_7B = "Salesforce/blip2-opt-2.7b"
MODEL_ID_OPT_6_7B = "Salesforce/blip2-opt-6.7b"
MODEL_ID_FLAN_T5_XL = "Salesforce/blip2-flan-t5-xl"
MODEL_ID_FLAN_T5_XXL = "Salesforce/blip2-flan-t5-xxl"
MODEL_ID = os.getenv("MODEL_ID", MODEL_ID_FLAN_T5_XXL)
assert MODEL_ID in [MODEL_ID_OPT_2_7B, MODEL_ID_OPT_6_7B, MODEL_ID_FLAN_T5_XL, MODEL_ID_FLAN_T5_XXL]

if torch.cuda.is_available():
    processor = AutoProcessor.from_pretrained(MODEL_ID)
    model = Blip2ForConditionalGeneration.from_pretrained(MODEL_ID, device_map="auto", load_in_8bit=True)


@spaces.GPU
def generate_caption(
    image: PIL.Image.Image,
    decoding_method: str = "Nucleus sampling",
    temperature: float = 1.0,
    length_penalty: float = 1.0,
    repetition_penalty: float = 1.5,
    max_length: int = 50,
    min_length: int = 1,
    num_beams: int = 5,
    top_p: float = 0.9,
) -> str:
    inputs = processor(images=image, return_tensors="pt").to(device, torch.float16)
    generated_ids = model.generate(
        pixel_values=inputs.pixel_values,
        do_sample=decoding_method == "Nucleus sampling",
        temperature=temperature,
        length_penalty=length_penalty,
        repetition_penalty=repetition_penalty,
        max_length=max_length,
        min_length=min_length,
        num_beams=num_beams,
        top_p=top_p,
    )
    result = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
    return result


@spaces.GPU
def answer_question(
    image: PIL.Image.Image,
    prompt: str,
    decoding_method: str = "Nucleus sampling",
    temperature: float = 1.0,
    length_penalty: float = 1.0,
    repetition_penalty: float = 1.5,
    max_length: int = 50,
    min_length: int = 1,
    num_beams: int = 5,
    top_p: float = 0.9,
) -> str:
    inputs = processor(images=image, text=prompt, return_tensors="pt").to(device, torch.float16)
    generated_ids = model.generate(
        **inputs,
        do_sample=decoding_method == "Nucleus sampling",
        temperature=temperature,
        length_penalty=length_penalty,
        repetition_penalty=repetition_penalty,
        max_length=max_length,
        min_length=min_length,
        num_beams=num_beams,
        top_p=top_p,
    )
    result = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
    return result


def postprocess_output(output: str) -> str:
    if output and output[-1] not in string.punctuation:
        output += "."
    return output


def chat(
    image: PIL.Image.Image,
    text: str,
    decoding_method: str = "Nucleus sampling",
    temperature: float = 1.0,
    length_penalty: float = 1.0,
    repetition_penalty: float = 1.5,
    max_length: int = 50,
    min_length: int = 1,
    num_beams: int = 5,
    top_p: float = 0.9,
    history_orig: list[str] = [],
    history_qa: list[str] = [],
) -> tuple[list[tuple[str, str]], list[str], list[str]]:
    history_orig.append(text)
    text_qa = f"Question: {text} Answer:"
    history_qa.append(text_qa)
    prompt = " ".join(history_qa)

    output = answer_question(
        image=image,
        prompt=prompt,
        decoding_method=decoding_method,
        temperature=temperature,
        length_penalty=length_penalty,
        repetition_penalty=repetition_penalty,
        max_length=max_length,
        min_length=min_length,
        num_beams=num_beams,
        top_p=top_p,
    )
    output = postprocess_output(output)
    history_orig.append(output)
    history_qa.append(output)

    chat_val = list(zip(history_orig[0::2], history_orig[1::2]))
    return chat_val, history_orig, history_qa


examples = [
    [
        "images/house.png",
        "How could someone get out of the house?",
    ],
    [
        "images/flower.jpg",
        "What is this flower and where is it's origin?",
    ],
    [
        "images/pizza.jpg",
        "What are steps to cook it?",
    ],
    [
        "images/sunset.jpg",
        "Here is a romantic message going along the photo:",
    ],
    [
        "images/forbidden_city.webp",
        "In what dynasties was this place built?",
    ],
]

