feat: gen_task1 v3

src/action_state/gen_task1.py

@@ -4,9 +4,9 @@
 注：{angle}就是 xx degrees, {direction} 就是 clockwise / anticlockwise
 
 1.
-The object in the image <image_start>[image_1]<image_end> remains **static**. Imagine a camera rotating around this object. The direction of rotation is defined from a **top-down bird's-eye view**.
+The {object} in the image <image_start>[image_1]<image_end> remains **static**. Imagine a camera rotating around this {object}. The direction of rotation is defined from a **top-down bird's-eye view**.
 
-Please identify the view of the object after the camera rotates {angle} {direction} based on this top-down perspective, and select the correct answer.
+Please identify the view of the {object} after the camera rotates {angle} {direction} based on this top-down perspective, and select the correct answer.
 
 A. <image_start>[image_A]<image_end>
 B. <image_start>[image_B]<image_end>
@@ -15,11 +15,11 @@ D. <image_start>[image_D]<image_end>
 
 
 2.
-Given the initial view of a static object: <image_start>[image_1]<image_end>.
+Given the initial view of a **static** {object}: <image_start>[image_1]<image_end>.
 
-Imagine looking at the setup from a bird's-eye view (from directly above) to determine the direction. Now, move the camera {angle} {direction} around the object.
+Imagine looking at the setup from a bird's-eye view (from directly above) to determine the direction. Now, move the camera {angle} {direction} around the {object}.
 
-Which of the following images shows what the object looks like from this new position?
+Which of the following images shows what the {object} looks like from this new position?
 
 A. <image_start>[image_A]<image_end>
 B. <image_start>[image_B]<image_end>
@@ -36,6 +36,7 @@ import argparse
 import numpy as np
 from scipy.spatial.transform import Rotation as SciRotation
 from tqdm import tqdm
+import math
 
 # --- 1. Utils 函数 ---
 
@@ -43,29 +44,27 @@ def get_relative_horizontal_rotation_matrix(R1, R2):
     """
     输入两个旋转矩阵，输出相对 Yaw 角度 (Degree)。
     """
-    # R_rel = R2 @ R1.T
     R_rel = R2 @ R1.T
-    
-    # 转换为欧拉角提取水平分量 (Y轴)
     r = SciRotation.from_matrix(R_rel)
     euler_angles = r.as_euler('xyz', degrees=True) 
     horizontal_rotation = euler_angles[1]
-    
     return horizontal_rotation
 
 # --- 2. 核心生成类 ---
 
 class CO3DQuestionGenerator:
-    def __init__(self, root_path, category):
+    def __init__(self, root_path, category, image_prefix="data/"):
         self.root_path = root_path
-        # 移除末尾斜杠
         if self.root_path.endswith('/'):
             self.root_path = self.root_path[:-1]
             
         self.category = category
+        self.image_prefix = image_prefix
+        
+        if self.image_prefix and not self.image_prefix.endswith('/'):
+            self.image_prefix += '/'
         
         # 预加载 annotations
-        # 假设结构是 root/data/original/category/frame_annotations.json
         json_path = os.path.join(root_path, 'data', 'original', category, 'frame_annotations.json')
         
         if not os.path.exists(json_path):
@@ -76,9 +75,8 @@ class CO3DQuestionGenerator:
         with open(json_path, 'r') as f:
             self.annotations = json.load(f)
         
-        # 整理数据结构：Sequence -> Frame Dict
+        # 整理数据结构
         self.seq_data = {}
-        
         for item in self.annotations:
             seq_name = item['sequence_name']
             frame_idx = item['frame_number']
@@ -92,12 +90,6 @@ class CO3DQuestionGenerator:
             }
 
     def format_path(self, raw_path):
-        """
-        处理图片路径：
-        1. 移除 root_path 前缀（如果是绝对路径）。
-        2. 确保路径以 'data/' 开头。
-        """
-        # 1. 处理绝对路径转相对
         if raw_path.startswith(self.root_path):
             rel_path = raw_path[len(self.root_path):]
             if rel_path.startswith('/'):
@@ -105,21 +97,12 @@ class CO3DQuestionGenerator:
         else:
             rel_path = raw_path
 
-        # 2. 确保以 data/ 开头
-        # 如果原始 path 是 "motorcycle/..."，拼接成 "data/motorcycle/..."
-        # 如果原始 path 已经是 "data/original/..."，则保持不变
-        if not rel_path.startswith('data/'):
-            # 这里假设如果不是 data 开头，就需要补全 data/ 前缀
-            # 注意：这里根据你的需求，如果原始是 motorcycle/xxx，变成 data/motorcycle/xxx
-            # 如果原始是 original/motorcycle/xxx (比较少见)，也会变成 data/original/...
-            rel_path = os.path.join('data', rel_path)
+        if self.image_prefix and rel_path.startswith(self.image_prefix):
+            return rel_path
             
