File size: 4,687 Bytes
b6f2d77 | 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 | """VLM2Vec-V2.0 编码封装（Q3 决策：帧级 image emb + clip 级 video emb 两级）。
复用 MAGIC-video 的 VLM2VecV2EmbeddingModel wrapper（权重已在本地 HF 缓存）：
- encode_frames: 关键帧 JPEG 批量图像编码（wrapper.encode_image 内部真批量）
- encode_clip: 一个 clip 的关键帧序列当作视频编码 → 1 个 clip 级向量
（复用已保存的关键帧 JPEG，不重复解码视频；帧数补齐为偶数
以满足 Qwen2-VL temporal patch=2）
- encode_texts: 检索 query 文本编码（推理阶段 memory_store 调用）
"""
import logging
import sys
import numpy as np
import torch
from config import MAGIC_SRC, setup_hf_env
logger = logging.getLogger(__name__)
setup_hf_env()
if MAGIC_SRC not in sys.path:
sys.path.insert(0, MAGIC_SRC)
class MemoryEncoder:
def __init__(self, model_name: str = "VLM2Vec/VLM2Vec-V2.0",
device: str = "cuda", clip_max_pixels: int = 360 * 420):
from worldmm.embedding.vlm2vecv2 import VLM2VecV2EmbeddingModel
self.model = VLM2VecV2EmbeddingModel(model_name=model_name, device=device)
self.device = device
self.clip_max_pixels = clip_max_pixels
# 视频/图像的 instruction 常量（与 MMEB 协议一致）
from worldmm.embedding.VLM2Vec.src.model.processor import (
QWEN2_VL, VLM_IMAGE_TOKENS, VLM_VIDEO_TOKENS,
)
self._img_token = VLM_IMAGE_TOKENS[QWEN2_VL]
self._vid_token = VLM_VIDEO_TOKENS[QWEN2_VL]
# ---------------------------------------------------------- 帧级编码
def encode_frames(self, image_paths: list, batch_size: int = 16) -> np.ndarray:
"""关键帧 JPEG → (n, D) 归一化向量。"""
embs = []
for i in range(0, len(image_paths), batch_size):
batch = image_paths[i:i + batch_size]
out = self.model.encode_image(
batch,
query_text=f"{self._img_token} Represent the given video frame.",
)
embs.append(out)
return np.concatenate(embs, axis=0) if embs else np.zeros((0, 1536), np.float32)
# --------------------------------------------------------- clip 级编码
def encode_clip(self, frames: list, clip_seconds: float) -> np.ndarray:
"""一个 clip 的帧序列 → (1, D) 视频级向量。
frames: 图片路径列表 或 PIL.Image 列表 或 JPEG bytes 列表
（v2 用 clip 内均匀 8 帧的内存 JPEG，qwen_vl_utils.fetch_image 原生支持 PIL）。
"""
import io
from PIL import Image
from worldmm.embedding.VLM2Vec.src.model.vlm_backbone.qwen2_vl.qwen_vl_utils import (
process_vision_info,
)
items = []
for f in frames:
if isinstance(f, bytes):
items.append(Image.open(io.BytesIO(f)).convert("RGB"))
elif isinstance(f, str):
items.append(f"file://{f}")
else:
items.append(f) # PIL.Image
if len(items) == 1:
items = items * 2 # temporal patch=2 需要至少 2 帧
if len(items) % 2 == 1:
items.append(items[-1])
messages = [{
"role": "user",
"content": [
{
"type": "video",
"video": items,
"max_pixels": self.clip_max_pixels,
"fps": max(len(items) / max(clip_seconds, 1.0), 0.1),
},
{"type": "text", "text": "Describe this video."},
],
}]
_, video_inputs = process_vision_info(messages)
inputs = self.model.processor(
text=f"{self._vid_token} Represent the given video.",
videos=video_inputs,
return_tensors="pt",
)
inputs = {k: v.to(self.device) for k, v in inputs.items()}
if "pixel_values_videos" in inputs:
inputs["pixel_values_videos"] = inputs["pixel_values_videos"].unsqueeze(0)
if "video_grid_thw" in inputs:
inputs["video_grid_thw"] = inputs["video_grid_thw"].unsqueeze(0)
with torch.no_grad():
with torch.autocast(device_type=self.device.split(":")[0], dtype=torch.bfloat16):
out = self.model.model(qry=inputs)["qry_reps"]
return out.float().cpu().numpy()
# ------------------------------------------------------------ 文本编码
def encode_texts(self, texts: list) -> np.ndarray:
"""检索 query → (n, D)。推理阶段 search_vis 用。"""
return self.model.encode_text(texts)
|