File size: 44,582 Bytes
857c2e9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 | import time
from swift.llm import (
get_model_tokenizer, get_template, ModelType, load_dataset, EncodePreprocessor
)
from swift.utils import seed_everything
from swift.tuners import Swift
from typing import Any, Dict, List, Mapping, Optional, Tuple, Union,Literal
from swift.llm import RequestConfig,InferRequest,TemplateInputs
import torch
from loguru import logger
import copy
from PIL import Image
import sys
import os
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
from data_processing_vlm import DataProcessor,trojectory_example_prompt,denormalize_with_params
from video_tool import images_get_from_video, video_trajectory
import tqdm
import math
from collections import defaultdict
import heapq
from scipy.spatial import cKDTree
from scipy.stats import spearmanr
import random
import numpy as np
def clip_one(image):
width, height = image.size
if width > height:
left = (width - height) // 2
right = left + height
top = 0
bottom = height
else:
top = (height - width) // 2
bottom = top + width
left = 0
right = width
square_image = image.crop((left, top, right, bottom))
image=square_image
return image
def to_device(data: Any, device: Union[str, torch.device, int]) -> Any:
"""Move inputs to a device"""
if isinstance(data, Mapping):
return type(data)({k: to_device(v, device) for k, v in data.items()})
elif isinstance(data, (tuple, list)):
return type(data)(to_device(v, device) for v in data)
elif isinstance(data, torch.Tensor):
return data.to(device=device)
else:
return data
class GAC_model():
def __init__(self,tag='critic'):
self.tag=tag
self.temperature=1
self.max_tokens=10240
self.do_sample=True
self.top_logprobs=10
self.logprobs=True
self.top_k=None
self.system_prompt=None
self.action_data={}
self.critic_data={}
self.sft_dataset=None
self.dataclient=DataProcessor()
self.dataclient.prompt_templete['v3']="Image-1: <image>\nImage-2: <image>\nCompare two images and evaluate whether the second image is closer to achieving task objectives compared to the first image. + score means the second image is closer, - score means the first image is closer\nResponse the relative progressing of target task follow <score>. The target task is: <task> {} </task> <score>"
self.dataclient.prompt_templete['v3_think']="0% <image>\nThis image is the trajectory beginning of the following two images\nImage-1: <image>\nImage-2: <image>\nCompare two images and evaluate whether the second image is closer to achieving task objectives compared to the first image. + score means the second image is closer, - score means the first image is closer\nResponse the relative progressing of target task follow <score>. The target task is: <task> {} </task> <score>"
def _songling_process(self,action):
action=copy.deepcopy(action)
for i in range(len(action)):
if i in [0, 1, 2]:
action[i]=action[i]//1000
elif i in [3,4,5]:
action[i]=action[i]//1000
elif i ==6:
pass
else:
pass
return action
def format_state(self,state,gripper_format=False):
state=self._songling_process(state)
if gripper_format:
state_open=round((state[6]//10000)/7.0,1)
state[6]=state_open
else:
state[6]=state[6]//1000
return state
def get_score_prompt(self,task,trajectory_len=0,think=False):
"two or len+3 image"
if trajectory_len>0:
trajectory_prompt=trojectory_example_prompt(list(range(trajectory_len)),task=task)
full_prompt=trajectory_prompt+self.dataclient.prompt_templete['v3_think'].format(task)
else:
full_prompt=self.dataclient.prompt_templete['v3'].format(task)
if think:
full_prompt+=self.dataclient.prompt_templete['think']
return full_prompt
def get_done_prompt(self,task):
"one image"
return self.dataclient.prompt_templete["task_done"].format(task)
def get_in_context_done_prompt(self,task=None):
"two image"
if task:
return self.dataclient.prompt_templete["context_task_done"].format(task,task)
else:
return self.dataclient.prompt_templete["image_done"]
def get_task_prompt(self):
"two image"
return self.dataclient.prompt_templete["task_vqa"]
def get_action_inverse_prompt(self):
"two image"
return self.dataclient.prompt_templete["action_inverse"]
def get_action_score_prompt(self,task,state,action):
"one image"
format_fuction=self.dataclient.action_format['songling']
action_score_prompt=self.dataclient.prompt_templete['task_action_score'].format(format_fuction(state),format_fuction(action),task)
return action_score_prompt
def get_action_prompt(self,task,view_num=1,position_output=False,simple=False,state='type',output_num=1,think=False):
if simple:
format_fuction=self.dataclient.action_format_simple['songling']
action_key='task_action_simple'
if position_output:
action_key='task_position_simple'
else:
format_fuction=self.dataclient.action_format['songling']
action_key='task_action'
if position_output:
action_key='task_position'
"view_num(max 3) image"