-        return rel_path
+        return os.path.join(self.image_prefix, rel_path)
 
     def format_angle_direction(self, angle):
-        """
-        将带符号的角度转换为绝对值角度和方向字符串。
-        """
         angle = (angle + 180) % 360 - 180
         direction = "clockwise" if angle > 0 else "anticlockwise"
         abs_angle = abs(angle)
@@ -127,35 +110,44 @@ class CO3DQuestionGenerator:
 
     def verify_angles(self, start_R, target_R, distractor_Rs, min_angle, max_angle, min_interval):
         """
-        验证生成的题目是否符合约束条件。
+        验证所有选定帧之间的角度互斥性。
         """
-        # 1. 计算目标角度
+        # 1. 计算 Target 相对于 Start 的角度
         target_yaw = get_relative_horizontal_rotation_matrix(start_R, target_R)
-        abs_target_yaw = abs(target_yaw)
         
-        # 检查目标角度范围
-        if not (min_angle <= abs_target_yaw <= max_angle):
+        # 检查 Target 是否在合法范围内 (绝对值)
+        if not (min_angle <= abs(target_yaw) <= max_angle):
             return False, target_yaw, []
 
-        # 2. 检查干扰项
+        # 2. 计算所有 Distractors 相对于 Start 的角度
         distractor_yaws = []
         for d_R in distractor_Rs:
             d_yaw = get_relative_horizontal_rotation_matrix(start_R, d_R)
             distractor_yaws.append(d_yaw)
-            
-            # 计算干扰项和目标项的角度差 (考虑圆周)
-            diff = abs(d_yaw - target_yaw)
-            diff = min(diff, 360 - diff)
-            
-            if diff < min_interval:
-                return False, target_yaw, []
+
+        # 3. 全局互斥检查 (Global Separation Check)
+        # 我们需要确保以下集合中任意两个角度之差都 >= min_interval:
+        # [Start(0度), Target, Distractor1, Distractor2, Distractor3]
+        
+        all_angles = [0.0, target_yaw] + distractor_yaws
+        
+        # 双重循环检查任意两两组合
+        for i in range(len(all_angles)):
+            for j in range(i + 1, len(all_angles)):
+                angle_a = all_angles[i]
+                angle_b = all_angles[j]
                 
+                # 计算圆周上的最小差值
+                diff = abs(angle_a - angle_b)
+                diff = min(diff, 360 - diff)
+                
+                if diff < min_interval:
+                    # 只要有一对冲突，该组合即为无效
+                    return False, target_yaw, []
+        
         return True, target_yaw, distractor_yaws
 
     def generate_samples_for_sequence(self, sequence_name, num_samples, min_angle, max_angle, min_interval):
-        """
-        为一个 Sequence 生成指定数量的样本。
-        """
         if sequence_name not in self.seq_data:
             return []
 
@@ -167,28 +159,25 @@ class CO3DQuestionGenerator:
 
         generated_samples = []
         attempts = 0
-        max_attempts = num_samples * 20 # 避免死循环
+        # 增加尝试次数，因为全互斥条件比较严格
+        max_attempts = num_samples * 5000
         
         while len(generated_samples) < num_samples and attempts < max_attempts:
             attempts += 1
             
-            # 1. 随机选择起始帧
             start_idx = random.choice(frame_indices)
             start_R = frames_dict[start_idx]['R']
             
-            # 2. 随机选择目标帧
             possible_targets = [f for f in frame_indices if f != start_idx]
             target_idx = random.choice(possible_targets)
             target_R = frames_dict[target_idx]['R']
             
-            # 3. 随机选择干扰项
             remaining = [f for f in frame_indices if f != start_idx and f != target_idx]
             if len(remaining) < 3: continue
             
             distractor_indices = random.sample(remaining, 3)
             distractor_Rs = [frames_dict[d]['R'] for d in distractor_indices]
             
-            # 4. 验证
             is_valid, target_yaw, distractor_yaws = self.verify_angles(
                 start_R, target_R, distractor_Rs, 
                 min_angle, max_angle, min_interval
@@ -208,18 +197,14 @@ class CO3DQuestionGenerator:
         
         angle_deg, direction_str = self.format_angle_direction(target_yaw)
         
-        # 获取并格式化路径
         img_start = self.format_path(frames_dict[start_idx]['path'])
         
