Spaces:
Runtime error
Runtime error
File size: 17,166 Bytes
ec50e73 792e854 92f5f56 a2496cb 34d389e 92f5f56 488a53d ec50e73 dc6da18 bf512c7 ec50e73 bf512c7 1430cb0 ec50e73 dc6da18 ec50e73 792e854 ec50e73 34d389e db79ccc f84e07f ec50e73 4273b28 ec50e73 0113778 f9f9d0b 0113778 34d389e 192577c f9f9d0b 0113778 f9f9d0b ec50e73 02dcc48 38ae6ff 02dcc48 38ae6ff 02dcc48 38ae6ff 02dcc48 38ae6ff 02dcc48 8d96ce6 2a3ed29 07eac8b 62100e3 07eac8b 02dcc48 ec50e73 0a45780 ec50e73 adf07d5 ec50e73 34d389e bf512c7 ec50e73 23c98b1 65a7de2 bc1a623 65a7de2 34d389e 23c98b1 ec50e73 adf07d5 dc6da18 ec50e73 adf07d5 ec50e73 4273b28 1430cb0 ec50e73 adf07d5 34d389e 4273b28 dc6da18 adf07d5 4273b28 adf07d5 4273b28 dc6da18 34d389e dc6da18 ec50e73 adf07d5 4273b28 dc6da18 34d389e dc6da18 ec50e73 dc6da18 4273b28 dc6da18 34d389e dc6da18 38fdad8 adf07d5 4273b28 adf07d5 dc6da18 adf07d5 34d389e adf07d5 ec50e73 adf07d5 34d389e dc6da18 4273b28 34d389e adf07d5 dc6da18 adf07d5 dc6da18 adf07d5 ec50e73 dc6da18 adf07d5 dc6da18 adf07d5 34d389e 6c67beb 914f129 23c98b1 bc1a623 23c98b1 bc1a623 7cae97c bc1a623 7cae97c 914f129 bc1a623 914f129 bc1a623 23c98b1 adf07d5 84228e7 adf07d5 76f94cd adf07d5 76f94cd adf07d5 ec50e73 91e4b1d 6aa3f3a 91e4b1d 742ceb6 91e4b1d 1430cb0 34d389e 1430cb0 34d389e 1430cb0 34d389e 1430cb0 34d389e 1430cb0 71338f5 34d389e 6aa3f3a 91e4b1d ec50e73 adf07d5 ec50e73 6aa3f3a ec50e73 6aa3f3a adf07d5 34d389e ec50e73 adf07d5 ec50e73 3edb153 34d389e 3edb153 adf07d5 ec50e73 1bfcc16 adf07d5 ec50e73 91e4b1d 02dcc48 adf07d5 00da7f2 34d389e ec50e73 1430cb0 914f129 23c98b1 1430cb0 23c98b1 1430cb0 914f129 23c98b1 1430cb0 23c98b1 1430cb0 ec50e73 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 |
import os
import subprocess
# Install flash attention
subprocess.run(
"pip install flash-attn --no-build-isolation",
env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
shell=True,
)
import copy
import spaces
import time
import torch
from threading import Thread
from typing import List, Dict, Union
import urllib
from urllib.parse import urlparse
from PIL import Image
import io
import pandas as pd
import datasets
import json
import requests
import gradio as gr
from transformers import AutoProcessor, TextIteratorStreamer
from transformers import Idefics2ForConditionalGeneration
DEVICE = torch.device("cuda")
MODELS = {
"idefics2-8b-chatty": Idefics2ForConditionalGeneration.from_pretrained(
"HuggingFaceM4/idefics2-8b-chatty",
torch_dtype=torch.bfloat16,
_attn_implementation="flash_attention_2",
trust_remote_code=True,
token=os.environ["HF_AUTH_TOKEN"],
).to(DEVICE),
}
PROCESSOR = AutoProcessor.from_pretrained(
"HuggingFaceM4/idefics2-8b",
token=os.environ["HF_AUTH_TOKEN"],
)
SYSTEM_PROMPT = [
{
"role": "system",
"content": [
{
"type": "text",
"text": "The following is a conversation between Idefics2, a highly knowledgeable and intelligent visual AI assistant created by Hugging Face, referred to as Assistant, and a human user called User. In the following interactions, User and Assistant will converse in natural language, and Assistant will do its best to answer Userβs questions. Assistant has the ability to perceive images and reason about the content of visual inputs. Assistant was built to be respectful, polite and inclusive. It knows a lot, and always tells the truth. When prompted with an image, it does not make up facts.",
},
],
}
]
examples_path = os.path.dirname(__file__)
EXAMPLES = [
[
{
"text": "What's in the image?",
"files": [f"{examples_path}/example_images/plant_bulb.webp"],
}
],
[
{
"text": "What's funny about this image?",
"files": [f"{examples_path}/example_images/pope_doudoune.webp"],
}
],
[
{
"text": "Why is this image cute",
"files": [