image_prompt=self.dataclient.image_prompt_templete[view_num]
full_prompt=image_prompt+self.dataclient.prompt_templete[action_key].format(format_fuction(state,state=position_output),task)
if output_num>1:
full_prompt+=f'*{output_num}'
if think:
full_prompt+=self.dataclient.prompt_templete['think']
return full_prompt
def get_fast_action_prompt(self,task,view_num=1,position_output=False,simple=False,state='type',output_num=1,think=False):
full_prompt = self.get_action_prompt(task,view_num,position_output,simple,state,output_num,think)
full_prompt+='<chunks>'
return full_prompt
def set_config(self):
self.request_config = RequestConfig(
max_tokens=self.max_tokens,
temperature=self.temperature,
top_k=self.top_k,
logprobs=self.logprobs,
top_logprobs=self.top_logprobs,
# repetition_penalty=args.repetition_penalty,
# stop=args.stop_words,
# stream=True
)
self.model.generation_config.max_new_tokens = self.max_tokens
self.model.generation_config.do_sample=self.do_sample
self.model.generation_config.temperature=self.temperature
def _get_internvl2_per_token_logps(self, model, inputs):
from trl.trainer.utils import selective_log_softmax
logits_to_keep = inputs['logits_to_keep']
input_ids = inputs['input_ids']
inputs = {
k: v
for k, v in inputs.items() if k not in
['logits_to_keep', 'completion_mask', 'ref_per_token_logps', 'advantages', 'old_per_token_logps']
}
_, inputs = self.template.pre_forward_hook(self.model, None, inputs)
logits = model(**inputs).logits
# exclude the last logit: it corresponds to the next token pred
logits = logits[:, -(logits_to_keep + 1):-1, :]
logits = logits / self.temperature
input_ids = input_ids[:, -logits_to_keep:]
return selective_log_softmax(logits, input_ids)
def init_model(self,model_path,model_type='internvl2',device_map:str = 'auto',torch_dtype=torch.bfloat16,adapter: str = None):
"""
Args:
device_map: ['auto', 'cuda:0',...]
"""
template_type = model_type
# print(f'template_type: {template_type}')
self.model, tokenizer = get_model_tokenizer(model_id_or_path=model_path,
model_type=model_type,
torch_dtype=torch_dtype,
device_map=device_map,
attn_impl = 'flash_attn')
self.template = get_template(template_type, tokenizer)
if adapter:
self.model = Swift.from_pretrained(self.model, adapter, adapter_name=None)
from swift.llm import PtEngine
from swift.plugin import InferStats
self.engine = PtEngine.from_model_template(self.model, self.template, max_batch_size=0)
self.infer_stats = InferStats()
seed_everything(42)
logger.success("model initialized successfully")
def chat(self,infer_requests):
start_t=time.time()
response_list = self.engine.infer(
infer_requests, template=self.template, request_config=self.request_config, metrics=[self.infer_stats])
end_t=time.time()
infer_time=end_t-start_t
return response_list,infer_time
def results_format(self,response_list,infer_requests,rich=False):
infer_requests=copy.deepcopy(infer_requests)
answers=[]
for i in range(len(response_list)):
if rich:
rich_answer=''
for one in response_list[i].choices[0].logprobs['content'][:-1]:
if one['token'].isdigit():
temp_num=0
temp_weight=0
top_prob=math.e**one['top_logprobs'][0]['logprob']
for one_tops in one['top_logprobs']:
top_num=one_tops['token']
if top_num.isdigit():
prob=math.e**one_tops['logprob']
if prob>top_prob*0.1:
temp_num+=float(top_num)*prob
temp_weight+=prob
rich_answer+="{:.1f}".format(temp_num/temp_weight)
else:
rich_answer+=one['token']
answers.append(rich_answer)
else:
answers.append(response_list[i].choices[0].message.content)
infer_requests[i].messages.append({'role': 'assistant', 'content': response_list[i].choices[0].message.content})
return answers,infer_requests
def fast_results_format(self,response_list,infer_requests,tokenizer,time_horizon=10,action_dim=7):
infer_requests=copy.deepcopy(infer_requests)
answers=[]
for i in range(len(response_list)):
temp=[]
text=[]
for one in response_list[i].choices[0].logprobs['content'][:-1]:
ids=tokenizer.encode(one['token'],add_special_tokens=False)
if 92537-ids[0]<=2048:
temp.extend(ids)
# temp.extend(tokenizer.encode(one['token'],add_special_tokens=False))
temp=[[92537-one for one in temp]]
fast_chunk=tokenizer.fasttokenizer.decode(temp,time_horizon=time_horizon,action_dim=action_dim)
infer_requests[i].messages.append({'role': 'assistant', 'content': response_list[i].choices[0].message.content})
return fast_chunk,infer_requests
def denormalize_fastchunk(self,fast_chunk,params):
return denormalize_with_params(fast_chunk,params)
def set_system_prompt(self,system_prompt=None):
if system_prompt is not None:
self.system_prompt=system_prompt
elif system_prompt=='default':
self.system_prompt=None
else:
self.system_prompt='You are a visual-language assistant designed to interpret spatial and task-related information from images and text. Provide precise, context-aware responses and actionable guidance to assist in achieving task objectives.'
