update

Author: wenerme

.vscode/settings.json

@@ -3,7 +3,7 @@
   "editor.minimap.enabled": true,
   "editor.wordWrapColumn": 120,
   "editor.defaultFormatter": "esbenp.prettier-vscode",
-   "[javascript]": {
+  "[javascript]": {
     "editor.defaultFormatter": "esbenp.prettier-vscode"
   },
   "[markdown]": {
@@ -13,6 +13,12 @@
   "[xml]": {
     "editor.defaultFormatter": "DotJoshJohnson.xml"
   },
+  "[yaml]": {
+    "editor.defaultFormatter": "kennylong.kubernetes-yaml-formatter"
+  },
+  "[helm]": {
+    "editor.defaultFormatter": "kennylong.kubernetes-yaml-formatter"
+  },
   "[solidity]": {
     "editor.defaultFormatter": "JuanBlanco.solidity"
   },

notes/ai/ai-awesome.md

@@ -26,21 +26,30 @@ tags:
     - MIT, TS
     - AI-powered search engine
     - alternative to Perplexity
+  - [zaidmukaddam/scira](https://github.com/zaidmukaddam/scira)
+    - MIT, TS
+    - MiniPerplx -> scira
+    - minimalistic AI-powered search engine
 - Web/SDK/UI
   - [vercel/ai](https://github.com/vercel/ai)
     - Apache-2.0, TS
     - npm:ai
     - Build AI-powered applications with React, Svelte, Vue, and Solid
   - [xenova/transformers.js](https://github.com/xenova/transformers.js)
     - Apache-2.0, JS
-- Client/Desktop
+- Client/Desktop/App
   - [chatbox](./service/chatbox.md)
     - GPLv3, TypeScript, React, Electron
     - 支持 Desktop, iOS, Android, Web
   - [mckaywrigley/chatbot-ui](https://github.com/mckaywrigley/chatbot-ui)
     - MIT
   - [CherryHQ/cherry-studio](https://github.com/CherryHQ/cherry-studio)
     - ~~Apache-2.0~~, TS
+  - [mainframecomputer/fullmoon-ios](https://github.com/mainframecomputer/fullmoon-ios)
+    - MIT, Swift
+  - [languine-ai/languine](https://github.com/languine-ai/languine)
+    - AGPLv3, TS
+    - CLI, 用于 i10n
 - WebUI/Chatbot
   - [vercel/ai-chatbot](https://github.com/vercel/ai-chatbot)
     - MIT, React
@@ -428,3 +437,9 @@ tags:
   - Tiered Discount Pricing
   - Graduated Discount Pricing
 - https://help.aliyun.com/zh/isi/product-overview/billing-10
+
+## Learning
+
+> Reading, Learning, Tutorials, 教程, 阅读资料
+
+- https://github.com/NielsRogge/Transformers-Tutorials

notes/ai/ai-faq.md

@@ -11,7 +11,7 @@ tags:
   - 量化 - Quantization
     - 降低精度
     - 减少内存占用和计算需求
-    - 例如 FP32 -> INT8, INT4, BIT1.5
+    - 例如 FP32,FP16 -> INT8 [-128, 127], INT4 [-8, 7], BIT1.5, FP16, BF16
   - 蒸馏 - Distillation
     - 将大型模型的知识转移到较小的模型中，实现性能接近的同时降低计算成本。
     - 例如 Teacher-Student
@@ -27,6 +27,19 @@ tags:
 - MoE - Mixture of Experts
   - 减少内存和计算需求
 
+| size |   FP32 1 |   int8 |   int4 | perf                                                                         |
+| ---- | -------: | -----: | -----: | ---------------------------------------------------------------------------- |
+| 1B   |   ~4-6GB | ~2-3GB | ~1-2GB | 适用于简单任务，如基础问答、文本分类，但易出现无意义输出，复杂任务表现不佳。 |
+| 3B   | ~12-18GB | ~6-9GB | ~3-4GB | 能处理中等复杂度任务，如简单对话、文本摘要，性能适中，但仍有限制。           |
+| 7B   |    ~28GB |  ~14GB |   ~7GB | 在大多数NLP任务上表现良好，如机器翻译、情感分析，性能与资源消耗较平衡。      |
+| 13B  |    ~52GB |  ~26GB |  ~13GB | 具备较高准确性和生成质量，适合专业领域应用，如法律、金融等。                 |
+| 30B  |   ~120GB |  ~60GB |  ~30GB | 可处理复杂任务，如多轮对话、代码生成，性能接近人类水平。                     |
+| 65B  |   ~260GB | ~130GB |  ~65GB | 顶级模型，适用于前沿研究和高端应用，具备极强的语言理解和生成能力。           |
+
+## Train
+
+- Train Your Own O1 Preview Model Within $450
+  - https://news.ycombinator.com/item?id=43125430
 
