Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,272 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import gradio as gr
|
3 |
+
from transformers import AutoProcessor, Idefics2ForConditionalGeneration, TextIteratorStreamer
|
4 |
+
from threading import Thread
|
5 |
+
import re
|
6 |
+
import time
|
7 |
+
from PIL import Image
|
8 |
+
import torch
|
9 |
+
import spaces
|
10 |
+
|
11 |
+
PROCESSOR = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b")
|
12 |
+
|
13 |
+
model = Idefics2ForConditionalGeneration.from_pretrained(
|
14 |
+
"HuggingFaceM4/idefics2-8b",
|
15 |
+
torch_dtype=torch.bfloat16,
|
16 |
+
_attn_implementation="flash_attention_2",
|
17 |
+
trust_remote_code=True).to("cuda")
|
18 |
+
|
19 |
+
|
20 |
+
|
21 |
+
def turn_is_pure_media(turn):
|
22 |
+
return turn[1] is None
|
23 |
+
def format_user_prompt_with_im_history_and_system_conditioning(
|
24 |
+
user_prompt, chat_history
|
25 |
+
):
|
26 |
+
"""
|
27 |
+
Produces the resulting list that needs to go inside the processor.
|
28 |
+
It handles the potential image(s), the history and the system conditionning.
|
29 |
+
"""
|
30 |
+
resulting_messages = copy.deepcopy([])
|
31 |
+
resulting_images = []
|
32 |
+
|
33 |
+
# Format history
|
34 |
+
for turn in chat_history:
|
35 |
+
if not resulting_messages or (resulting_messages and resulting_messages[-1]["role"] != "user"):
|
36 |
+
resulting_messages.append(
|
37 |
+
{
|
38 |
+
"role": "user",
|
39 |
+
"content": [],
|
40 |
+
}
|
41 |
+
)
|
42 |
+
|
43 |
+
if turn_is_pure_media(turn):
|
44 |
+
media = turn[0][0]
|
45 |
+
resulting_messages[-1]["content"].append({"type": "image"})
|
46 |
+
resulting_images.append(Image.open(media))
|
47 |
+
else:
|
48 |
+
user_utterance, assistant_utterance = turn
|
49 |
+
resulting_messages[-1]["content"].append(
|
50 |
+
{"type": "text", "text": user_utterance.strip()}
|
51 |
+
)
|
52 |
+
resulting_messages.append(
|
53 |
+
{
|
54 |
+
"role": "assistant",
|
55 |
+
"content": [
|
56 |
+
{"type": "text", "text": user_utterance.strip()}
|
57 |
+
]
|
58 |
+
}
|
59 |
+
)
|
60 |
+
|
61 |
+
# Format current input
|
62 |
+
if not user_prompt["files"]:
|
63 |
+
resulting_messages.append(
|
64 |
+
{
|
65 |
+
"role": "user",
|
66 |
+
"content": [
|
67 |
+
{"type": "text", "text": user_prompt['text']}
|
68 |
+
],
|
69 |
+
}
|
70 |
+
)
|
71 |
+
else:
|
72 |
+
# Choosing to put the image first (i.e. before the text), but this is an arbiratrary choice.
|
73 |
+
resulting_messages.append(
|
74 |
+
{
|
75 |
+
"role": "user",
|
76 |
+
"content": [{"type": "image"}] * len(user_prompt['files']) + [
|
77 |
+
{"type": "text", "text": user_prompt['text']}
|
78 |
+
]
|
79 |
+
}
|
80 |
+
)
|
81 |
+
for im in user_prompt["files"]:
|
82 |
+
print(im)
|
83 |
+
if isinstance(im, str):
|
84 |
+
|
85 |
+
resulting_images.extend([Image.open(im)])
|
86 |
+
elif isinstance(im, dict):
|
87 |
+
resulting_images.extend([Image.open(im['path'])])
|
88 |
+
|
89 |
+
|
90 |
+
return resulting_messages, resulting_images
|
91 |
+
|
92 |
+
|
93 |
+
def extract_images_from_msg_list(msg_list):
|
94 |
+
all_images = []
|
95 |
+
for msg in msg_list:
|
96 |
+
for c_ in msg["content"]:
|
97 |
+
if isinstance(c_, Image.Image):
|
98 |
+
all_images.append(c_)
|
99 |
+
return all_images
|
100 |
+
|
101 |
+
@spaces.GPU(duration=180)
|
102 |
+
def model_inference(
|
103 |
+
user_prompt,
|
104 |
+
chat_history,
|
105 |
+
decoding_strategy,
|
106 |
+
temperature,
|
107 |
+
max_new_tokens,
|
108 |
+
repetition_penalty,
|
109 |
+
top_p,
|
110 |
+
):
|
111 |
+
if user_prompt["text"].strip() == "" and not user_prompt["files"]:
|
112 |
+
gr.Error("Please input a query and optionally image(s).")