def _process_image_to_pil(self, image_input: Union[str, Image.Image]) -> Image.Image:
"""
将各种格式的图像输入转换为448x448的PIL.Image
Args:
image_input: 图像URL、文件路径、base64编码字符串或PIL.Image对象
Returns:
PIL.Image: 调整为448x448尺寸的PIL图像
"""
pil_image = None
if isinstance(image_input, Image.Image):
pil_image = image_input
elif isinstance(image_input, str):
if image_input.startswith(('http://', 'https://')):
response = requests.get(image_input)
pil_image = Image.open(BytesIO(response.content))
elif image_input.startswith('data:image'):
header, encoded = image_input.split(',', 1)
image_data = base64.b64decode(encoded)
pil_image = Image.open(BytesIO(image_data))
elif len(image_input) > 100 and not '/' in image_input and not '\\' in image_input:
try:
image_data = base64.b64decode(image_input)
pil_image = Image.open(BytesIO(image_data))
except:
pil_image = Image.open(image_input)
else:
pil_image = Image.open(image_input)
else:
raise ValueError(f"不支持的图像输入类型: {type(image_input)}")
if pil_image.mode != 'RGB':
pil_image = pil_image.convert('RGB')
pil_image = pil_image.resize((448, 448), Image.Resampling.LANCZOS)
return pil_image
def get_infer_requests(self,prompt,images:List[Union[str, Image.Image]] = None):
"""
Args:
prompt: 提示文本
images: 可以是图像URL、文件路径、base64编码字符串或PIL.Image对象
"""
if type(prompt)==str:
prompt=[prompt]
images=[images]
infer_requests=[]
for i in range(len(prompt)):
one_input=[]
if self.system_prompt:
one_input.append(
{
'role': 'system',
'content': self.system_prompt
}
)
one_input.append(
{
'role': 'user',
'content': prompt[i]
}
)
processed_images = None
if images[i] is not None:
if isinstance(images[i], list):
processed_images = [self._process_image_to_pil(img) for img in images[i] if img is not None]
else:
processed_images = [self._process_image_to_pil(images[i])]
infer_requests.append(TemplateInputs(
messages=one_input,
images=processed_images,
))
# infer_requests.append(TemplateInputs(
# messages=one_input,
# images=images[i],
# ))
return infer_requests
def get_logprobs(self,infer_requests,return_mask=False,digit=False):
old_mode=self.template.mode
self.template.set_mode('train')
mini_batch_encoded_inputs = [self.template.encode(infer_request) for infer_request in infer_requests]
mini_batch_encoded_inputs = to_device(
self.template.data_collator(mini_batch_encoded_inputs), self.model.device)
labels = mini_batch_encoded_inputs.pop('labels')
logits_to_keep = (labels.shape[-1] - (torch.ne(labels, -100).int().argmax(-1))).max().item()
mini_batch_encoded_inputs['logits_to_keep'] = logits_to_keep
mini_batch_encoded_inputs['completion_mask'] = labels[:, -logits_to_keep:] != -100
per_token_logps = self._get_internvl2_per_token_logps(self.model, mini_batch_encoded_inputs)
self.template.set_mode(old_mode)
if return_mask:
if digit:
tokenizer = self.template.tokenizer
digit_completion_mask = torch.zeros_like(labels[:, -logits_to_keep:], dtype=torch.bool)
for i in range(labels.shape[0]):
for j in range(labels.shape[1] - logits_to_keep, labels.shape[1]):
if labels[i, j] != -100:
token_id = labels[i, j].item()
token_str = tokenizer.decode([token_id]).strip()
if re.search(r'[0-9+\-]', token_str):
digit_completion_mask[i, j - (labels.shape[1] - logits_to_keep)] = True
return per_token_logps, mini_batch_encoded_inputs['completion_mask'],digit_completion_mask
else:
return per_token_logps,mini_batch_encoded_inputs['completion_mask']
else:
return per_token_logps
def get_in_context_done(self,task:str,first_image:List[Image.Image],n_pre_image:List[Image.Image],now_image:List[Image.Image],ref_image_list:List[Image.Image],ref_num=9,rich=False):
"""
In context 的方式获取done
提供一段参考ref_image_list
first_image是你的轨迹的第一张图,这里为list是可以batch推理
n_pre_image是上一个时刻
now_image是当前时刻
"""
if ref_image_list is not None:
ref_images=[ref_image_list[0]]
delta=(len(ref_image_list)-1)/(ref_num-1)
for i in range(1,ref_num):
ref_images.append(ref_image_list[int(i*delta)])
else:
ref_num=0
