JunX

# Media - 多媒体 - Audio/Image/Video/...

# ML - 机器学习/深度学习/...

# 大模型微调 - Adapter - Prompting - Prefix-Tuning - P-Tuing V2 - LoRA - Dreambooth ## Reference - [LoRA: Low-Rank Adaptation of Large Language Models 简读](https://zhuanlan.zhihu.com/p/514033873) - [Adapter/Prompt文献整理](https://zhuanlan.zhihu.com/p/554959952) - [PEFT...

[Bug]: openvino export INT8_PTQ for winclip

1

### Describe the bug https://github.com/openvinotoolkit/anomalib/blob/3fe55842cf7bf1024ed43022d5859600052a4ed0/src/anomalib/models/components/base/export_mixin.py#L251 **as the [nncf api documents](https://openvinotoolkit.github.io/nncf/autoapi/nncf/#nncf.quantize):** - Is the `model_type` should be set **Transformer** for WinCLIP : `model_type=nncf.parameters.ModelType.TRANSFORMER` ### Dataset MVTec ### Model Other (please specify...

### Describe the bug **`compress_to_fp16` is False, but the comment say that `fp16 compression is enabled by default `**, it is confused. https://github.com/openvinotoolkit/anomalib/blob/e5b91d6a32c469ec0b9c4c161272b68114ba6fa3/src/anomalib/models/components/base/export_mixin.py#L155-L164 https://github.com/openvinotoolkit/anomalib/blob/e5b91d6a32c469ec0b9c4c161272b68114ba6fa3/src/anomalib/models/components/base/export_mixin.py#L253-L255 ### Dataset MVTec ### Model Other...

fix winclip openvino inference output without mask

4

## 📝 Description WinClip OpenVINO Inference output is only the classification score. - 🛠️ Fixes #1641 ## ✨ Changes Select what type of change your PR is: - [x] 🐞...

# FiLo - 增加了 LLMs 来生成异常描述信息 - 增加了 Grounding DINO 来生成 位置信息 - CLIP Backbone: `ViT-L-14-336px` - GroundingDINO - `Swin-T` ## Arch - FG-Des - Adaptively Learned Text Templates -...

# 知识蒸馏 - Knowledge Distillation 知识蒸馏 - Teacher/Student Model - Teacher Model 的输出 作为 Soft Target Training Student Model - Student 学习 Teacher 泛化能力 - 用途 - 模型压缩(小模型学习大模型的泛化能力) - [History](https://github.com/junxnone/aiwiki/issues/336)...

MO KD History

1

# KD History ## History **Pattern** ``` - 2020-05 **[Name]** paper_full_name [[Paper](paperlink)] [[Code](codelink)] [[Note](notelink)] #tags ``` - 2006 Model Compression [[Paper](http://www.cs.cornell.edu/~caruana/compression.kdd06.pdf)] [[Code](codelink)] [[Note](notelink)] - 2015-03 Distilling the Knowledge in a...

# Distilling the Knowledge in a Neural Network - `Soft Target` 比 `Hard Target` 包含更多的信息 - 不同类之间的关系(相似性) ## Motivation - 背景/存在的问题 - 粗暴地使用one-hot编码丢失了类间和类内关于相似性的额外信息 - 做了什么尝试/分别有什么效果 - 使用带温度的softmax函数 - 应用领域 ##...