致谢

2023年7月7日
修改自 SmilingWolf/wd-v1-4-moat-tagger-v2
sha256==8452cddf280b952281b6e102411c50e981cb2908

我使用这个模型进行图像特征提取，用于图像聚类 https://github.com/WSH032/image-deduplicate-cluster-webui

我做了什么？

我将模型的输出调整到了最后四层。
并将 keras 转换为 onnx。

环境

工具在此 https://github.com/WSH032/wd-v1-4-tagger-feature-extractor-tutorials

感谢 Colab

onnx == 1.14.0
tf2onnx == 1.14.0
tensorflow == 2.12.0

详情

模型详情

# modified from \"SmilingWolf/wd-v1-4-moat-tagger-v2\"
# 8452cddf280b952281b6e102411c50e981cb2908

# 输入 ['input_1']
# 输出 ['predictions_sigmoid', 'predictions_dense', 'predictions_norm', 'predictions_globalavgpooling']  # 最左边是最外层

[[input]]
name = \"input_1\"  # 原始模型就有
shape = [ \"None\", 448, 448, 3,]
dtype = \"float32\"

[[output]]
name = \"predictions_sigmoid\"  # 原始模型就有
shape = [ \"None\", 9083,]
dtype = \"float32\"

[[output]]
name = \"predictions_dense\"
shape = [ \"None\", 9083,]
dtype = \"float32\"

[[output]]
name = \"predictions_norm\"
shape = [ \"None\", 1024,]
dtype = \"float32\"

[[output]]
name = \"predictions_globalavgpooling\"
shape = [ \"None\", 1024,]
dtype = \"float32\"

关于 wd14_tags.toml 的详情

它修改自 wd-v1-4-moat-tagger-v2/selected_tags.csv

[rating] 对应 selected_tags.csv 中的 category == 9
[general] 对应 selected_tags.csv 中的 category == 0
[character] 对应 selected_tags.csv 中的 category == 4

关于 candidate_labels_scores_*.npz 的详情

import numpy as np
import pandas as pd
import toml

with open(\"wd14_tags.toml\", \"r\") as f:
    general_tags = toml.load(f)[\"tags\"][1][\"tags\"]  # 0 -> rating, 1 -> general, 2 -> characters

with np.load(\"candidate_labels_scores_safetensors.npz\") as data:
    candidate_labels = data[\"candidate_labels\"]  # 类似于 `wd14_tags.toml` 中的 `[candidate_labels]`
    scores = data[\"scores\"]

df = pd.DataFrame(
    scores,
    index=candidate_labels,
    columns=general_tags,
)

该分数由 sileod/deberta-v3-base-tasksource-nli 推理得出
sha256 == 6a7865dd24917225ec499fad77e91b97baedf7da