one_prompt=self.get_score_prompt(task=task,trajectory_len=ref_num,think=True)
batch_prompt=[]
batch_image=[]
for i in range(len(now_image)):
batch_prompt.append(one_prompt)
batch_image.append(ref_images+[first_image[i],n_pre_image[i],now_image[i]])
infer_requests=self.get_infer_requests(prompt=batch_prompt,images=batch_image)
response_list,infer_time=self.chat(infer_requests)
answers_list,complete_requests_list=self.results_format(response_list,infer_requests,rich=rich)
think_pre_value_list=[]
think_post_value_list=[]
think_critic_list=[]
for one in answers_list:
one_critic=one.split('</think>')[1]
one_pre_value=one.split('first image progressing: ')[1].split('%')[0]
one_post_value=one.split('second image progressing: ')[1].split('%')[0]
think_critic_list.append("{:.1f}".format((100-float(one_pre_value))*float(one_critic)/100.0))
think_pre_value_list.append(one_pre_value)
think_post_value_list.append("{:.3f}".format(float(one_post_value)/100.0))
batch_done=think_post_value_list
batch_related_critic=think_critic_list
return batch_done,batch_related_critic
def web_trajectory_critic(self, task_description, main_video_path, reference_video_path=None, batch_num=20, ref_num=9,
think=False, skip=1, rich=False, reverse_eval=False, output_path=None, fps=None, frame_skip=False, addition_scale=1,
bias=0, positive_clip=0, negative_clip=0, related_critic=False, done_flag=False, in_context_done=False, done_threshold=False, video_output=True):
list_video=images_get_from_video(main_video_path)
ref_list=None
if reference_video_path is None:
ref_num=0
else:
ref_list=images_get_from_video(reference_video_path)
if done_flag:
if in_context_done:
done_ref=ref_list
else:
done_ref=None
done_list=self.get_trajectory_done(task=task_description,image_list=list_video,ref_image_list=done_ref,batch_num=batch_num,rich=True,ref_num=ref_num,threshold=done_threshold)
if frame_skip:
done_list=[done_list[i] for i in range(0,len(done_list),skip)]
else:
done_list=None
critic_list, think_value_list=self.get_trajectory_critic(task=task_description,image_list=list_video,ref_image_list=ref_list,batch_num=batch_num,ref_num=ref_num,think=think,skip=skip,rich=rich,reverse_eval=reverse_eval,frame_skip=frame_skip,addition_scale=addition_scale,bias=bias,positive_clip=positive_clip,negative_clip=negative_clip,related_critic=related_critic)
value_list=think_value_list
if video_output:
if ref_num>0:
ref_image_paths=ref_list
else:
ref_image_paths=None
if fps:
pass
else:
fps=5.0/skip
video_path = video_trajectory(
traj_id='temp',
view='0',
image_paths=[],
image_objects=list_video,
done_list=done_list,
critic_list=critic_list,
value_list=value_list,
task=task_description,
output_path=output_path,
fps=fps,
ref_image_paths=ref_image_paths,
n_num=ref_num
)
else:
video_path=None
return video_path,value_list,critic_list,done_list
def web_trajectory_done(self, task_description, main_image_path, reference_image_path=None, rich=False):
done_list=self.get_trajectory_done(task=task_description,image_list=[main_image_path],batch_num=1,rich=rich)
return f"Task: {task_description}\nCritic Score: {done_list[0]}"
def get_trajectory_task(self, image_list):
one_prompt=self.get_task_prompt()
infer_requests=self.get_infer_requests(prompt=one_prompt,images=[image_list[0],image_list[-1]])
response_list,infer_time=self.chat(infer_requests)
answers_list,complete_requests_list=self.results_format(response_list,infer_requests)
return answers_list
def get_trajectory_done(self,task:str, image_list:List[Image.Image], ref_image_list:List[Image.Image]=None,
batch_num:int=20,rich=False,ref_num=9,threshold=0,skip=1,goal_image:Image.Image=None):
"""
输入一条trajectory的所有图片,输出每张图片的done,0~1的突变值
当有ref_image_list时,进行过程判断,输出0~1的渐变值
当有gaol_image时,进行最终状态判断,输出0~1的突变值(与当有ref_image_list冲突),task=None时只依靠图片
"""
batch_prompt=[]
batch_image=[]
done_list=[]
if skip>1:
image_list=[image_list[i] for i in range(0,len(image_list),skip)]