|
113 |
+
|
114 |
+
if user_prompt["text"].strip() == "" and user_prompt["files"]:
|
115 |
+
gr.Error("Please input a text query along the image(s).")
|
116 |
+
|
117 |
+
|
118 |
+
streamer = TextIteratorStreamer(
|
119 |
+
PROCESSOR.tokenizer,
|
120 |
+
skip_prompt=True,
|
121 |
+
timeout=5.,
|
122 |
+
)
|
123 |
+
|
124 |
+
# Common parameters to all decoding strategies
|
125 |
+
# This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies
|
126 |
+
generation_args = {
|
127 |
+
"max_new_tokens": max_new_tokens,
|
128 |
+
"repetition_penalty": repetition_penalty,
|
129 |
+
"streamer": streamer,
|
130 |
+
}
|
131 |
+
|
132 |
+
assert decoding_strategy in [
|
133 |
+
"Greedy",
|
134 |
+
"Top P Sampling",
|
135 |
+
]
|
136 |
+
if decoding_strategy == "Greedy":
|
137 |
+
generation_args["do_sample"] = False
|
138 |
+
elif decoding_strategy == "Top P Sampling":
|
139 |
+
generation_args["temperature"] = temperature
|
140 |
+
generation_args["do_sample"] = True
|
141 |
+
generation_args["top_p"] = top_p
|
142 |
+
|
143 |
+
# Creating model inputs
|
144 |
+
resulting_text, resulting_images = format_user_prompt_with_im_history_and_system_conditioning(
|
145 |
+
user_prompt=user_prompt,
|
146 |
+
chat_history=chat_history,
|
147 |
+
)
|
148 |
+
prompt = PROCESSOR.apply_chat_template(resulting_text, add_generation_prompt=True)
|
149 |
+
inputs = PROCESSOR(text=prompt, images=resulting_images if resulting_images else None, return_tensors="pt")
|
150 |
+
inputs = {k: v.to("cuda") for k, v in inputs.items()}
|
151 |
+
generation_args.update(inputs)
|
152 |
+
|
153 |
+
|
154 |
+
thread = Thread(
|
155 |
+
target=model.generate,
|
156 |
+
kwargs=generation_args,
|
157 |
+
)
|
158 |
+
thread.start()
|
159 |
+
|
160 |
+
print("Start generating")
|
161 |
+
acc_text = ""
|
162 |
+
for text_token in streamer:
|
163 |
+
time.sleep(0.04)
|
164 |
+
acc_text += text_token
|
165 |
+
if acc_text.endswith("<end_of_utterance>"):
|
166 |
+
acc_text = acc_text[:-18]
|
167 |
+
yield acc_text
|
168 |
+
print("Success - generated the following text:", acc_text)
|
169 |
+
print("-----")
|
170 |
+
BOT_AVATAR = "IDEFICS_logo.png"
|
171 |
+
|
172 |
+
# Hyper-parameters for generation
|
173 |
+
max_new_tokens = gr.Slider(
|
174 |
+
minimum=8,
|
175 |
+
maximum=1024,
|
176 |
+
value=512,
|
177 |
+
step=1,
|
178 |
+
interactive=True,
|
179 |
+
label="Maximum number of new tokens to generate",
|
180 |
+
)
|
181 |
+
repetition_penalty = gr.Slider(
|
182 |
+
minimum=0.01,
|
183 |
+
maximum=5.0,
|
184 |
+
value=1.2,
|
185 |
+
step=0.01,
|
186 |
+
interactive=True,
|
187 |
+
label="Repetition penalty",
|
188 |
+
info="1.0 is equivalent to no penalty",
|
189 |
+
)
|
190 |
+
decoding_strategy = gr.Radio(
|
191 |
+
[
|
192 |
+
"Greedy",
|
193 |
+
"Top P Sampling",
|
194 |
+
],
|
195 |
+
value="Greedy",
|
196 |
+
label="Decoding strategy",
|
197 |
+
interactive=True,
|
198 |
+
info="Higher values is equivalent to sampling more low-probability tokens.",
|
199 |
+
)
|
200 |
+
temperature = gr.Slider(
|
201 |
+
minimum=0.0,
|
202 |
+
maximum=5.0,
|
203 |
+
value=0.4,
|
204 |
+
step=0.1,
|
205 |
+
interactive=True,
|
206 |
+
label="Sampling temperature",
|
207 |
+
info="Higher values will produce more diverse outputs.",
