File size: 10,224 Bytes
60fde3b | 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 | import os
import re
import copy
import logging
import traceback
from tqdm import tqdm
from tenacity import (
retry,
stop_after_attempt,
wait_random_exponential,
retry_if_exception_type,
before_sleep_log
)
from src.utils.image_utils import parse_multimodal_text, parse_multimodal_text_for_vllm
# 导入 Prompts
from src.llm_generation.prompts import (
TASK1_SYSTEM_PROMPT, TASK1_STEP_GENERATION_PROMPT, TASK1_FINAL_VERIFICATION_PROMPT,
TASK3_SYSTEM_PROMPT, TASK3_STEP_GENERATION_PROMPT, TASK3_FINAL_VERIFICATION_PROMPT
)
# 设置 logging
logger = logging.getLogger(__name__)
class CoTGenerator:
def __init__(self, client, image_root, model_name="gpt-4o"):
self.client = client
self.image_root = image_root
self.model_name = model_name
# 判断是否为 vllm,其和API采用两种图片加载方式
self.is_vllm = "VLLMClient" in client.__class__.__name__
self.prompt_map = {
"camera_view_prediction": {
"system": TASK1_SYSTEM_PROMPT,
"step": TASK1_STEP_GENERATION_PROMPT,
"final": TASK1_FINAL_VERIFICATION_PROMPT
},
"camera_view_ordering": {
"system": TASK3_SYSTEM_PROMPT,
"step": TASK3_STEP_GENERATION_PROMPT,
"final": TASK3_FINAL_VERIFICATION_PROMPT
}
}
def _get_content(self, text, img_map):
"""将 prompt 中的相对路径占位符替换为绝对路径格式"""
def repl(match):
content = match.group(1)
rel_path = img_map.get(content, content)
abs_path = os.path.join(self.image_root, rel_path)
return f"<image_start>[{abs_path}]<image_end>"
text_abs = re.sub(r"<image_start>\[(.*?)\]<image_end>", repl, text)
if self.is_vllm:
return parse_multimodal_text_for_vllm(text_abs)
else:
return parse_multimodal_text(text_abs)
@retry(
stop=stop_after_attempt(5),
wait=wait_random_exponential(min=1, max=60),
before_sleep=before_sleep_log(logger, logging.WARNING),
reraise=True
)
def _call_llm_with_retry(self, messages):
return self.client.call_chat(messages, model=self.model_name)
def process_single_entry(self, entry):
# 确定 task 种类
# task1: camera view prediction
# task3: camera view ordering
task_type = entry.get('task', 'camera_view_prediction')
# 根据种类选择对应的 prompt
prompts = self.prompt_map.get(task_type)
if not prompts:
raise ValueError(f"Unknown task type: {task_type}")
meta = entry['oracle_meta']
img_map = entry['images']
chain = meta['chain']
# 1. 初始化
# system prompt: system prompt template + "The question is: {question}"
history_messages = [{"role": "system", "content": prompts["system"]}]
# 把 question 加入历史对话中
question = self._get_content("**Question:\n**" + entry["question"], img_map)
history_messages.append({"role": "user", "content": question})
full_trace_parts = []
cot_list = []
current_angle_so_far = 0.0
# 2. 循环生成 Chain (Think 1...N)
for i, step in enumerate(chain):
# 这一步的目标图 I_k
next_key = step['result_image_key']
step_action = step['action']
# --- A. 准备 Context ---
# 这里的 context 是:历史消息 + 这一轮的新指令
current_turn_messages = copy.deepcopy(history_messages)
# 如果是第二步以后,需要给模型上一步的结果;将上一步的结果作为 User 的新输入放进去
if i > 0:
# 拿到上一步的 image 占位符,如 reasoning_image_1
prev_key = chain[i-1]['result_image_key']
# 构造 text instruction
# instruction: 这是上一步的结果,请继续
step_instruction_text = f"Here is the result of the previous step: <image_start>[{prev_key}]<image_end>. Proceed to the next step."