# zero-shot testing
if ref_image_list is None and goal_image is None:
# for i in tqdm.tqdm(range(len(image_list)),desc='done processing'):
for i in range(len(image_list)):
one_prompt=self.get_done_prompt(task=task)
batch_prompt.append(one_prompt)
batch_image.append([image_list[i]])
if (i+1) % batch_num==0 or i==len(image_list)-1:
infer_requests=self.get_infer_requests(prompt=batch_prompt, images=batch_image)
response_list,infer_time=self.chat(infer_requests)
answers_list,complete_requests_list=self.results_format(response_list,infer_requests,rich=rich)
# print(f'infer_time:{infer_time}s')
# print(f'answers_list:{answers_list}')
done_list.extend(answers_list)
batch_prompt=[]
batch_image=[]
# ref image as example context
elif ref_image_list:
first_image=[]
n_pre_image=[]
now_image=[]
# for i in tqdm.tqdm(range(1,len(image_list)),desc='done processing'):
for i in range(1,len(image_list)):
first_image.append(image_list[0])
n_pre_image.append(image_list[i-1])
now_image.append(image_list[i])
if (i+1) % batch_num==0 or i==len(image_list)-1:
batch_done,batch_related_critic=self.get_in_context_done(task=task,first_image=first_image,n_pre_image=n_pre_image,now_image=now_image,ref_image_list=ref_image_list,ref_num=ref_num,rich=rich)
# print(f'answers_list:{batch_done}')
done_list.extend(batch_done)
first_image=[]
n_pre_image=[]
now_image=[]
done_list=[0]+done_list
# goal image as example context
else:
# for i in tqdm.tqdm(range(len(image_list)),desc='done processing'):
for i in range(len(image_list)):
one_prompt=self.get_in_context_done_prompt(task=task)
batch_prompt.append(one_prompt)
batch_image.append([goal_image,image_list[i]])
if (i+1) % batch_num==0 or i==len(image_list)-1:
infer_requests=self.get_infer_requests(prompt=batch_prompt,images=batch_image)
response_list,infer_time=self.chat(infer_requests)
answers_list,complete_requests_list=self.results_format(response_list,infer_requests,rich=rich)
# print(f'infer_time:{infer_time}s')
# print(f'answers_list:{answers_list}')
done_list.extend(answers_list)
batch_prompt=[]
batch_image=[]
done_list = [float(one) if float(one) > threshold else 0.0 for one in done_list]
return done_list
def get_trajectory_critic(self,task:str, image_list:List[Image.Image], ref_image_list:List[Image.Image]=None,
batch_num:int=20, ref_num=9, think=False, skip=1, rich=False, reverse_eval=False, frame_skip=True,
addition_scale=1, bias=0, related_critic=False, positive_clip=0, negative_clip=0, value_simple=True):
"""
输入一条trajectory的所有图片,输出每张图片的critic和processing value
可以给一条参考轨迹
"""
batch_prompt=[]
batch_image=[]
critic_list=[]
value_list=[]
if ref_image_list is not None:
ref_images=[ref_image_list[0]]
delta=(len(ref_image_list)-1)/(ref_num-1)
for i in range(1,ref_num):
ref_images.append(ref_image_list[int(i*delta)])
else:
ref_num=0
if frame_skip:
select_idx=range(skip,len(image_list),skip)
else:
select_idx=range(skip,len(image_list))
# for i in tqdm.tqdm(select_idx,desc='critic processing'):
for i in select_idx:
one_prompt=self.get_score_prompt(task=task,trajectory_len=ref_num,think=think)
batch_prompt.append(one_prompt)
if ref_image_list is not None:
if reverse_eval:
batch_image.append(ref_images+[image_list[0],image_list[i],image_list[i-skip]])
else:
batch_image.append(ref_images+[image_list[0],image_list[i-skip],image_list[i]])
else:
if reverse_eval:
batch_image.append([image_list[i],image_list[i-skip]])
else:
batch_image.append([image_list[i-skip],image_list[i]])
if (len(batch_prompt)) % batch_num==0 or len(critic_list)+len(batch_prompt)==len(select_idx):
infer_requests=self.get_infer_requests(prompt=batch_prompt,images=batch_image)
response_list,infer_time=self.chat(infer_requests)
answers_list,complete_requests_list=self.results_format(response_list,infer_requests,rich=rich)
# print(f'infer_time:{infer_time}s')
print(f'answers_list:{answers_list}')
critic_list.extend(answers_list)
batch_prompt=[]
