image

CG

Dreamworks animation (DWA) usd-manager

https://github.com/dreamworksanimation/usdmanager

# virtualenv
python310 -m virtualenv venv
.\venv\scripts\activate
 
#  branch: python3
 
python setup.py install
pip install PySide2
 
python .\build\scripts-3.10\usdmanager

AYON

AYON Frontend

REM
REM ayon-frontend
REM If you don,t have yarn, you can install it by `npm install -g yarn`
REM
 
cd C:\src\ayon\ayon-frontend
 
yarn install
yarn build

AYON launcher

#
# ayon-launcher
# 1. On Windows, You need install iscc.exe by scoop install inno-setup
# 2. python39 -m virtualenv ./venv
# 
 
cd c:\src\ayon\ayon-launcher
 
# proxy setup if needed.
 
.\venv\scripts\activate
 
./tools/manage.ps1 create-env
./tools/manage.ps1 install-runtime-dependencies
./tools/manage.ps1 build
./tools/manage.ps1 make-installer

AYON Dependencies Tools

# 创建.env文件, 这个apiKey是从项目文件,找一张有图片的,在浏览器按F12,定位元素到这种图片,在url中会显示token,将token填入下方
: AYON_API_KEY=
: AYON_SERVER_URL=http://<addr>:<port>
 
cd C:\src\ayon\ayon-dependencies-tool
 
# Change python version same as your project, for me for now, it's python3.9.13, my system has 3 versions of python, 3.9,3.10,3.11, and installed by scoop, commonly I use virtualenv, but I removed the default scoop-shims of python.exe,python3.exe, and only keep the python39,python310,python311 on my system.
 
.\start.ps1 install
.\start.ps1 create -b <bundle_name>

AYON addons

# from self
# 1. use https://github.com/ynput/ayon-addon-template to create your own addon or clone the official addon repository
# 2. create a python virtualenv that version compatibility with your project (for now 24.01.17, needs python3.9.13), after create it, activate it. then run
poetry run python create_package.py
# from openpype (all)
cd c:\src\ayon\openpype
./.poetry/bin/poetry run python ./server_addon/create_ayon_addons.py

AYON Documentation

# Install yarn if needed
# npm install -g yarn
yarn add docusaurus --dev
cd website
yarn start

AYON Docker

(for windows Docker Desktop) - Recommand

# powershell
cd C:\src\ayon\ayon-docker
 
# If you have one, you can speed it up
$env:http_proxy = "http://localhost:10809"
$env:https_proxy = "http://localhost:10809"
 
# from template.json to server
#  exection only once need.
.\manage.ps1 setup
 
# Use official docker image
.\manage.ps1 update
 
# build whole things (from local)
.\manage.ps1 build

AYON Backend

(for windows WSL)

#!/bin/bash
 
# Windows (WSL)
 
#
# On Windows (Outside WSL) - Port Forward
# - Use `wsl -- ifconfig eth0` get wsl-ip, then:
# netsh interface portproxy add v4tov4 listenport=5000 listenaddress=0.0.0.0 connectport=5000 connectaddress=<wsl-ip>
# netsh interface portproxy show all
# netsh interface portproxy delete v4tov4 listenport=5000 listenaddress=0.0.0.0
#
 
#
# On WSL - Install Project (user environment, root is ok.)
# - If you have better network for pip, You don't need this line.
# export http_proxy=http://<address>:<port>
# export https_proxy=http://<address>:<port>
# pip install -U pip
# pip install poetry
# poetry config virtualenvs.create false
# python3 -m virtualenv /home/<user>/venv-ayon
# source /home/<user>/venv-ayon/bin/activate
# cd /mnt/c/src/ayon/ayon-backend
# poetry install --no-interaction --no-ansi
#
 
#
# On noth (Manual)
# - Windows
# wsl --shutdown -d Ubuntu-22.04
# wsl -d Ubuntu-22.04
# - WSL (ayon.pid, backend-server need, for non-root user only.)
# sudo touch /var/run/ayon.pid
# sudo chown <user> /var/run/ayon.pid
#
 