-        # 准备选项数据：(路径, 角度, 是否正确)
-        # 正确选项
         options_data = [{
             "path": self.format_path(frames_dict[target_idx]['path']),
             "angle": target_yaw,
             "is_correct": True
         }]
         
-        # 干扰选项
         for idx, yaw in zip(distractor_indices, distractor_yaws):
             options_data.append({
                 "path": self.format_path(frames_dict[idx]['path']),
@@ -227,10 +212,8 @@ class CO3DQuestionGenerator:
                 "is_correct": False
             })
         
-        # 打乱选项
         random.shuffle(options_data)
         
-        # 构建选项字典和元数据
         images_dict = {"image_1": img_start}
         option_labels = ['A', 'B', 'C', 'D']
         option_angles_meta = {}
@@ -244,24 +227,27 @@ class CO3DQuestionGenerator:
             if opt["is_correct"]:
                 correct_answer_label = label
 
-        # 构建 Question
+        # --- 修改点：将 object 替换为 category ---
+        # 处理 category 名称，去掉下划线，使其更自然 (例如 tv_monitor -> tv monitor)
+        cat_name = self.category.replace('_', ' ')
+        
         template_id = random.choice([1, 2])
         question = ""
         if template_id == 1:
-            question = f"""The object in the image <image_start>[image_1]<image_end> remains **static**. Imagine a camera rotating around this object. The direction of rotation is defined from a **top-down bird's-eye view**.
+            question = f"""The {cat_name} in the image <image_start>[image_1]<image_end> remains **static**. Imagine a camera rotating around this {cat_name}. The direction of rotation is defined from a **top-down bird's-eye view**.
 
-Please identify the view of the object after the camera rotates {angle_deg} degrees {direction_str} based on this top-down perspective, and select the correct answer.
+Please identify the view of the {cat_name} after the camera rotates {angle_deg} degrees {direction_str} based on this top-down perspective, and select the correct answer.
 
 A. <image_start>[image_A]<image_end>
 B. <image_start>[image_B]<image_end>
 C. <image_start>[image_C]<image_end>
 D. <image_start>[image_D]<image_end>"""
         else:
-            question = f"""Given the initial view of a static object: <image_start>[image_1]<image_end>.
+            question = f"""Given the initial view of a **static** {cat_name}: <image_start>[image_1]<image_end>.
 
-Imagine looking at the setup from a bird's-eye view (from directly above) to determine the direction. Now, move the camera {angle_deg} degrees {direction_str} around the object.
+Imagine looking at the setup from a bird's-eye view (from directly above) to determine the direction. Now, move the camera {angle_deg} degrees {direction_str} around the {cat_name}.
 
-Which of the following images shows what the object looks like from this new position?
+Which of the following images shows what the {cat_name} looks like from this new position?
 
 A. <image_start>[image_A]<image_end>
 B. <image_start>[image_B]<image_end>
@@ -280,69 +266,96 @@ D. <image_start>[image_D]<image_end>"""
                 "angle_degrees": angle_deg,
                 "direction": direction_str,
                 "raw_yaw": target_yaw,
-                "option_angles": option_angles_meta # 新增：记录每个选项的角度
+                "option_angles": option_angles_meta
             },
             "gt_answer": correct_answer_label
         }
 
-# --- 3. 主程序 ---
+# --- 3. 辅助函数：保存与切分 ---
+
+def save_split_data(data_list, output_dir, split_name, max_items):
+    if not data_list:
+        return
+
+    split_dir = os.path.join(output_dir, split_name)
+    os.makedirs(split_dir, exist_ok=True)
+    
+    num_files = math.ceil(len(data_list) / max_items)
+    
+    print(f"Saving {len(data_list)} items to {split_name} (split into {num_files} files)...")
+    
+    for i in range(num_files):
+        start_idx = i * max_items
+        end_idx = min((i + 1) * max_items, len(data_list))
+        chunk = data_list[start_idx:end_idx]
+        
+        file_name = f"{split_name}_{i+1}.jsonl"
+        file_path = os.path.join(split_dir, file_name)
+        
+        with open(file_path, 'w') as f:
+            for item in chunk:
+                f.write(json.dumps(item) + '\n')
+
+# --- 4. 主程序 ---
 
 def main():
-    parser = argparse.ArgumentParser(description="Generate CO3D Rotation Questions")
+    parser = argparse.ArgumentParser(description="Generate CO3D Rotation Questions (JSONL with Splits)")
     
-    # 路径参数
     parser.add_argument("--root_path", type=str, required=True, help="Dataset root path")
-    parser.add_argument("--output_file", type=str, default="co3d_questions.json", help="Output JSON file path")
+    parser.add_argument("--output_dir", type=str, default="output_dataset", help="Output directory")
+    parser.add_argument("--image_prefix", type=str, default="data/", help="Prefix for image paths")
     