batch_image=[]
if think:
think_pre_value_list=[]
think_post_value_list=[]
think_critic_list=[]
for one in critic_list:
one_critic=one.split('</think>')[1]
one_pre_value=one.split('first image progressing: ')[1].split('%')[0]
one_post_value=one.split('second image progressing: ')[1].split('%')[0]
think_critic_list.append(one_critic)
think_pre_value_list.append(one_pre_value)
think_post_value_list.append(one_post_value)
if reverse_eval:
temp=think_pre_value_list
think_pre_value_list=think_post_value_list
think_post_value_list=temp
think_critic_list=[0-float(one) for one in think_critic_list]
think_pre_value_list=think_pre_value_list+think_post_value_list[-skip:]
think_post_value_list=think_pre_value_list[:skip]+think_post_value_list
critic_list=think_critic_list
if frame_skip:
pass
else:
critic_list=[float(one)/skip for one in critic_list]
critic_list=[float(one)/addition_scale for one in critic_list]
value_list=self.critic_to_value_simple(critic_list,simple=value_simple)
else:
if reverse_eval:
critic_list=[0-float(one) for one in critic_list]
if frame_skip:
pass
else:
critic_list=[float(one)/skip for one in critic_list]
critic_list=[float(one)/addition_scale for one in critic_list]
value_list=self.critic_to_value_simple(critic_list,simple=value_simple)
if related_critic:
critic_list=[value_list[i]-value_list[i-1] for i in range(1,len(value_list))]
if bias!=0:
critic_list=[one+bias for one in critic_list]
if positive_clip!=0:
critic_list=[one if (one<0 or one>positive_clip) else 0 for one in critic_list]
if negative_clip!=0:
critic_list=[one if (one>0 or one<-negative_clip) else 0 for one in critic_list]
if related_critic:
value_list=[0]
for one in critic_list:
value_list.append(value_list[-1]+one)
else:
value_list=self.critic_to_value_simple(critic_list,simple=value_simple)
#这里的value也是done，越接近100完成度越高
return critic_list,value_list
def magic_smooth(self,task,file_path,hz=10,value_simple='mix_f',ref_image_list=None,ref_num=9,think=False,max_skip=3):
import os
import pickle
import gzip
import subprocess
from video_tool import read_data_and_create_video,images_get_from_video
if type(ref_image_list) == str:
if ref_image_list.endswith(".pkl.gz"):
ref_image_list = read_data_and_create_video(ref_image_list)
else:
ref_image_list=images_get_from_video(ref_image_list)
image_list=read_data_and_create_video(file_path, camera_name=None, create_video=False, output_path=None, fps=hz, start_frame=0, end_frame=None)
skip=int(hz/2)
critic_list,value_list=self.get_trajectory_critic(task=task,image_list=image_list,ref_image_list=ref_image_list,batch_num=10,ref_num=ref_num,think=think,skip=skip,rich=True,reverse_eval=False,frame_skip=True,addition_scale=1,bias=0,related_critic=False,positive_clip=0,negative_clip=0,value_simple=value_simple)
skip_idx_range=[]
skip_idxs=[]
i=0
while i <len(critic_list):
if float(critic_list[i])<=0:
skip_window=2
while i+skip_window<len(critic_list) and skip_window<=max_skip:
two_image=[image_list[0],image_list[i*skip],image_list[(i+skip_window)*skip]]
one_critic_list,_=self.get_trajectory_critic(task=task,image_list=two_image,ref_image_list=ref_image_list,batch_num=10,ref_num=ref_num,think=think,skip=1,rich=True,reverse_eval=False,frame_skip=True,addition_scale=1,bias=0,related_critic=False,positive_clip=0,negative_clip=0,value_simple=value_simple)
if float(one_critic_list[-1])<0:
skip_window+=1
else:
break
skip_idx_range.append([i,i+skip_window])
skip_idxs.extend(range(i*skip+1,(i+skip_window)*skip))
i+=skip_window
print("skip_window:",skip_window)
else:
i+=1
with gzip.open(file_path, 'rb') as f:
original_data = pickle.load(f)
base_dir = os.path.dirname(file_path)
file_name = os.path.splitext(os.path.splitext(os.path.basename(file_path))[0])[0] # 去掉.pkl.gz
output_dir = os.path.join(base_dir, f"{file_name}_smoothed")
os.makedirs(output_dir, exist_ok=True)
new_data=[one for i,one in enumerate(original_data) if i not in skip_idxs]