#
# ! Need yarn build frontend first.
#
 
# sh /mnt/c/src/ayon/ayon-backend.sh
 
. /home/<user>/venv-ayon/bin/activate
 
echo <password> | sudo -S touch /var/run/ayon.pid
echo <password> | sudo -S chown <user> /var/run/ayon.pid
 
# host_ip=$(cat /etc/resolv.conf |grep "nameserver" |cut -f 2 -d " ")
# export ALL_PROXY="http://$host_ip:<port>"
 
export AYON_postgres_url=postgres://<username>:<password>@<addr>:<port>/ayon
export AYON_redis_url=redis://<username>:<password>@<addr>:<port>
 
export AYON_frontend_dir=/mnt/c/src/ayon/ayon-frontend/dist
export AYON_addons_dir=/mnt/c/src/ayon/ayon-addons/dist
export AYON_log_file=/mnt/c/src/ayon/log-backend.log
export AYON_SETTINGS_TEMPLATE=/mnt/c/src/ayon/template.json
export AYON_force_create_admin=1
export AYON_auth_pass_pepper=<secure_password>
 
export PYTHONBUFFERED=1
export LOGLEVEL=info
 
cd /mnt/c/src/ayon/ayon-backend
 
gunicorn \
  -k uvicorn.workers.UvicornWorker \
  --log-level ${LOGLEVEL} \
  --timeout 120 \
  -b :5000 \
  ayon_server.api.server:app

AYON Openpype (Deprecated)

更详细: CG Pipeline

GitHub - ynput/OpenPype: Main OpenPype repository and AYON openpype addon codebase

开源Pipeline管理工具。

# python version 3.9.13
$Env:http_proxy = "http://localhost:10809"
$Env:https_proxy = "http://localhost:10809"
.\tools\create_env.ps1
.\tools\fetch_thirdparty_libs.ps1
.\tools\build.ps1

Armory

ArmorCore

cd armorcore
# Unpack `v8\libraries\win32\release\v8_monolithz` using 7-Zip - Extract Here (exceeds 100MB)
.\Kinc\make.bat -g direct3d11

ArmorPaint

GitHub - armory3d/armortools: 3D Content Creation Tools

贴图绘制软件。

# 解压 `..\armorcore\v8\libraries\win32\release\v8_monolithz` using 7-Zip - Extract Here (exceeds 100MB)
 
# 生成工程
cd c:/src/armortools/armorpaint
..\armorcore\Kinc\make --from ..\armorcore -g direct3d11
 
# 在"build\ArmorPaint.sln"中打开生成的 Visual Studio 项目针对 x64 和Release进行构建和运行
 
# Copy build\x64\Release\ArmorPaint.exe to build\krom to run ArmorPaint.exe directly

Armory3D

GitHub - armory3d/armory: 3D Engine with Blender Integration

一个游戏引擎,可以在 Blender 中构建游戏。

ArmorLab

AI驱动的纹理创作工具,图片或文字->贴图。

cd armortools/armorlab
git clone https://github.com/armory3d/onnx_bin onnx
# Unpack `models.zip` from https://github.com/armory3d/armorai/releases into `Assets/models` using 7-Zip - Extract Here
 
# Unpack `..\armorcore\v8\libraries\win32\release\v8_monolithz` using 7-Zip - Extract Here (exceeds 100MB)
..\armorcore\Kinc\make --from ..\armorcore -g direct3d11