 ## AI vs ML vs DL
 

notes/ai/ai-glossary.md

@@ -43,6 +43,38 @@ tags:
 | Speech Synthesis | 语音合成 |
 | Voice Synthesis  | 语音合成 |
 
+## 精度 {#precision}
+
+| type | byte |        dynamic | 训练中常见用途                         | GPU支持性        |
+| ---- | ---- | -------------: | -------------------------------------- | ---------------- |
+| FP64 | 8    | 极高（~10³⁰⁸） | 科学计算、极端精度需求，极少用于DL训练 | 较弱，性能低     |
+| FP32 | 4    |    高（~10³⁸） | 中小型模型训练，混合精度中的关键操作   | 广泛支持         |
+| FP16 | 2    |     低（~10⁴） | 大模型训练（需损失缩放），推理优化     | Tensor Core 加速 |
+| BF16 | 2    |    高（~10³⁸） | 大模型训练主流，数值稳定               | A100/H100 优化   |
+
+- FP64 - Double Precision 双精度
+- FP32 - Single Precision 单精度
+- FP16 - Half Precision 半精度
+- BF16 / Bfloat16
+  - Brain Floating Point 16-bit
+  - by Google Brain
+  - 保留了 FP32 的指数范围（8位指数），减少尾数（7位）
+- Float
+  - s 符号位（Sign bit）
+  - e 指数（Exponent）
+  - m 尾数（Mantissa，或称为有效数/分数）
+
+$$
+\text{FP} = (-1)^s \times 2^{e-\text{Bias}} \times (1 + m)
+$$
+
+- 1 + m
+  - 更明确地分开隐含位（1）和存储的小数部分（m）
+
+$$
+\text{FP32} = (-1)^s \times 2^{e - 127} \times (1 + m)
+$$
+
 ## LLM 参数
 
 - temperature

notes/ai/ml/jupyter/README.md

@@ -5,23 +5,28 @@ title: Jupyter
 # Project Jupyter
 
 - https://jupyter.org/
+  - Kernel
+  - Binder
+- UI/用户界面
+  - [jupyterlab/jupyterlab](https://github.com/jupyterlab/jupyterlab)
+    - BSD-3, TS, Python
+    - 多文档, 插件, 内置文件浏览, 交互输出, 更高级的文本编辑
+    - next-generation web-based user interface for Project Jupyter
+  - [jupyter/notebook](https://github.com/jupyter/notebook)
+    - 特点
+      - document-centric
+    - Notebook v7
+      - JupyterLab components for the frontend
+      - Jupyter Server for the Python server
+    - Classic Notebook v6
+      - [jupyter/nbclassic](https://github.com/jupyter/nbclassic)
+  - nbclassic
+  - Jupyter Console
+  - Qt console
   - Voilà
     - 笔记本转换为独立的交互式网页应用程序
-  - Binder
 - [jupyter/jupyter](https://github.com/jupyter/jupyter)
   - metapackage for installation, docs and chat
-- [jupyterlab/jupyterlab](https://github.com/jupyterlab/jupyterlab)
-  - BSD-3, TS, Python
-  - 多文档, 插件, 内置文件浏览, 交互输出, 更高级的文本编辑
-  - next-generation web-based user interface for Project Jupyter
-- [jupyter/notebook](https://github.com/jupyter/notebook)
-  - 特点
-    - document-centric
-  - Notebook v7
-    - JupyterLab components for the frontend
-    - Jupyter Server for the Python server
-  - Classic Notebook v6
-    - [jupyter/nbclassic](https://github.com/jupyter/nbclassic)
 - [Jupyter-kernels](https://github.com/jupyter/jupyter/wiki/Jupyter-kernels)
   - kernels -> 执行环境
 - [jupyter-server/jupyter_server](https://github.com/jupyter-server/jupyter_server)

notes/ai/ml/labeling.md

@@ -22,6 +22,12 @@ tags:
   - https://github.com/KKKSQJ/DeepLearning/tree/master/others/label_convert
   - https://openvinotoolkit.github.io/datumaro/latest/docs/data-formats/formats/index.html
 - [datumaro](./datumaro.md)
+- box
+  - xywh
+  - xywhn
+  - xyxy
+  - xyxyn
+  - cycywh
 