# 使用 _get_content 方法直接将图片嵌入,并得到 content
step_content = self._get_content(step_instruction_text, img_map)
current_turn_messages.append({"role": "user", "content": step_content})
# --- B. 准备 Task 3 特有的匹配提示 ---
# 默认值为空字符串 (适配 Task 1)
match_status_hint_text = ""
if task_type == "camera_view_ordering":
is_current_view_match = False
matched_label = None
# Look Back 逻辑:检查“当前看到的图”(即上一步的结果)是否是 Key Frame
if i == 0:
is_current_view_match = False
else:
prev_step = chain[i-1]
if prev_step.get('is_key_frame'):
is_current_view_match = True
# 当前看到的图,对应的是要 order 的哪个 view
matched_label = prev_step['matched_display_label']
# 提示模型,注意当前 view 与 matched_label 的图是 match 的
if is_current_view_match:
match_status_hint_text = f"The view you are looking at RIGHT NOW corresponds to **Candidate Image {matched_label}**. Explicitly identify this match."
else:
match_status_hint_text = "The view you are looking at RIGHT NOW is an intermediate transition. It does NOT match any candidate image."
current_step_direction = step_action.get('direction', '')
if current_step_direction == 'anticlockwise':
# 逆时针 = 摄像机向右移 = 画面向左移
physics_rule_text = "Camera moves RIGHT -> View shifts LEFT. (Features on the Right move to Center; Center moves to Left)."
elif current_step_direction == 'clockwise':
# 顺时针 = 摄像机向左移 = 画面向右移
physics_rule_text = "Camera moves LEFT -> View shifts RIGHT. (Features on the Left move to Center; Center moves to Right)."
else:
physics_rule_text = "Maintain current view."
# --- C. 构造参数字典 ---
step_format_kwargs = {
# Task 1 & 3 通用
"total_degrees": meta.get('angle_degrees', 0),
"direction": meta.get('direction', step_action.get('direction', '')),
"current_angle_so_far": current_angle_so_far,
"step_degrees": step_action.get('degrees', 0),
"step_direction": step_action.get('direction', ''),
"physics_rule": physics_rule_text,
"next_image_content": self._get_content(f"<image_start>[{next_key}]<image_end>", img_map),
# Task 3 特有 (Task 1 会自动忽略这个多余参数)
"match_status_hint": match_status_hint_text,
# 辅助
"next_image_key": img_map.get(next_key, next_key)
}
# --- D. 渲染 Prompt ---
step_prompt_template = prompts["step"]
cheat_prompt = step_prompt_template.format(**step_format_kwargs)
current_turn_messages.append({"role": "user", "content": self._get_content(cheat_prompt, img_map)})
# --- E. 调用模型 ---
reasoning_text = self._call_llm_with_retry(current_turn_messages)
cot_list.append(reasoning_text)
think_block = f"<think>{reasoning_text}</think>"
image_block = f"<image_start>[{next_key}]<image_end>"
full_trace_parts.append(think_block)
full_trace_parts.append(image_block)
# --- F. 更新 History ---
# TODO: assistant 历史中不放 image ,其已经保留在 user query 中
assistant_content = f"{think_block}"
history_messages.append({"role": "assistant", "content": assistant_content})
current_angle_so_far = step_action.get('total_angle_so_far', 0.0)
# 3. 最终验证
current_turn_messages = copy.deepcopy(history_messages)
final_key = chain[-1]['result_image_key']
final_format_kwargs = {
# common args
"correct_label": meta.get('correct_label', ""),
# task1 args
"total_degrees": meta.get('angle_degrees', 0),
"final_img_abs_path": img_map.get(final_key, ""),
"direction": meta.get("direction"),
# task3 args
"correct_sequence_str": meta.get('correct_sequence_str', "")
}
final_prompt = prompts["final"].format(**final_format_kwargs)
current_turn_messages.append({"role": "user", "content": self._get_content(final_prompt, img_map)})
final_response_text = self._call_llm_with_retry(current_turn_messages)
cot_list.append(final_response_text)
full_trace_parts.append(f"<think>{final_response_text}</think>")
entry['reasoning'] = "".join(full_trace_parts)
entry['oracle_meta']['cot'] = cot_list
return entry
def process_batch(self, entries):
results = []
logging.basicConfig(level=logging.INFO)
for entry in tqdm(entries, desc="Generating CoT"):
try:
res = self.process_single_entry(entry)
results.append(res)
except Exception as e:
logger.error(f"Failed to process entry {entry.get('id', 'unknown')} (Error: {e})")
print("="*30 + " ERROR TRACEBACK " + "="*30)
traceback.print_exc()
print("="*77)
return results
|