OpenUSD

GitHub - PixarAnimationStudios/OpenUSD: Universal Scene Description

数据交换格式。

构建

# python version 3.10.11
# run `x64 Native Tools Command Prompt for VS 2022`
# for download usd-src
set http_proxy=http://localhost:10809
set https_proxy=http://localhost:10809
# venv
python -m pip install virtualenv
python -m virtualenv C:/src/OpenUSD/venv
call C:/src/OpenUSD/venv/Scripts/activate
# for usdview
pip install PyOpenGL
pip install PySide6
pip install opencolorio
pip install numpy
# for usdSchemaGen
pip install jinja2
# build dry run, to build remove -n option
python .\build_scripts\build_usd.py --materialx --draco --alembic --hdf5 --openimageio --opencolorio --embree --openvdb --ptex --prman --test --build-variant=relwithdebuginfo --src=c:\data\build\usd-src --build-args=USD,"-DPXR_ENABLE_VULKAN_SUPPORT=FALSE -DRENDERMAN_LOCATION=C:/data/plugins/RenderManProServer-25.2" c:\data\build\usd-build –n

Schema

usdGenSchema [schemaPath] [codeGenPath]

Blender

Blender Developer Wiki

开源全功能的DCC软件。

git clone https://projects.blender.org/blender/blender.git
cd C:\src\blender-git\blender
make update
make full nobuild

QuiltiX

GitHub - PrismPipeline/QuiltiX: QuiltiX is a graphical node editor to edit, and author MaterialX based materials of 3D assets

基于节点的MaterialX材质编辑。

python version 3.9.13

# via pypi (install)
pip install QuiltiX
pip install git+https://github.com/PrismPipeline/OpenUSD_build.git@23.11-win-mtlx-1.38.8
 
# run
python -m QuiltiX

gatling

GitHub - pablode/gatling: Hydra-enabled GPU path tracer that supports MaterialX and MDL

Hydra光追渲染器。

cmake -B build -Wno-dev -DUSD_ROOT=%E_USD_LOCATION% -DMDL_ROOT=%E_MDLSDK_DIR% -DCMAKE_INSTALL_PREFIX=C:/data/build/gatling -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release --target INSTALL -- /M:8

xComp

GitHub - gugenstudio/xComp: best render comparing software in town!

cmake -B build -DCMAKE_TOOLCHAIN_FILE="%E_VCPKG_DIR%\scripts\buildsystems\vcpkg.cmake"
 
cmake --build build --target xcomp --config Release
 
# Working-dir on 
#   C:/src/xComp/apps/xcomp/../debug_dir
# Copy _bin/Release/* to Working-dir

AI

SD-WebUI-EasyPhoto

GitHub - aigc-apps/sd-webui-EasyPhoto: 📷 EasyPhoto | Your Smart AI Photo Generator.

输入了一些右下方的图片训练出Lora后,就可以扩散风格到人物肖像上。

1702548443361

ComfyUI

GitHub - comfyanonymous/ComfyUI: The most powerful and modular stable diffusion GUI with a graph/nodes interface.

SD的图形用户界面,具有节点图。

1702519000724

Stable-diffusion-Webui

GitHub - AUTOMATIC1111/stable-diffusion-webui: Stable Diffusion web UI

# python version 3.10.11
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
 
.\webui-user.bat

config

webui-user.bat

set COMMANDLINE_ARGS=--ckpt-dir <path> --lora-dir <path>

Krita-ai-diffusion

GitHub - Acly/krita-ai-diffusion: Streamlined interface for generating images with AI in Krita. Inpaint and outpaint with optional text prompt, no tweaking required.

在 Krita 中使用AI生成图像的一个插件。

MiniGPT-4

GitHub - Vision-CAIR/MiniGPT-4: Open-sourced codes for MiniGPT-4 and MiniGPT-v2 (https://minigpt-4.github.io, https://minigpt-v2.github.io/)

1.我使用的Vicuna-7B,low_resource=False, 最少GPU显存 16GB。如果是True,我这里使用了10GB。 2.安装CUDA Toolkit 12/11。

1702518989918

Windows的支持

Windows not supported · Issue #28 · Vision-CAIR/MiniGPT-4 · GitHub

# 也许可以不用conda,使用virtualenv
git clone https://github.com/Vision-CAIR/MiniGPT-4.git
cd MiniGPT-4
##conda env create -f environment.yml
##conda activate minigptv
 
