Update some files

src/__init__.py

@@ -0,0 +1,5 @@
+# src/__init__.py
+# 空文件，用于将 src 文件夹标记为 Python 包。
+# 这样可以通过 from src import model_loader 等方式进行模块导入。
+
+__version__ = "0.1.0"

src/model_loader.py

@@ -0,0 +1,62 @@
+# src/model_loader.py
+import os
+import math
+import torch
+from transformers import AutoModel, AutoTokenizer, AutoConfig
+from huggingface_hub import snapshot_download
+
+MODEL_NAME = "OpenGVLab/InternVL3-14B"
+CACHE_DIR = "/data/internvl3_model"
+
+# === 自动分配模型层到多张 GPU（InternVL3 建议方式） ===
+def split_model(model_path):
+    device_map = {}
+    world_size = torch.cuda.device_count()
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    num_layers = config.llm_config.num_hidden_layers
+
+    num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
+    num_layers_per_gpu = [num_layers_per_gpu] * world_size
+    num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
+
+    layer_cnt = 0
+    for i, num_layer in enumerate(num_layers_per_gpu):
+        for _ in range(num_layer):
+            device_map[f'language_model.model.layers.{layer_cnt}'] = i
+            layer_cnt += 1
+
+    # 固定组件放在 GPU 0
+    for key in [
+        'vision_model', 'mlp1',
+        'language_model.model.tok_embeddings',
+        'language_model.model.embed_tokens',
+        'language_model.output',
+        'language_model.model.norm',
+        'language_model.model.rotary_emb',
+        'language_model.lm_head',
+        f'language_model.model.layers.{num_layers - 1}'
+    ]:
+        device_map[key] = 0
+
+    return device_map
+
+# === 模型加载函数 ===
+def load_model():
+    if not os.path.exists(CACHE_DIR):
+        print("⏬ First run: downloading model to persistent storage...")
+        snapshot_download(repo_id=MODEL_NAME, local_dir=CACHE_DIR)
+    else:
+        print("✅ Loaded model from persistent cache.")
+
+    device_map = split_model(CACHE_DIR)
+    tokenizer = AutoTokenizer.from_pretrained(CACHE_DIR, trust_remote_code=True)
+    model = AutoModel.from_pretrained(
+        CACHE_DIR,
+        torch_dtype=torch.bfloat16,
+        low_cpu_mem_usage=True,
+        use_flash_attn=False,  # 或者True，如果确认安装好FlashAttention
+        trust_remote_code=True,
+        device_map=device_map
+    ).eval()
+
+    return tokenizer, model

src/prompt.py

@@ -0,0 +1,7 @@
+# src/prompt.py
+
+def build_video_prompt(num_frames: int) -> str:
+    """构建适用于 InternVL3 的单轮 AR 视频评估提示语。"""
+    frame_descriptors = ''.join([f"Frame{i+1}: <image>\n" for i in range(num_frames)])
+    final_prompt = frame_descriptors + "Evaluate the quality of AR occlusion and rendering in the uploaded video."
+    return final_prompt

src/video_utils.py

@@ -0,0 +1,94 @@
+# src/video_utils.py
+import numpy as np
+import torch
+from PIL import Image
+from decord import VideoReader, cpu
+import torchvision.transforms as T
+from torchvision.transforms.functional import InterpolationMode
+from src.prompt import build_video_prompt
+
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+
+# === 构建标准图像预处理 transform ===
+def build_transform(input_size=448):
+    return T.Compose([
+        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+        T.ToTensor(),
+        T.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
+    ])
+
+# === InternVL3 视频帧采样策略 ===
+def get_frame_indices(num_frames, total_frames):
+    indices = np.linspace(0, total_frames - 1, num_frames, dtype=int)
+    return indices
+
+# === 从视频中提取图像帧并预处理成 patch tensor ===
+def process_video_for_internvl3(video_path, num_segments=8, max_patch_per_frame=1, input_size=448):
+    vr = VideoReader(video_path, ctx=cpu(0))
+    total_frames = len(vr)
+    frame_indices = get_frame_indices(num_segments, total_frames)
+    transform = build_transform(input_size)
+
+    pixel_values_list, num_patches_list = [], []
+    for idx in frame_indices:
+        img = Image.fromarray(vr[idx].asnumpy()).convert("RGB")
+        patches = dynamic_preprocess(img, image_size=input_size, max_num=max_patch_per_frame)
+        patch_tensors = [transform(tile) for tile in patches]
+        patch_tensor = torch.stack(patch_tensors)
+        pixel_values_list.append(patch_tensor)
+        num_patches_list.append(patch_tensor.shape[0])
+
+    pixel_values = torch.cat(pixel_values_list, dim=0).to(torch.bfloat16).cuda()
+    prompt = build_video_prompt(len(num_patches_list))
+
+    return pixel_values, num_patches_list, prompt
+
+# === 图像切片为 patch 区块 ===
+def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=True):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # 构造备选分块比率
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1)
+        if i * j <= max_num and i * j >= min_num
+    )
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    best_ratio = find_closest_aspect_ratio(aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+    target_width = image_size * best_ratio[0]
+    target_height = image_size * best_ratio[1]
+    blocks = best_ratio[0] * best_ratio[1]
+
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+
+    return processed_images
+
+# === 找出最接近原图比例的块切方案 ===
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect = ratio[0] / ratio[1]
+        diff = abs(aspect_ratio - target_aspect)
+        if diff < best_ratio_diff or (diff == best_ratio_diff and area > 0.5 * image_size**2 * ratio[0] * ratio[1]):
+            best_ratio_diff = diff
+            best_ratio = ratio
+    return best_ratio