 ## YOLO
 

notes/ai/ml/ml-faq.md

@@ -297,3 +297,12 @@ LD_LIBRARY_PATH=/opt/conda/lib/python3.10/site-packages/nvidia/cudnn/lib/:$LD_LI
 - https://github.com/pytorch/pytorch/issues/104591
 
 ## Placeholder shape mismatches (expected 1 vs got tensorData with 2240) at dimIdx = 0
+
+## transforms.Normalize
+
+```py
+# 基于 ImageNet 的均值和标准差
+transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
+```
+
+- https://stackoverflow.com/a/58151903/1870054

notes/ai/ml/pytorch/pytorch-cookbook.md

@@ -37,6 +37,10 @@ device = torch.device((
 print(f"Using {device} device")
 ```
 
+```py
+device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+```
+
 ```py
 # 不推荐
 torch.set_default_device(device)

notes/ai/ocr/doclayout-yolo.md

@@ -19,3 +19,58 @@ tags:
   - table_footnote
   - isolate_formula
   - formula_caption
+
+```python
+#%%
+from doclayout_yolo import YOLOv10
+from huggingface_hub import hf_hub_download
+
+# model = YOLOv10.from_pretrained("juliozhao/DocLayout-YOLO-DocStructBench")
+filepath = hf_hub_download(repo_id="juliozhao/DocLayout-YOLO-DocStructBench",
+                           filename="doclayout_yolo_docstructbench_imgsz1024.pt")
+model = YOLOv10(filepath)
+#%%
+import torch
+
+device = torch.device((
+  "cuda"
+  if torch.cuda.is_available()
+  else "mps"
+  if torch.backends.mps.is_available()
+  else "cpu"
+))
+det_res = model.predict("input/2003-D30-000-0013.jpg", imgsz=1024, device=device)
+
+#%%
+import cv2
+
+annotated_frame = det_res[0].plot(pil=True, line_width=5, font_size=20)
+cv2.imwrite("result.jpg", annotated_frame)
+#%%
+result = det_res[0]
+table_class_id = None
+for id, name in det_res[0].names.items():
+  if name == "table":
+    table_class_id = id
+    break
+table_class_id
+#%%
+table_boxes = []
+for i, cls in enumerate(result.boxes.cls):
+  if int(cls.item()) == table_class_id:
+    # 获取表格边界框 [x1, y1, x2, y2]
+    box = result.boxes.xyxy[i].cpu().numpy()
+    # 获取置信度
+    conf = result.boxes.conf[i].item()
+    table_boxes.append({
+      "xywh": box.tolist(),
+      "confidence": conf
+    })
+table_boxes
+```
+
+
+```tsx
+
+
+```

notes/ai/ocr/ocr-awesome.md

@@ -81,5 +81,17 @@ tags:
   - license plates
   - https://news.ycombinator.com/item?id=37384327
 - https://github.com/kba/awesome-ocr
+- [deepdoctection/deepdoctection](https://github.com/deepdoctection/deepdoctection)
+  - Apache-2.0, Python
+  - 用到了很多东西，可以作为参考
+- LayoutLM
+  - https://huggingface.co/docs/transformers/en/model_doc/layoutlm
+  - https://medium.com/@shivarama/layoutlmv3-from-zero-to-hero-part-1-85d05818eec4
+- DiT - Document Image Text
+  - https://github.com/microsoft/unilm/tree/master/dit
+- [microsoft/unilm](https://github.com/microsoft/unilm)
+  - MIT, Python
+  - Large-scale Self-supervised Pre-training Across Tasks, Languages, and Modalities
+  - Unified Language Model Pre-training
 - 商业
   - https://doc2x.noedgeai.com/