new_images=[one for i,one in enumerate(image_list) if i not in skip_idxs]
data_file = os.path.join(output_dir, f"{file_name}_smoothed.pkl.gz")
with gzip.open(data_file, 'wb') as f:
pickle.dump(new_data, f)
video_file = os.path.join(output_dir, f"{file_name}_smoothed.mp4")
temp_dir = os.path.join(output_dir, f"temp")
os.makedirs(temp_dir, exist_ok=True)
try:
for j, img in enumerate(new_images):
frame_file = os.path.join(temp_dir, f"frame_{j:06d}.png")
if hasattr(img, 'save'):
img.save(frame_file)
else:
from PIL import Image
if isinstance(img, np.ndarray):
Image.fromarray(img).save(frame_file)
else:
Image.fromarray(np.array(img)).save(frame_file)
ffmpeg_cmd = [
'ffmpeg', '-y',
'-framerate', str(hz),
'-i', os.path.join(temp_dir, 'frame_%06d.png'),
'-c:v', 'libx264',
'-pix_fmt', 'yuv420p',
'-crf', '23',
video_file
]
subprocess.run(ffmpeg_cmd, check=True, capture_output=True)
print(f"Created video: {video_file}")
except Exception as e:
print(f"Error creating video for segment {i}: {e}")
finally:
import shutil
if os.path.exists(temp_dir):
shutil.rmtree(temp_dir)
return new_data
def magic_split(self,task,file_path,hz=10,value_simple='mix_f',done_assist=False,ref_image_list=None):
import os
import pickle
import gzip
import subprocess
from video_tool import read_data_and_create_video,images_get_from_video
from magic_detect import adaptive_peak_valley_detection,plot_results
if type(ref_image_list) == str:
if ref_image_list.endswith(".pkl.gz"):
ref_image_list = read_data_and_create_video(ref_image_list)
else:
ref_image_list=images_get_from_video(ref_image_list)
skip=int(hz/2)
done_len=4
gap_merge_len=4#一般看数据的平均记录周期
image_list=read_data_and_create_video(file_path, camera_name=None, create_video=False, output_path=None, fps=hz, start_frame=0, end_frame=None)
critic_list,value_list=self.get_trajectory_critic(task=task,image_list=image_list,ref_image_list=ref_image_list,batch_num=10,ref_num=9,think=False,skip=skip,rich=True,reverse_eval=False,frame_skip=True,addition_scale=1,bias=0,related_critic=False,positive_clip=0,negative_clip=0,value_simple=value_simple)
if done_assist:
done_list=self.get_trajectory_done(task=task,image_list=image_list,ref_image_list=None,batch_num=10,ref_num=2,rich=True,threshold=0.9,skip=skip)
assist_peaks={}
peak_window=[]
done_list=done_list+[0]
for k in range(len(done_list)):
if done_list[k]>=0.95:
peak_window.append(k)
else:
if len(peak_window)>=done_len:
assist_peaks[peak_window[done_len-1]]=peak_window
peak_window=[]
peak_list=list(assist_peaks.keys())
index=0
while index<len(peak_list)-1:
if assist_peaks[peak_list[index+1]][0]-assist_peaks[peak_list[index]][-1]<=gap_merge_len:
assist_peaks[peak_list[index]].extend(assist_peaks[peak_list[index+1]])
del assist_peaks[peak_list[index+1]]
del peak_list[index+1]
else:
index+=1
result = adaptive_peak_valley_detection(
value_list,
assist_peaks=peak_list,
window_size=None, # Adaptive window size
verbose=True,
min_distance=10,
prominence_threshold=0.1,
outlier_sensitivity=3
)
else:
result = adaptive_peak_valley_detection(
value_list,
window_size=None, # Adaptive window size
verbose=True,
min_distance=10,
prominence_threshold=0.01,
outlier_sensitivity=1.5
)
with gzip.open(file_path, 'rb') as f:
original_data = pickle.load(f)
base_dir = os.path.dirname(file_path)
file_name = os.path.splitext(os.path.splitext(os.path.basename(file_path))[0])[0] # 去掉.pkl.gz
output_dir = os.path.join(base_dir, f"{file_name}_segments")
os.makedirs(output_dir, exist_ok=True)
plot_results(value_list, result,save_path=os.path.join(output_dir, "peak_valley.png"))
peaks = [one*skip for one in result['peaks']]
valleys = [one*skip for one in result['valleys']]
ascending_dir = os.path.join(output_dir, "ascending")
descending_dir = os.path.join(output_dir, "descending")
os.makedirs(ascending_dir, exist_ok=True)
os.makedirs(descending_dir, exist_ok=True)