-    # 筛选参数
-    parser.add_argument("--category", type=str, default=None, help="Specific category (e.g., motorcycle). If None, iterate all.")
+    parser.add_argument("--category", type=str, default=None, help="Specific category. If None, iterate all.")
     parser.add_argument("--num_samples", type=int, default=1, help="Number of samples per sequence")
     
-    # 角度约束参数
-    parser.add_argument("--min_angle", type=float, default=40.0, help="Minimum rotation angle for correct answer")
-    parser.add_argument("--max_angle", type=float, default=140.0, help="Maximum rotation angle for correct answer")
-    parser.add_argument("--min_interval", type=float, default=25.0, help="Minimum angle difference between correct answer and distractors")
+    parser.add_argument("--min_angle", type=float, default=40.0)
+    parser.add_argument("--max_angle", type=float, default=140.0)
+    parser.add_argument("--min_interval", type=float, default=25.0)
     
-    # 随机种子
-    parser.add_argument("--seed", type=int, default=42, help="Random seed")
+    parser.add_argument("--train_ratio", type=float, default=0.8)
+    parser.add_argument("--val_ratio", type=float, default=0.1)
+    parser.add_argument("--test_ratio", type=float, default=0.1)
+    parser.add_argument("--max_items", type=int, default=10000)
+    
+    parser.add_argument("--seed", type=int, default=42)
 
     args = parser.parse_args()
 
-    # 1. 设置随机种子
+    total_ratio = args.train_ratio + args.val_ratio + args.test_ratio
+    if abs(total_ratio - 1.0) > 1e-5:
+        print(f"Warning: Ratios sum to {total_ratio}, not 1.0. They will be normalized.")
+        args.train_ratio /= total_ratio
+        args.val_ratio /= total_ratio
+        args.test_ratio /= total_ratio
+
     random.seed(args.seed)
     np.random.seed(args.seed)
     
-    # 2. 确定要处理的 Categories
     categories = []
     if args.category:
         categories = [args.category]
     else:
-        # 自动遍历 root/data/original 下的所有文件夹
         data_dir = os.path.join(args.root_path, 'data', 'original')
         if os.path.exists(data_dir):
             categories = [d for d in os.listdir(data_dir) if os.path.isdir(os.path.join(data_dir, d))]
-            categories.sort() # 排序保证顺序一致
-            print(f"Found {len(categories)} categories: {categories}")
+            categories.sort()
+            print(f"Found {len(categories)} categories.")
         else:
             print(f"Error: Data directory not found: {data_dir}")
             return
 
     all_results = []
     
-    # 3. 遍历 Category
     for cat in categories:
-        print(f"\nProcessing category: {cat}")
-        generator = CO3DQuestionGenerator(args.root_path, cat)
+        generator = CO3DQuestionGenerator(args.root_path, cat, args.image_prefix)
         
         if not hasattr(generator, 'seq_data') or not generator.seq_data:
             continue
             
         sequences = list(generator.seq_data.keys())
-        sequences.sort() # 排序保证顺序一致
+        sequences.sort()
         
-        # 4. 遍历 Sequence
-        # 使用 tqdm 显示进度
-        for seq in tqdm(sequences, desc=f"Sequences in {cat}"):
+        for seq in tqdm(sequences, desc=f"Processing {cat}", leave=False):
             samples = generator.generate_samples_for_sequence(
                 seq, 
                 args.num_samples, 
@@ -352,12 +365,25 @@ def main():
             )
             all_results.extend(samples)
 
-    # 5. 保存结果
     print(f"\nTotal samples generated: {len(all_results)}")
-    print(f"Saving to {args.output_file}...")
-    with open(args.output_file, 'w') as f:
-        json.dump(all_results, f, indent=4)
-    print("Done.")
+    
+    random.shuffle(all_results)
+    
+    n_total = len(all_results)
+    n_train = int(n_total * args.train_ratio)
+    n_val = int(n_total * args.val_ratio)
+    
+    train_data = all_results[:n_train]
+    val_data = all_results[n_train : n_train + n_val]
+    test_data = all_results[n_train + n_val :]
+    
+    print(f"Split counts: Train={len(train_data)}, Val={len(val_data)}, Test={len(test_data)}")
+    
+    save_split_data(train_data, args.output_dir, "train", args.max_items)
+    save_split_data(val_data, args.output_dir, "val", args.max_items)
+    save_split_data(test_data, args.output_dir, "test", args.max_items)
+    
+    print(f"All done. Output saved to {args.output_dir}")
 
 if __name__ == "__main__":
     main()