Skip to content

macos-setup-Silicon.md

Latest commit

 

History

History
304 lines (214 loc) · 7.33 KB

macos-setup-Silicon.md

File metadata and controls

304 lines (214 loc) · 7.33 KB

1. 环境信息

  • 硬件：Apple MacBook Pro 14寸，Apple M1 Pro 芯片，16GB 内存
  • 操作系统：macOS Sequoia 15.1
  • Unity 版本：2023.2.20f1
  • ML-Agents（Unity端）：com.unity.ml-agents 3.0.0（通过 Unity Package Manager 官方源安装）
  • Python 版本：3.8.x（x86_64 架构，需通过 Rosetta 终端运行）
  • mlagents（Python端）：0.28.0
  • PyTorch：1.8.0
  • Miniforge/Miniconda：x86_64 版本

2. 安装与配置步骤

2.1 安装 x86_64 版 Miniforge3（必须用 Rosetta 终端）

  1. 用 Rosetta 打开终端
    Finder > 应用程序 > 实用工具 > 右键"终端" > 显示简介 > 勾选"使用 Rosetta 打开"

  2. 下载 x86_64 版 Miniforge3 安装包
    https://github.com/conda-forge/miniforge/releases/download/23.11.0-0/Miniforge3-MacOSX-x86_64.sh

  3. 安装 Miniforge3

    cd ~/Downloads
bash Miniforge3-MacOSX-x86_64.sh

  4. 安装完成后，重启 Rosetta 终端

  5. 验证 Python 架构

    python -c "import platform; print(platform.machine())"

    输出应为 x86_64

2.2 创建 Python 3.8 虚拟环境

conda create -n mlagents-028 python=3.8
conda activate mlagents-028

2.3 安装依赖包（严格指定版本）

conda install pytorch==1.8.0 torchvision==0.9.0 cpuonly -c pytorch
pip install mlagents==0.28.0 mlagents-envs==0.28.0
pip install protobuf==3.20.3
pip install numpy==1.19.5
pip install six

2.4 Unity 端 ML-Agents 安装

  1. 打开 Unity 项目
  2. 打开 Package Manager（菜单栏 Window > Package Manager）
  3. 切换到 Unity Registry，搜索 ml-agents
  4. 选择并安装 com.unity.ml-agents 3.0.0 版本
    • 如未显示，点击左上角"+" > "Add package by name..."
    • Name: com.unity.ml-agents
    • Version: 3.0.0
  5. 确认 ML-Agents 3.0.0 安装成功，且不是 local 版本

2.5 Unity 场景配置

  • Behavior Parameters 组件
    • Behavior Name 与 trainer_config.yaml 配置文件中的 key 完全一致（区分大小写）
    • Behavior Type 设为 Default
  • 场景可正常运行，Agent 能运动，无报错

2.6 训练配置文件

  • 在 Unity 项目目录下准备好 trainer_config.yaml，内容示例：

    gewu:
  trainer_type: ppo
  max_steps: 20000000
  batch_size: 2048
  buffer_size: 20480
  learning_rate: 0.0003
  network_settings:
    hidden_units: 512
    num_layers: 3

2.7 启动训练

  1. Unity 编辑器中点击 Play，保持场景运行

  2. 在终端（已激活 mlagents-028 环境，且在项目目录下）运行：

    mlagents-learn trainer_config.yaml --run-id=run1 --force

  3. 观察终端和 Unity Console 日志，确认训练正常进行

2.8 训练中断与恢复

  • 训练过程中可随时 Ctrl+C 停止，ML-Agents 会自动保存 checkpoint
  • 恢复训练时，使用相同的 --run-id，ML-Agents 会自动从最近的 checkpoint 继续

3. 参考命令汇总

# 用 Rosetta 终端
bash Miniforge3-MacOSX-x86_64.sh

# 创建并激活环境
conda create -n mlagents-028 python=3.8
conda activate mlagents-028

# 安装依赖
conda install pytorch==1.8.0 torchvision==0.9.0 cpuonly -c pytorch
pip install mlagents==0.28.0 mlagents-envs==0.28.0
pip install protobuf==3.20.3
pip install numpy==1.19.5
pip install six