|
208 |
+
)
|
209 |
+
top_p = gr.Slider(
|
210 |
+
minimum=0.01,
|
211 |
+
maximum=0.99,
|
212 |
+
value=0.8,
|
213 |
+
step=0.01,
|
214 |
+
interactive=True,
|
215 |
+
label="Top P",
|
216 |
+
info="Higher values is equivalent to sampling more low-probability tokens.",
|
217 |
+
)
|
218 |
+
|
219 |
+
|
220 |
+
chatbot = gr.Chatbot(
|
221 |
+
label="Idefics2",
|
222 |
+
avatar_images=[None, BOT_AVATAR],
|
223 |
+
# height=750,
|
224 |
+
)
|
225 |
+
|
226 |
+
|
227 |
+
with gr.Blocks(fill_height=True, css=".message-wrap.svelte-1lcyrx4>div.svelte-1lcyrx4 img { width: auto; max-width: 30%; height: auto; max-height: 30%; }") as demo:
|
228 |
+
decoding_strategy.change(
|
229 |
+
fn=lambda selection: gr.Slider(
|
230 |
+
visible=(
|
231 |
+
selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"]
|
232 |
+
)
|
233 |
+
),
|
234 |
+
inputs=decoding_strategy,
|
235 |
+
outputs=temperature,
|
236 |
+
)
|
237 |
+
decoding_strategy.change(
|
238 |
+
fn=lambda selection: gr.Slider(
|
239 |
+
visible=(
|
240 |
+
selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"]
|
241 |
+
)
|
242 |
+
),
|
243 |
+
inputs=decoding_strategy,
|
244 |
+
outputs=repetition_penalty,
|
245 |
+
)
|
246 |
+
decoding_strategy.change(
|
247 |
+
fn=lambda selection: gr.Slider(visible=(selection in ["Top P Sampling"])),
|
248 |
+
inputs=decoding_strategy,
|
249 |
+
outputs=top_p,
|
250 |
+
)
|
251 |
+
examples = [{"text": "How many items are sold?", "files":["./example_images/docvqa_example.png"]},
|
252 |
+
{"text": "What is this UI about?", "files":["./example_images/s2w_example.png"]},
|
253 |
+
{"text": "I want to go somewhere similar to the one in the photo. Give me destinations and travel tips.", "files":["./example_images/travel_tips.jpg"]},
|
254 |
+
{"text": "Can you tell me a very short story based on this image?", "files":["./example_images/chicken_on_money.png"]},
|
255 |
+
{"text": "Where is this pastry from?", "files":["./example_images/baklava.png"]},
|
256 |
+
{"text": "How much percent is the order status?", "files":["./example_images/dummy_pdf.png"]},
|
257 |
+
{"text":"As an art critic AI assistant, could you describe this painting in details and make a thorough critic?.", "files":["./example_images/art_critic.jpg"]}
|
258 |
+
]
|
259 |
+
description = "Try [IDEFICS2-8B](https://huggingface.co/HuggingFaceM4/idefics2-8b), the instruction fine-tuned IDEFICS2 in this demo. 💬 IDEFICS2 is a state-of-the-art vision language model in various benchmarks. To get started, upload an image and write a text prompt or try one of the examples. You can also play with advanced generation parameters. To learn more about IDEFICS2, read [the blog](https://huggingface.co/blog/idefics2). Note that this model is not as chatty as the upcoming chatty model, and it will give shorter answers."
|
260 |
+
|
261 |
+
|
262 |
+
gr.ChatInterface(
|
263 |
+
fn=model_inference,
|
264 |
+
chatbot=chatbot,
|
265 |
+
examples=examples,
|
266 |
+
description=description,
|
267 |
+
title="Idefics2 Playground 🐶 ",
|
268 |
+
multimodal=True,
|
269 |
+
additional_inputs=[decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p],
|
270 |
+
)
|
271 |
+
|
272 |
+
demo.launch(debug=True)
|