# 卸载不支持GPU版本的包
pip uninstall bitsandbytes
pip uninstall torch
# 重新安装torch与bitsandbytes
##这个不更新啦,使用下面那个
##pip install git+https://github.com/Keith-Hon/bitsandbytes-windows.git
pip install scipy
pip install bitsandbytes==0.41.2 --index-url=https://jllllll.github.io/bitsandbytes-windows-webui
 
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
 
---
 
python demo_v2.py --cfg-path eval_configs/minigptv2_eval.yaml  --gpu-id 0

FastChat

GitHub - lm-sys/FastChat: An open platform for training, serving, and evaluating large language models. Release repo for Vicuna and Chatbot Arena. GPU16GB显存,加载了vicuna-7b和vicuna-13b。

git clone https://github.com/lm-sys/FastChat.git
cd FastChat
# use venv
pip3 install --upgrade pip  # enable PEP 660 support
pip3 install -e ".[model_worker,webui]"
 
---
 
# cli
python -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5-16k --load-8bit --cpu-offloading
 
# gradio
python -m fastchat.serve.controller
python -m fastchat.serve.model_worker --model-path lmsys/vicuna-7b-v1.5-16K --load-8bit --cpu-offloading
python -m fastchat.serve.gradio_web_server

Ye

ub

python开发环境

python310 -m virtualenv C:\src\Ye\src\ub\venv
# pip install dependenties with requirements.txt

Dev

(Tools)

pyside-setup

vcpkg

# download & install
git clone https://github.com/microsoft/vcpkg
.\vcpkg\bootstrap-vcpkg.bat
 
# qt6
...

(Python)

conda

# Shell Register
conda init cmd.exe
# Shell Unregister
conda init cmd.exe --reverse
 
# Update
conda update -n base -c defaults conda
## Or minimize update use
conda install conda=23.11.0

pyenv

Till the 2023-12-08,Using scoop install pyenv causing error core compoment install error. So we use this command Invoke-WebRequest -UseBasicParsing -Uri "https://raw.githubusercontent.com/pyenv-win/pyenv-win/master/pyenv-win/install-pyenv-win.ps1" -OutFile "./install-pyenv-win.ps1"; &"./install-pyenv-win.ps1" from Installing and setting up Python - Everything I Know

在这之后,需要执行 pyenv update 去更新仓库。

Building Ye and blender warning. say python not found.

pipx

GitHub - pypa/pipx: Install and Run Python Applications in Isolated Environments

# python version 3.10.11
python -m pip install --user pipx
 
# add executable to environment path
python -m pipx ensurepath
 
pipx install pycowsay
 

poetry

GitHub - python-poetry/poetry: Python packaging and dependency management made easy

Install and Upgrade

python -m pipx install poetry
python -m pipx upgrade poetry
 
# Or
# create virtualenv, then (Recommand)
pip install poetry
 
# Or
 
pip install pipx
pipx install poetry
pipx upgrade poetry

(Version Control)

  1. gitbucket
scoop bucket add java
scoop install temurin11-jdk
 
set GITBUCKET_PORT=12122
set GITBUCKET_HOST=0.0.0.0
set GITBUCKET_HOME=<your_data_path> # The default is $HOME/gitbucket
java -jar gitbucket.war

https://github.com/gitbucket/gitbucket

  1. git
  2. SnowFS

  1. TortoiseSVN
# 重置当前仓库
svn revert --recursive .
# 忽略文件(项目级)
图形界面 TortoiseSVN -> Properties -> New -> Other -> svn:global-ignores
# 忽略文件(目录级)
选中一个文件/目录 -> TortoiseSVN -> Properties -> New -> Other -> svn:ignore
  1. VisualSVN Server