f"{examples_path}/example_images/kittens-cats-pet-cute-preview.jpg"
],
}
],
[
{
"text": "Describe the image",
"files": [f"{examples_path}/example_images/baguettes_guarding_paris.png"],
}
],
[
{
"text": "What's unusual about this image?",
"files": [f"{examples_path}/example_images/dragons_playing.png"],
}
],
[
{
"text": "Read what's written on the paper",
"files": [f"{examples_path}/example_images/paper_with_text.png"],
}
],
[
{
"text": "Can this happen in real life?",
"files": [f"{examples_path}/example_images/elephant_spider_web.webp"],
}
],
[
{
"text": "Can you explain this meme?",
"files": [f"{examples_path}/example_images/running_girl_meme.webp"],
}
],
]
API_TOKEN = os.getenv("HF_AUTH_TOKEN")
HF_WRITE_TOKEN = os.getenv("HF_WRITE_TOKEN")
# IDEFICS_LOGO = "https://huggingface.co/spaces/HuggingFaceM4/idefics_playground/resolve/main/IDEFICS_logo.png"
BOT_AVATAR = "IDEFICS_logo.png"
# Chatbot utils
def turn_is_pure_media(turn):
return turn[1] is None
def load_image_from_url(url):
with urllib.request.urlopen(url) as response:
image_data = response.read()
image_stream = io.BytesIO(image_data)
image = Image.open(image_stream)
return image
def img_to_bytes(image_path):
image = Image.open(image_path).convert(mode='RGB')
buffer = io.BytesIO()
image.save(buffer, format="JPEG")
img_bytes = buffer.getvalue()
image.close()
return img_bytes
def format_user_prompt_with_im_history_and_system_conditioning(
user_prompt, chat_history
) -> List[Dict[str, Union[List, str]]]:
"""
Produces the resulting list that needs to go inside the processor.
It handles the potential image(s), the history and the system conditionning.
"""
resulting_messages = copy.deepcopy(SYSTEM_PROMPT)
resulting_images = []
for resulting_message in resulting_messages:
if resulting_message["role"] == "user":
for content in resulting_message["content"]:
if content["type"] == "image":
resulting_images.append(load_image_from_url(content["image"]))
# Format history
for turn in chat_history:
if not resulting_messages or (
resulting_messages and resulting_messages[-1]["role"] != "user"
):
resulting_messages.append(
{
"role": "user",
"content": [],
}
)
if turn_is_pure_media(turn):
media = turn[0][0]
resulting_messages[-1]["content"].append({"type": "image"})
resulting_images.append(Image.open(media))
else:
user_utterance, assistant_utterance = turn
resulting_messages[-1]["content"].append(
{"type": "text", "text": user_utterance.strip()}
)
resulting_messages.append(
{
"role": "assistant",
"content": [{"type": "text", "text": user_utterance.strip()}],
}
)
# Format current input
if not user_prompt["files"]:
resulting_messages.append(
{
"role": "user",
"content": [{"type": "text", "text": user_prompt["text"]}],
}
)
else:
# Choosing to put the image first (i.e. before the text), but this is an arbiratrary choice.
resulting_messages.append(
{
"role": "user",
"content": [{"type": "image"}] * len(user_prompt["files"])
+ [{"type": "text", "text": user_prompt["text"]}],
}
)
resulting_images.extend([Image.open(path) for path in user_prompt["files"]])
return resulting_messages, resulting_images
def extract_images_from_msg_list(msg_list):
all_images = []
for msg in msg_list:
for c_ in msg["content"]:
if isinstance(c_, Image.Image):
all_images.append(c_)
return all_images
@spaces.GPU(duration=180)
def model_inference(
user_prompt,
chat_history,
model_selector,
decoding_strategy,
temperature,
max_new_tokens,
repetition_penalty,
top_p,
):
if user_prompt["text"].strip() == "" and not user_prompt["files"]:
gr.Error("Please input a query and optionally image(s).")
if user_prompt["text"].strip() == "" and user_prompt["files"]:
gr.Error("Please input a text query along the image(s).")