video_dir = os.path.join(output_dir, "videos")
os.makedirs(video_dir, exist_ok=True)
all_points = []
for p in peaks:
all_points.append((p, 'peak'))
for v in valleys:
all_points.append((v, 'valley'))
all_points.sort(key=lambda x: x[0])
segments = []
for i, (point_idx, point_type) in enumerate(all_points):
if i == 0:
prev_type='valley'
start_idx = 0
else:
start_idx=all_points[i-1][0]
prev_type = all_points[i-1][1]
end_idx = point_idx
if prev_type == 'valley' and point_type == 'peak':
segment_type = 'ascending'
elif prev_type == 'peak' and point_type == 'valley':
segment_type = 'descending'
else:
continue
data_segment = original_data[start_idx:end_idx]
image_segment = image_list[start_idx:end_idx]
segments.append({
'type': segment_type,
'start': start_idx,
'end': end_idx,
'data': data_segment,
'images': image_segment
})
if segments and segments[-1]['type'] == 'descending':
segments.pop()
for i, segment in enumerate(segments):
segment_type = segment['type']
task_name = self.get_trajectory_task(segment['images'])[0].split('</task>')[0].strip().replace(' ', '_').replace('.', '_').replace(',', '_')
if segment_type == 'ascending':
data_file = os.path.join(ascending_dir, f"{task_name}_{i:03d}.pkl.gz")
else:
data_file = os.path.join(descending_dir, f"{task_name}_{i:03d}.pkl.gz")
with gzip.open(data_file, 'wb') as f:
pickle.dump(segment['data'], f)
video_file = os.path.join(video_dir, f"{segment_type}_{task_name}_{i:03d}.mp4")
temp_dir = os.path.join(output_dir, f"temp_{i}")
os.makedirs(temp_dir, exist_ok=True)
try:
for j, img in enumerate(segment['images']):
frame_file = os.path.join(temp_dir, f"frame_{j:06d}.png")
if hasattr(img, 'save'):
img.save(frame_file)
else:
from PIL import Image
if isinstance(img, np.ndarray):
Image.fromarray(img).save(frame_file)
else:
Image.fromarray(np.array(img)).save(frame_file)
ffmpeg_cmd = [
'ffmpeg', '-y',
'-framerate', str(hz),
'-i', os.path.join(temp_dir, 'frame_%06d.png'),
'-c:v', 'libx264',
'-pix_fmt', 'yuv420p',
'-crf', '23',
video_file
]
subprocess.run(ffmpeg_cmd, check=True, capture_output=True)
print(f"Created video: {video_file}")
except Exception as e:
print(f"Error creating video for segment {i}: {e}")
finally:
import shutil
if os.path.exists(temp_dir):
shutil.rmtree(temp_dir)
print(f"Processing complete. Segments saved to: {output_dir}")
print(f"Total segments: {len(segments)}")
print(f"Ascending segments: {sum(1 for s in segments if s['type'] == 'ascending')}")
print(f"Descending segments: {sum(1 for s in segments if s['type'] == 'descending')}")
return segments
def critic_to_value_simple(self,critic_list,simple=True):
"""
将critic计算为0-100的value,输入需是一条轨迹按顺序的critic
"""
value_list=[0]
for i in range(len(critic_list)):
if float(critic_list[i])>0 or simple is True:
value_list.append(value_list[-1]+(100-value_list[-1])*float(critic_list[i])/100.0)
else:
if simple=='mix_f':
if value_list[-1]>50:
value_list.append(max(10,100-max((100-value_list[-1]),1.0)/(100+float(critic_list[i]))*100.0))
else:
value_list.append(value_list[-1]+value_list[-1]*float(critic_list[i])/100.0)
else:
value_list.append(100-max((100-value_list[-1]),1.0)/(100+float(critic_list[i]))*100.0)
return value_list
def compute_voc(self,values_list):
"""
计算Value-Order Correlation (VOC)
参数:
predicted_values: 模型预测的值序列，形状为(T,)，其中T是时间步数
返回:
voc: Value-Order Correlation值，范围从-1到1
"""
T = len(values_list)
time_order = np.arange(T)
correlation, _ = spearmanr(values_list, time_order)
return correlation
def compute_negative_rate(self,critic_list):
"""
计算逆序动作的比例
"""
negative_critic=[one for one in critic_list if one<0]
return float(len(negative_critic))/len(critic_list)
def import_external_file(file_path: str):
import importlib
file_path = os.path.abspath(os.path.expanduser(file_path))
py_dir, py_file = os.path.split(file_path)
assert os.path.isdir(py_dir), f'py_dir: {py_dir}'
sys.path.insert(0, py_dir)
return importlib.import_module(py_file.split('.', 1)[0]) |