INFO

SVN -> Import 可以将当前目录的所有文件上传项到SVNServer的版本库。 SVN -> Checkout 输入仓库地址与路径可以检出。

ERROR

1.SVN连接不到VisualSVN服务,检查是不是开启了代理。%appdata%\Subversion\servers 编辑这个文件,注释掉下方两行。

#http-proxy-host = 127.0.0.1
#http-proxy-port = 10809

2.SVN图标在Winodws11下不显示。原因是SVN的图标在注册表中靠后了。(Windows按照空格来排),(Explorer空白处右键 ) -> TortoiseSVN -> Settings -> Icon Overlays -> Overlay Handles -> Start registry Editor进行编辑,增加多几个空格使其排名靠前。

(Database Connection)

Mongosh - https://www.mongodb.com/try/download/compass Mysqlsh - https://dev.mysql.com/downloads/shell/

for database install on Linux, see OS-Linux

(Cloud Services)

我们缺少的不是工具,而是真正产生价值的点子。

Vercel Cloudflare fly.io

(AI)

概念

  • pytorch: Start Locally | PyTorch

  • tensorflow: GitHub - tensorflow/tensorflow: An Open Source Machine Learning Framework for Everyone

  • ONNX:Open Neural Network Exchange(开放式神经网络交换)ONNX | Home

  • OpenVINO - 从框架中优化深度学习模型。

  • RL-LLM - 大模型+强化学习

  • Gradient descent 梯度下降 - (鞍点,局部最优)

    • derivate - 求导,learning rate 学习速率
    • sgd 随机梯度下降,Momentum动量,Adam (替代随机梯度下降算法)
    • partial derivate 偏微分
    • gradient 梯度 - 所有的偏微分的向量 e.g.
    • Convex function 凸函数 - 全局最优解
    • 梯度离散/爆炸 - 误差积累导致,前层得不到更新
  • Backpropagation 反向传播

  • Activation Functions 激活函数 - 阈值响应机制

    • Sigmoid -
    • Tanh -
    • Rectified Linear Unit (ReLU) -
    • Leaky ReLU
    • SELU - 两个函数的合并
    • Softmax - [Logits Scores -> Softmax -> Probabilities], 概率和为1
    • Softplus - ReLU函数的平滑版本
  • Typical Loss

    • Mean Squared Error (MSE) 均方误差
    • Cross Entropy Loss 交叉熵
    • Hinge Loss
  • Himmelblau Function函数 - 数学家构造的特殊函数,用来测试深度学习算法是否能够收敛到局部最小值。

  • Chain rule链式法则 - 1702375241410 求解过程是

  • Dropout regularization - 为了防止过拟合。Dropout的思想是对于网络的每一层,随机的丢弃一些单元。

  • PCA - 主元分析,用于提取数据的主要特征分量。主要用于数据降维。例如两维数据房价与面积,房价与面积的正相关性很大,其中一维的数据是冗余的。

  • Sequence 序列

    • Word Embedding
    • Word2vec
    • GloVe
    • One-hot: [0,1,0,0,…] - 稀疏,高维
  • AutoEncoders - 1702399346359 中间是Neck,可以升/降维。

    • Denoising AutoEncoders 去噪自编码器
    • Dropout AutoEncoders
    • Adversarial AutoEncoders 对抗自编码器
    • Variational AutoEncoder (VAE) 变分自编码器
  • Gradient Clipping - 用于缓解梯度爆炸。

  • Zeno/One/Few-Shot Learning 零/一/少样本学习。例如,当我们只有一张狮子的照片时,一次样本学习可以帮助我们将新的狮子图像正确分类。

  • RLHF(Reinforcement Learning from Human Feedback,人类反馈强化学习)。是GPT-4/ChatGPT与GPT-3.5的主要区别。

  • Anomaly Detection 异常侦测

  • LoRA,英文全称Low-Rank Adaptation of Large Language Models,直译为大语言模型的低阶适应,这是微软的研究人员为了解决大语言模型微调而开发的一项技术。

