| --- |
| license: apache-2.0 |
| base_model: |
| - heathcliff01/Kaloscope |
| --- |
| |
|
|
| onnx conversion of kaloscope model. |
| also extra pytorch ema only version (`best_checkpoint_ema.pth`) of the model which is roughly 4 times smaller and outputs same results |
|
|
| conversion script and implementation script soon:tm: if i dont forget it after testing if the models actually work (onnx model might be broken or my impl is wrong not sure) |
| The ""issue"" with the onnx models is that even when using timm+lsnet impl like the pytorch inference code does im getting correct results but with rather large confidence differences which might be expected but 19%->27.8% seems a bit large |
|
|
| **onnx model name info because im still testing around** (optimization/constant folding) enabled for all of them) |
| - `lsnet_xl_artist-opset13`: first onnx export, dynamo=False, dynamic_axes=`dynamic_axes={ |
| "input": {0: "batch_size", 2: "height", 3: "width"}, |
| "output": {0: "batch_size"}, |
| }`, opset_version=13 |
| - `lsnet_xl_artist-dynamo-opset18_merged.onnx`: dynamo=True, dynamic_axes=None, opset_version=None (torch 2.8.0 used here defaults to 18 when None) |
| - 'merged' in name because torch.onnx.export gave me 2 files - .onnx.data and a small .onnx file - i used onnx_merge.py (todo: add link and upload file) to merge them |
| - `lsnet_xl_artist-dynamo-opset20_merged.onnx`: dynamo=True, dynamic_axes=None, opset_version=None (torch 2.9.0 used here defaults to 20 when None) |
| - `lsnet_xl_artist-dynamo-opset23_merged.onnx`: dynamo=True, dynamic_axes=None, opset_version=23 (torch 2.9.0 used) |
| |
| Note that newer opset versions might not work with older tooling. I haven't tested this. |
| **Using opset 18 is probably the best choice** - I don't see any score difference between the different opset except opset 13 but that one also didn't have dynamo enabled during export and 18 gives more backward compatibility. |
|
|
| --- |
|
|
|
|
| # LSNet 艺术家风格分类模型 Model Card |
|
|
| ## 模型概述 |
|
|
| **模型名称**: |
| Kaloscope Artist Style Classification Model |
| **模型版本**: v1.0 |
| **发布日期**: 2025年10月 |
| **模型类型**: 图像分类 (艺术家风格识别) |
| **架构**: LSNet (See Large, Focus Small) |
|
|
| ## 模型描述 |
|
|
| 本模型基于LSNet架构构建，专门用于识别和分类不同艺术家的绘画风格。LSNet是一个轻量级视觉模型，灵感来源于人类视觉系统的动态异尺度能力，即"看大局，聚焦细节"的特性。 |
|
|
| ### 架构特点 |
| - **设计理念**: 基于人类视觉系统的"See Large, Focus Small"原理 |
| - **模型系列**: 支持LSNet-T、LSNet-S、LSNet-B三种规模 |
| - **参数量**: 约100M参数 |
| - **优化目标**: 在保持高精度的同时实现高效推理 |
|
|
| ## 训练数据 |
|
|
| ### 数据来源 |
| - **数据集**: Danbooru数据集 (截止到2024年10月) |
| - **数据筛选**: 选取图像数量在50张以上的艺术家 |
| - **总分类数**: 31,770个艺术家类别 |
| - **数据采样策略**: |
| - 图像数量超过100张的艺术家：选取ID最靠后的100张图像 |
| - 图像数量50-100张的艺术家：使用全部图像 |
|
|
| ### 数据预处理 |
| - 图像尺寸: 224×224像素 |
| - 数据增强: 标准ImageNet预处理流程 |
| - 验证集划分: 5%的数据用于验证 |
|
|
| ## 训练配置 |
|
|
| ### 硬件环境 |
| - **GPU配置**: 8×H20 GPU |
| - **训练时长**: 80个epoch |
| - **批次大小**: 256 (每GPU) |
|
|
| ### 训练参数 |
| - **优化器**: AdamW |
| - **学习率调度**: Cosine Annealing |
| - **数据并行**: 分布式训练 (8卡) |
| - **模型参数量**: ~100M |
|
|
| ### 性能指标 |
| - **最终准确率**: 84.2% |
| - **验证方式**: Top-1准确率 |
| - **评估数据**: 验证集 (5%的数据) |
|
|
| ## 模型性能 |
|
|
| ### 分类性能 |
| | 指标 | 数值 | |
| |------|------| |
| | Top-1 准确率 | 84.2% | |
| | 总类别数 | 31,770 | |
| | 参数量 | ~100M | |
| | 训练轮数 | 80 epochs | |
|
|
| ### 推理性能 |
| - **输入格式**: RGB图像，224×224像素 |
| - **输出格式**: 31,770维概率分布 |
| - **推理速度**: 高效推理 (具体数值取决于硬件) |
|
|
| ## 使用方法 |
|
|
| ### 环境要求 |
| ```bash |
| pip install torch torchvision timm |
| ``` |
|
|
| ### 基本使用 |
| ```python |
| import torch |
| from timm.models import create_model |
| |
| # 加载模型 |
| model = create_model('lsnet_t_artist', pretrained=True, num_classes=31770) |
| model.eval() |
| |
| # 推理 |
| with torch.no_grad(): |
| output = model(input_tensor) |
| probabilities = torch.softmax(output, dim=1) |
| ``` |
| ### Comfyui内使用 |
| 安装comfyui节点：https://github.com/spawner1145/comfyui-lsnet |
| 下载本仓库模型即可使用 |
| ### 相关资源 |
| - **论文**: [LSNet: See Large, Focus Small](https://arxiv.org/abs/2503.23135) |
| - **代码仓库**: (https://github.com/spawner1145/lsnet-test) |
| - **预训练模型**: 可通过Hugging Face Hub获取 |
|
|
| ### 引用信息 |
| ```bibtex |
| @misc{wang2025lsnetlargefocussmall, |
| title={LSNet: See Large, Focus Small}, |
| author={Ao Wang and Hui Chen and Zijia Lin and Jungong Han and Guiguang Ding}, |
| year={2025}, |
| eprint={2503.23135}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CV}, |
| url={https://arxiv.org/abs/2503.23135}, |
| } |
| ``` |
|
|
| ## 更新日志 |
|
|
| ### v1.0 (2025年10月) |
| - 初始版本发布 |
| - 基于Danbooru数据集训练 |
| - 支持31,770个艺术家类别 |
| - 达到84.2%的分类准确率 |
|
|
| --- |
|
|
| **免责声明**: 本模型仅供研究和教育用途。在商业应用中使用时，请确保遵守相关法律法规和伦理准则。 |