streamer = TextIteratorStreamer(
PROCESSOR.tokenizer,
skip_prompt=True,
timeout=5.0,
)
# Common parameters to all decoding strategies
# This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies
generation_args = {
"max_new_tokens": max_new_tokens,
"repetition_penalty": repetition_penalty,
"streamer": streamer,
}
assert decoding_strategy in [
"Greedy",
"Top P Sampling",
]
if decoding_strategy == "Greedy":
generation_args["do_sample"] = False
elif decoding_strategy == "Top P Sampling":
generation_args["temperature"] = temperature
generation_args["do_sample"] = True
generation_args["top_p"] = top_p
# Creating model inputs
(
resulting_text,
resulting_images,
) = format_user_prompt_with_im_history_and_system_conditioning(
user_prompt=user_prompt,
chat_history=chat_history,
)
prompt = PROCESSOR.apply_chat_template(resulting_text, add_generation_prompt=True)
inputs = PROCESSOR(
text=prompt,
images=resulting_images if resulting_images else None,
return_tensors="pt",
)
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
generation_args.update(inputs)
# # The regular non streaming generation mode
# _ = generation_args.pop("streamer")
# generated_ids = MODELS[model_selector].generate(**generation_args)
# generated_text = PROCESSOR.batch_decode(generated_ids[:, generation_args["input_ids"].size(-1): ], skip_special_tokens=True)[0]
# return generated_text
# The streaming generation mode
thread = Thread(
target=MODELS[model_selector].generate,
kwargs=generation_args,
)
thread.start()
print("Start generating")
acc_text = ""
for text_token in streamer:
time.sleep(0.04)
acc_text += text_token
if acc_text.endswith("<end_of_utterance>"):
acc_text = acc_text[:-18]
yield acc_text
print("Success - generated the following text:", acc_text)
print("-----")
FEATURES = datasets.Features(
{
"model_selector": datasets.Value("string"),
"images": datasets.Sequence(datasets.Image(decode=True)),
"conversation": datasets.Sequence({"User": datasets.Value("string"), "Assistant": datasets.Value("string")}),
"decoding_strategy": datasets.Value("string"),
"temperature": datasets.Value("float32"),
"max_new_tokens": datasets.Value("int32"),
"repetition_penalty": datasets.Value("float32"),
"top_p": datasets.Value("int32"),
}
)
def flag_dope(
model_selector,
chat_history,
decoding_strategy,
temperature,
max_new_tokens,
repetition_penalty,
top_p,
):
images = []
conversation = []
for ex in chat_history:
if isinstance(ex[0], dict):
images.append(img_to_bytes(ex[0]["file"]["path"]))
else:
conversation.append({"User": ex[0], "Assistant": ex[1]})
data = {
"model_selector": [model_selector],
"images": [images],
"conversation": [conversation],
"decoding_strategy": [decoding_strategy],
"temperature": [temperature],
"max_new_tokens": [max_new_tokens],
"repetition_penalty": [repetition_penalty],
"top_p": [top_p],
}
try:
ds = datasets.load_dataset("HuggingFaceM4/dope-dataset-red-teaming", split="train", token=HF_WRITE_TOKEN)
new_data = datasets.Dataset.from_dict(data, features=FEATURES)
hf_dataset = datasets.concatenate_datasets([ds,new_data])
except Exception:
hf_dataset = datasets.Dataset.from_dict(data, features=FEATURES)
hf_dataset.push_to_hub( "HuggingFaceM4/dope-dataset-red-teaming", split="train", token=HF_WRITE_TOKEN, private=True)
def flag_problematic(
model_selector,
chat_history,
decoding_strategy,
temperature,
max_new_tokens,
repetition_penalty,
top_p,
):
images = []
conversation = []
for ex in chat_history:
if isinstance(ex[0], dict):
images.append(img_to_bytes(ex[0]["file"]["path"]))
else:
conversation.append({"User": ex[0], "Assistant": ex[1]})
data = {
"model_selector": [model_selector],
"images": [images],
"conversation": [conversation],
"decoding_strategy": [decoding_strategy],
"temperature": [temperature],
"max_new_tokens": [max_new_tokens],
"repetition_penalty": [repetition_penalty],
"top_p": [top_p],
}
try:
ds = datasets.load_dataset("HuggingFaceM4/problematic-dataset-red-teaming", split="train", token=HF_WRITE_TOKEN)