  • PGNet是一种新颖的框架,用于通过点收集操作实时阅读文本。 基于多任务学习的单步文本检测识别器。

  • SFT(Supervised Fine-Tuning)监督微调


  • AGI(通用人工智能)1.人工智慧是我们的目标。2.人类的能力是本能与学习。
    • Machine Learning - 1.是达成目标的手段。2.就是让机器找到一个函式(找到函式三步骤- a.设定范围(订出候选函式集合-Model),b.设定标准(评量函式好坏-Loss),c.达成目标(找出最好的函式-最佳化Optimization))。
      • Deep Learning 深度学习 1.是机器学习的其中一种方法
  • Supervised Learning
    • Linear Regression 线性回归
    • Logistic Regression 逻辑回归 - 线性回归的基础上加了一个激活函数(压缩函数)
    • Classification 分类问题
    • Semi-supervised Learning - 半监督学习。
  • Unsupervised Learning
  • Reinforcement Learning - 增强学习。游戏例如Dota,与环境交互反馈。
  • Representation Learning - 表征学习。
  • Structured (Generative) Learning - 结构(生成式)学习。
  • Meta Learning - 元学习,也叫做学会学习,learn to learn。人类学习新东西时会用到之前学的东西,比如你会玩LOL,那么对玩王者荣耀会很快。目前的深度学习遇到的问题是在遇到新的问题,即使比较类似的情况下,也要重新学习,这一差异就是MetaLearning的动机。
  • Transfer Learning 迁移学习 一种神经网络学习到了识别Cat这种物体,再使用这部分知识或者部分知识去识别X-Ray扫描。(Pre Training 预训练, Fine Turning 微调)
  • Multitask Learning - 神经网络同时执行多项任务,每个任务都会影响其他任务

  • Feedforward neural network (前馈神经网络) 也叫做感知机(Perceptron) - 二分类线性模型

  • SVM (Support vector machine) 支持向量网络 - 1702374445464

  • 卷积神经网络:局部感知 -

    • Weight sharing 权值共享 1702379993258
    • 池化 - Max pooling, Avg polling
      • Downsample, Upsample, Subsampling
  • RNN 循环神经网络 - (Hard to parallel) 语境信息上下文,引入memory的机制, image, 问题是很容易造成梯度离散或梯度爆炸。

    • LSTM - (Long Short-Term Memory) 长短时记忆 - 解决了记忆长度的问题。引入”闸门”的机制。1702394510046 闸门就是Sigmoid函数。输入门,遗忘(记忆)门。
  • CNN 卷积神经网络 - Using CNN to replace RNN (CNN Can parallel)

  • GAN 生成对抗性网络

  • Self-Attention Layer - Transformer

    • q:query (to match other)
    • k:key (to be matched)
    • v:information to be extracted
    • Positional Encoding ()
  • Multi-head Self-Attention (MHSA)

    • Norm (Layer Norm, Batch Norm, Instance Norm, Group Norm)
  • Transformer 1702297659308 这是一个seq2seq的model,左半部是Encoder,右半部是Decoder。

    • Universal Transformer - (横轴Positions使用Transformer,纵轴Depth使用RNN)
    • Vision Transformer - 是Google提出的基于Transformer的图片分类模型。1702523427973
  • Self-attention GAN - 用于影像处理的Transformer,一个pixel考虑全部pixel信息。

  • 扩散模型

    • Noising Process image
  • 多模态模型架构 1702558622286

    • 1)合并注意力架构(Merge-attention)
    • 2)共同注意力架构(Co-attention)
    • 3)交叉注意力架构(Cross-attention)
    • 4)三角Transformer架构(Tangled-transformer)
    • 5)模态间对比学习架构(Inter-Modality Contrastive Learning)

模型

  • 大型语言模型(LLM)- 文字接龙..

