model.predict 内部数据处理流程

[Zhang-O]

trafficstars

def demo2():
    model_name = "PP-ShiTuV2_rec"
    # model_name = "PP-ShiTuV2_rec_CLIP_vit_base"
    # model_name = "PP-ShiTuV2_rec_CLIP_vit_large"
    model_path = ""
    model = create_model(
        model_name=model_name,
        model_dir="../inference/latest_IR",  # 替换为你的模型路径
        # 不指定的话似乎是默认使用gpu，会占用显存，计算结果和指定gpu时一样，且与指定cpu时不同。 指定cpu时，计算结果和 paddle inference cpu结果
        # 一致. 但是指定gpu时，计算结果和 paddle inference gpu 结果不一致, 原因未知
        # device="gpu:0",
        device="cpu"
    )
    import cv2
    img_file = "./images/wangzai.jpg"
    img = cv2.imread(img_file)[:, :, ::-1]

    # while 1:
    output = model.predict([img_file] , batch_size=1)

    print(1)
    for res in output:
        res.print()
        # res.save_to_json("./fea_ext/output/res.json")
        feature1 = res["feature"]

    output = model.predict(img)
    for res in output:
        res.print()
        # res.save_to_json("./fea_ext/output/res.json")
        feature2 = res["feature"]

    import numpy as np
    xxx = np.sum(feature1 - feature2)
    print(xxx)  # 0.959705

xxx 不为0， 将

img = cv2.imread(img_file) 

这里修改后，xxx=0, predict 内部的逻辑是怎么样的，读取图片给到模型的时候 是什么通道顺序，内部是什么操作？ 我看 inference.yml 的内容是没有将BGR转换成RGB操作。 以下是inference.yml的内容

Global:
  model_name: PP-ShiTuV2_rec
Hpi:
  backend_configs:
    paddle_infer:
      trt_dynamic_shapes: &id001
        x:
        - - 1
          - 3
          - 224
          - 224
        - - 1
          - 3
          - 224
          - 224
        - - 8
          - 3
          - 224
          - 224
    tensorrt:
      dynamic_shapes: *id001
PreProcess:
  transform_ops:
  - ResizeImage:
      backend: cv2
      interpolation: bilinear
      return_numpy: false
      size:
      - 224
      - 224
  - NormalizeImage:
      channel_num: 3
      mean:
      - 0.485
      - 0.456
      - 0.406
      order: hwc
      scale: 0.00392156862745098
      std:
      - 0.229
      - 0.224
      - 0.225
  - ToCHWImage: null
PostProcess:
  NormalizeFeatures: null

[zhangyubo0722]

在输入到模型时时RGB格式

[TingquanGao]

The issue has no response for a long time and will be closed. You can reopen or new another issue if are still confused.

---

From Bot