with gr.Blocks(css="style.css") as demo:
    gr.Markdown(DESCRIPTION)
    gr.DuplicateButton(
        value="Duplicate Space for private use",
        elem_id="duplicate-button",
        visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
    )

    with gr.Group():
        image = gr.Image(type="pil")
        with gr.Tabs():
            with gr.Tab(label="Image Captioning"):
                caption_button = gr.Button("Caption it!")
                caption_output = gr.Textbox(label="Caption Output", show_label=False, container=False)
            with gr.Tab(label="Visual Question Answering"):
                chatbot = gr.Chatbot(label="VQA Chat", show_label=False)
                history_orig = gr.State(value=[])
                history_qa = gr.State(value=[])
                vqa_input = gr.Text(label="Chat Input", show_label=False, max_lines=1, container=False)
                with gr.Row():
                    clear_chat_button = gr.Button("Clear")
                    chat_button = gr.Button("Submit", variant="primary")
    with gr.Accordion(label="Advanced settings", open=False):
        text_decoding_method = gr.Radio(
            label="Text Decoding Method",
            choices=["Beam search", "Nucleus sampling"],
            value="Nucleus sampling",
        )
        temperature = gr.Slider(
            label="Temperature",
            info="Used with nucleus sampling.",
            minimum=0.5,
            maximum=1.0,
            step=0.1,
            value=1.0,
        )
        length_penalty = gr.Slider(
            label="Length Penalty",
            info="Set to larger for longer sequence, used with beam search.",
            minimum=-1.0,
            maximum=2.0,
            step=0.2,
            value=1.0,
        )
        repetition_penalty = gr.Slider(
            label="Repetition Penalty",
            info="Larger value prevents repetition.",
            minimum=1.0,
            maximum=5.0,
            step=0.5,
            value=1.5,
        )
        max_length = gr.Slider(
            label="Max Length",
            minimum=1,
            maximum=512,
            step=1,
            value=50,
        )
        min_length = gr.Slider(
            label="Minimum Length",
            minimum=1,
            maximum=100,
            step=1,
            value=1,
        )
        num_beams = gr.Slider(
            label="Number of Beams",
            minimum=1,
            maximum=10,
            step=1,
            value=5,
        )
        top_p = gr.Slider(
            label="Top P",
            info="Used with nucleus sampling.",
            minimum=0.5,
            maximum=1.0,
            step=0.1,
            value=0.9,
        )

    gr.Examples(
        examples=examples,
        inputs=[image, vqa_input],
    )

    caption_button.click(
        fn=generate_caption,
        inputs=[
            image,
            text_decoding_method,
            temperature,
            length_penalty,
            repetition_penalty,
            max_length,
            min_length,
            num_beams,
            top_p,
        ],
        outputs=caption_output,
        api_name="caption",
    )

    chat_inputs = [
        image,
        vqa_input,
        text_decoding_method,
        temperature,
        length_penalty,
        repetition_penalty,
        max_length,
        min_length,
        num_beams,
        top_p,
        history_orig,
        history_qa,
    ]
    chat_outputs = [
        chatbot,
        history_orig,
        history_qa,
    ]
    vqa_input.submit(
        fn=chat,
        inputs=chat_inputs,
        outputs=chat_outputs,
    ).success(
        fn=lambda: "",
        outputs=vqa_input,
        queue=False,
        api_name=False,
    )
    chat_button.click(
        fn=chat,
        inputs=chat_inputs,
        outputs=chat_outputs,
        api_name="chat",
    ).success(
        fn=lambda: "",
        outputs=vqa_input,
        queue=False,
        api_name=False,
    )
    clear_chat_button.click(
        fn=lambda: ("", [], [], []),
        inputs=None,
        outputs=[
            vqa_input,
            chatbot,
            history_orig,
            history_qa,
        ],
        queue=False,
        api_name="clear",
    )
    image.change(
        fn=lambda: ("", [], [], []),
        inputs=None,
        outputs=[
            caption_output,
            chatbot,
            history_orig,
            history_qa,
        ],
        queue=False,
    )

if __name__ == "__main__":
    demo.queue(max_size=10).launch()