new_data = datasets.Dataset.from_dict(data, features=FEATURES)
hf_dataset = datasets.concatenate_datasets([ds,new_data])
except Exception:
hf_dataset = datasets.Dataset.from_dict(data, features=FEATURES)
hf_dataset.push_to_hub( "HuggingFaceM4/problematic-dataset-red-teaming", split="train", token=HF_WRITE_TOKEN, private=True)
# Hyper-parameters for generation
max_new_tokens = gr.Slider(
minimum=8,
maximum=1024,
value=512,
step=1,
interactive=True,
label="Maximum number of new tokens to generate",
)
repetition_penalty = gr.Slider(
minimum=0.01,
maximum=5.0,
value=1.1,
step=0.01,
interactive=True,
label="Repetition penalty",
info="1.0 is equivalent to no penalty",
)
decoding_strategy = gr.Radio(
[
"Greedy",
"Top P Sampling",
],
value="Greedy",
label="Decoding strategy",
interactive=True,
info="Higher values is equivalent to sampling more low-probability tokens.",
)
temperature = gr.Slider(
minimum=0.0,
maximum=5.0,
value=0.4,
step=0.1,
visible=False,
interactive=True,
label="Sampling temperature",
info="Higher values will produce more diverse outputs.",
)
top_p = gr.Slider(
minimum=0.01,
maximum=0.99,
value=0.8,
step=0.01,
visible=False,
interactive=True,
label="Top P",
info="Higher values is equivalent to sampling more low-probability tokens.",
)
chatbot = gr.Chatbot(
label="Idefics2-Chatty",
avatar_images=[None, BOT_AVATAR],
height=450,
)
# Using Flagging for saving dope and problematic examples
# Dope examples flagging
# gr.Markdown("""## How to use?
# There are two ways to provide image inputs:
# - Using the image box on the left panel
# - Using the inline syntax: `text<fake_token_around_image><image:URL_IMAGE><fake_token_around_image>text`
# The second syntax allows inputting an arbitrary number of images.""")
image_flag = gr.Image(visible=False)
with gr.Blocks(
fill_height=True,
css=""".gradio-container .avatar-container {height: 40px width: 40px !important;}""",
) as demo:
gr.Markdown("# πΆ Idefics2-Chatty Playground πΆ")
gr.Markdown("In this demo you'll be able to chat with [Idefics2-8B-chatty](https://huggingface.co/HuggingFaceM4/idefics2-8b-chatty), a variant of [Idefics2-8B](https://huggingface.co/HuggingFaceM4/idefics2-8b-chatty) further fine-tuned on chat datasets")
gr.Markdown("If you want to learn more about Idefics2 and its variants, you can check our [blog post](https://huggingface.co/blog/idefics2).")
# model selector should be set to `visbile=False` ultimately
with gr.Row(elem_id="model_selector_row"):
model_selector = gr.Dropdown(
choices=MODELS.keys(),
value=list(MODELS.keys())[0],
interactive=True,
show_label=False,
container=False,
label="Model",
visible=False,
)
decoding_strategy.change(
fn=lambda selection: gr.Slider(
visible=(
selection
in [
"contrastive_sampling",
"beam_sampling",
"Top P Sampling",
"sampling_top_k",
]
)
),
inputs=decoding_strategy,
outputs=temperature,
)
decoding_strategy.change(
fn=lambda selection: gr.Slider(
visible=(
selection
in [
"contrastive_sampling",
"beam_sampling",
"Top P Sampling",
"sampling_top_k",
]
)
),
inputs=decoding_strategy,
outputs=repetition_penalty,
)
decoding_strategy.change(
fn=lambda selection: gr.Slider(visible=(selection in ["Top P Sampling"])),
inputs=decoding_strategy,
outputs=top_p,
)
gr.ChatInterface(
fn=model_inference,
chatbot=chatbot,
examples=EXAMPLES,
multimodal=True,
cache_examples=False,
additional_inputs=[
model_selector,
decoding_strategy,
temperature,
max_new_tokens,
repetition_penalty,
top_p,
],
)
with gr.Group():
with gr.Row():
with gr.Column(scale=1, min_width=50):
dope_bttn = gr.Button("Dopeπ₯")
with gr.Column(scale=1, min_width=50):
problematic_bttn = gr.Button("Problematicπ¬")
dope_bttn.click(
fn=flag_dope,
inputs=[
model_selector,
chatbot,
decoding_strategy,
temperature,
max_new_tokens,
repetition_penalty,
top_p,
],
outputs=None,
preprocess=False,
)
problematic_bttn.click(
fn=flag_problematic,
inputs=[
model_selector,
chatbot,
decoding_strategy,
temperature,
max_new_tokens,
repetition_penalty,
top_p,
],
outputs=None,
preprocess=False,
)
demo.launch()
|