# 启动训练
mlagents-learn trainer_config.yaml --run-id=run1 --force

4. 重要说明

  • 所有命令均需在 Rosetta 终端下执行，确保 x86_64 架构
  • Unity 端 ML-Agents 必须用官方源安装的 3.0.0 版本，切勿用 local 包
  • Python 端所有依赖需严格指定版本，避免兼容性问题

1. Environment Information

  • Hardware: Apple M1 Pro
  • Unity Version: 2023.2.20f1
  • ML-Agents (Unity side): com.unity.ml-agents 3.0.0 (installed via Unity Package Manager Registry)
  • Python Version: 3.8.x (x86_64 architecture, must run via Rosetta Terminal)
  • mlagents (Python side): 0.28.0
  • PyTorch: 1.8.0
  • Miniforge/Miniconda: x86_64 version

2. Installation & Configuration Steps

2.1 Install x86_64 Miniforge3 (Must use Rosetta Terminal)

  1. Open Terminal with Rosetta
    Finder > Applications > Utilities > Right-click "Terminal" > Get Info > Check "Open using Rosetta"

  2. Download x86_64 Miniforge3 installer
    https://github.com/conda-forge/miniforge/releases/download/23.11.0-0/Miniforge3-MacOSX-x86_64.sh

  3. Install Miniforge3

    cd ~/Downloads
bash Miniforge3-MacOSX-x86_64.sh

  4. After installation, restart Rosetta Terminal

  5. Verify Python architecture

    python -c "import platform; print(platform.machine())"

    Output should be x86_64

2.2 Create Python 3.8 Virtual Environment

conda create -n mlagents-028 python=3.8
conda activate mlagents-028

2.3 Install Dependencies (Strict Version)

conda install pytorch==1.8.0 torchvision==0.9.0 cpuonly -c pytorch
pip install mlagents==0.28.0 mlagents-envs==0.28.0
pip install protobuf==3.20.3
pip install numpy==1.19.5
pip install six

2.4 Install ML-Agents on Unity Side

  1. Open your Unity project
  2. Open Package Manager (Window > Package Manager)
  3. Switch to Unity Registry, search for ml-agents
  4. Select and install com.unity.ml-agents version 3.0.0
    • If not listed, click "+" > "Add package by name..."
    • Name: com.unity.ml-agents
    • Version: 3.0.0
  5. Ensure ML-Agents 3.0.0 is installed from Registry, not as a local package

2.5 Unity Scene Configuration

  • Behavior Parameters Component
    • Behavior Name must exactly match the key in trainer_config.yaml (case-sensitive)
    • Set Behavior Type to Default
  • Scene should run without errors, Agent should move as expected

2.6 Training Config File

  • Place trainer_config.yaml in your Unity project directory, for example:

    gewu:
  trainer_type: ppo
  max_steps: 20000000
  batch_size: 2048
  buffer_size: 20480
  learning_rate: 0.0003
  network_settings:
    hidden_units: 512
    num_layers: 3

2.7 Start Training

  1. Click Play in Unity Editor, keep scene running

  2. In terminal (mlagents-028 env activated, in project dir), run:

    mlagents-learn trainer_config.yaml --run-id=run1 --force

  3. Monitor terminal and Unity Console logs to ensure training is running

2.8 Pause and Resume Training

  • You can stop training anytime with Ctrl+C, ML-Agents will auto-save checkpoints.
  • To resume, use the same --run-id, ML-Agents will continue from the latest checkpoint.

3. Command Summary

# Use Rosetta Terminal
bash Miniforge3-MacOSX-x86_64.sh

# Create and activate environment
conda create -n mlagents-028 python=3.8
conda activate mlagents-028

# Install dependencies
conda install pytorch==1.8.0 torchvision==0.9.0 cpuonly -c pytorch
pip install mlagents==0.28.0 mlagents-envs==0.28.0
pip install protobuf==3.20.3
pip install numpy==1.19.5
pip install six

# Start training
mlagents-learn trainer_config.yaml --run-id=run1 --force

4. Important Notes

  • All commands must be run in Rosetta Terminal to ensure x86_64 architecture
  • ML-Agents in Unity must be installed from the official Registry, not as a local package
  • All Python dependencies must be strictly versioned to avoid compatibility issues