    • ChatGPT
    • nanoGPT 是一个用于对中等规模的生成式预训练 Transformer(GPT)进行训练和调优的框架。
    • Llama2
    • Mistral-7B
    • Mixtral 8x7B - MoE(Mixture of Experts)
    • WizardCoder
    • 羊驼系列
      • Vicuna -Based On LLaMa-13B
      • Alpaca -Based On LLaMa-7B
      • Guanaco - Based On QLoRA quantification
    • LaMDA (Language Models for Dialog Applications)
  • 多模态模型

    • GPT 生成式预训练变换模型
    • Macaw-LLM:具有图像、视频、音频和文本集成的多模态语言建模
    • CLIP ViT-L (Vision Transformer - Low resolution): Connecting text and images。
    • Stable Diffusion
    • SVD (Stable Video Diffusion)
    • Flamingo
    • LLaVA
  • 扩散模型

    • 连续型扩散模型

      • DDPM 去噪扩散概率模型(Denoising Diffusion Probabilistic Models)
      • DDIM 去燥扩散隐式模型 (Diffusion Denoising Implicit Model)
      • Improved Diffusion
      • Classifier Guidance 亦被称为Guided Diffusion
      • Classifier Free Guidance
      • DALL-E-2 - AI图形生成器
      • Google Imagen
      • LCM (Latent Consistency Model)
    • 离散型扩散模型

      • Multinomial Diffusion
      • D3PM
      • ImageBART
  • 语言表征模型

    • BERT
    • GPT
    • ELMO
  • 视觉模型

  • 卷积神经网络模型

    • ResNet (深度残差网络): 短路层(Skip to shortcut)
    • AlexNet
    • GoogLeNet - 22 layers
    • VGGNet-11/16/19 - 更小型的卷积核心。1x1 convolutions.
    • LeNet - 80年代
    • UNet - 为生物医学图像分割开发的卷积神经网络。
1702522306435
  • 其它模型
    • SMPL - A Skinned Multi Person Linear Model,用于人体动作捕捉。1704342182177

有哪些数据集

  • MNIST - 手写数字
  • ImageNet
  • CIFAR10/100 - 10类6000张照片。

资料

谷歌研究论坛Google Research Blog 词嵌入数据可视化Embedding projector - visualization of high-dimensional data LLM可视化 https://bbycroft.net/llm

李宏毅 吴恩达 Andrew NG

学习路径

  1. 10分鐘了解RNN的基本概念 [6AW80qmaAOk]
  2. Let’s build GPT: from scratch, in code, spelled out. [kCc8FmEb1nY]
  3. Multitask Learning (C3W2L08) [UdXfsAr4Gjw]
  4. Transfer Learning (C3W2L07) [yofjFQddwHE]
  5. 扩散模型是如何工作的-系列 [oSmlciqXOaU]
  6. PyTorch深度学习框架-龙良曲

应用

1702620460693
  • GPT
    • 文字冒险游戏 (ChatGPT) 1702572278800
    • 解决数学问题(FunSearch)
  • Diffusion
    • 可控的图像生成 (ControlNet) 1702572357330
    • 图像编辑(概念转换) (DreamBooth) 1702572462628
    • 图像编辑(文本驱动)
    • 图像修复/扩展
    • 图像的风格迁移 (BLIP-Diffusion) 1702572827664
    • 图像的交互式编辑 (DragGAN,DragDiffusion)

(Graphics)

概念

  • 场
    • 标量场 - 符号距离场(SDF), 密度场(Density Field), 散度场(描述一点是汇聚还是扩散)
      • 符号距离场 - 梯度即法线,大小即距离
    • 向量场 - 速度场,涡度场,加速度场
  • 场的操作
    • 标量场 - 梯度(输出:向量场),曲率,拉普拉斯(作用,平滑,算子:热扩散)
    • 向量场 - 散度(输出:标量场),旋度,归一化
    • RenormalizeSDF - 包面的几何只有表示内外而没有距离。需要做Eikonal equation(程函方程)得到带有距离的SDF
  • 场,几何,粒子
    • 场实际上是任意的几何与粒子的中间介质(中间表达)
  • 自由表面 (free surface),包面后的几何
  • Eikonal equation(程函方程) - 程函方程的一般形式为, 约束条件:边界为;是位置的函数, 是欧几里得范数,是从边界到Ω内部x的所需要的最短时间,F(x)是在x处耗费的时间。 1702320461230

(Robot)