nvidia-smi  # 查看驅(qū)動(dòng)版本和CUDA版本

# 創(chuàng)建名為pytorch的環(huán)境，指定Python版本
conda create -n pytorch python=3.11 -y

# 激活環(huán)境
conda activate pytorch

# 驗(yàn)證環(huán)境激活
which python  # 應(yīng)顯示 .../anaconda3/envs/pytorch/bin/python

# 在激活的環(huán)境中升級(jí)pip
python -m pip install --upgrade pip

# 檢查Python版本
python --version

# 檢查conda環(huán)境
conda info --envs

# 激活環(huán)境
conda activate pytorch

# 安裝CPU版本PyTorch
conda install pytorch torchvision torchaudio cpuonly -c pytorch -y

# 激活環(huán)境
conda activate pytorch

# 安裝CPU版本PyTorch
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

import torch
print(f"PyTorch版本: {torch.__version__}")
print(f"CUDA可用: {torch.cuda.is_available()}")  # 應(yīng)該返回False
print(f"設(shè)備: {torch.device('cpu')}")

# 檢查NVIDIA GPU
nvidia-smi

# 輸出示例:
# +-----------------------------------------------------------------------------+
# | NVIDIA-SMI 535.129.03   Driver Version: 535.129.03   CUDA Version: 12.2     |
# +-------------------------------+----------------------+----------------------+

# 激活環(huán)境
conda activate pytorch

# 安裝CUDA 12.1版本PyTorch
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia -y

# 激活環(huán)境
conda activate pytorch

# 安裝CUDA 12.1版本PyTorch
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# conda安裝
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia -y

# pip安裝
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

import torch
print(f"PyTorch版本: {torch.__version__}")
print(f"CUDA可用: {torch.cuda.is_available()}")
print(f"CUDA版本: {torch.version.cuda}")
print(f"GPU數(shù)量: {torch.cuda.device_count()}")
print(f"當(dāng)前GPU: {torch.cuda.get_device_name(0)}")
# 測(cè)試GPU計(jì)算
if torch.cuda.is_available():
    x = torch.randn(3, 3).cuda()
    y = torch.randn(3, 3).cuda()
    z = torch.mm(x, y)
    print("GPU計(jì)算測(cè)試成功!")
    print(f"結(jié)果設(shè)備: {z.device}")

# 1. 添加ROCm倉(cāng)庫(kù)（Ubuntu示例）
wget https://repo.radeon.com/rocm/rocm.gpg.key
sudo gpg --dearmor --output /usr/share/keyrings/rocm.gpg /tmp/rocm.gpg.key
echo 'deb [arch=amd64 signed-by=/usr/share/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/5.4.2 focal main' | sudo tee /etc/apt/sources.list.d/rocm.list

# 2. 安裝ROCm
sudo apt update
sudo apt install rocm-libs miopen-hip hipfft rccl

# 3. 創(chuàng)建并激活conda環(huán)境
conda create -n pytorch-rocm python=3.11 -y
conda activate pytorch-rocm

# 4. 安裝ROCm版本PyTorch
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2

import torch
print(f"PyTorch版本: {torch.__version__}")
print(f"ROCm可用: {torch.cuda.is_available()}")  # ROCm也使用cuda接口
print(f"設(shè)備數(shù)量: {torch.cuda.device_count()}")
if torch.cuda.is_available():
    print(f"設(shè)備名稱(chēng): {torch.cuda.get_device_name(0)}")
    # 測(cè)試計(jì)算
    x = torch.randn(1000, 1000).cuda()
    y = torch.randn(1000, 1000).cuda()
    z = torch.mm(x, y)
    print("ROCm計(jì)算測(cè)試成功!")

# Apple Silicon原生支持
conda create -n pytorch-m1 python=3.11 -y
conda activate pytorch-m1

# 安裝支持Metal的PyTorch版本
pip install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cpu

# 或者使用穩(wěn)定版本（自動(dòng)包含Metal支持）
pip install torch torchvision torchaudio

import torch
import torch.backends.mps
print(f"PyTorch版本: {torch.__version__}")
print(f"MPS可用: {torch.backends.mps.is_available()}")
print(f"MPS內(nèi)置: {torch.backends.mps.is_built()}")
if torch.backends.mps.is_available():
    device = torch.device("mps")
    x = torch.randn(3, 3, device=device)
    y = torch.randn(3, 3, device=device)
    z = torch.mm(x, y)
    print("Metal加速測(cè)試成功!")
    print(f"結(jié)果設(shè)備: {z.device}")
else:
    print("MPS不可用，使用CPU")

#!/usr/bin/env python3
"""
PyTorch環(huán)境完整驗(yàn)證腳本
"""
import torch
import torchvision
import torchaudio
def test_basic():
    """基礎(chǔ)功能測(cè)試"""
    print("=" * 50)
    print("基礎(chǔ)信息:")
    print(f"PyTorch版本: {torch.__version__}")
    print(f"TorchVision版本: {torchvision.__version__}")
    print(f"Torchaudio版本: {torchaudio.__version__}")
    print(f"Python版本: {torch.__version__.split('+')[0]}")
def test_cuda():
    """CUDA/GPU測(cè)試"""
    print("\n" + "=" * 50)
    print("CUDA/GPU信息:")
    if torch.cuda.is_available():
        print("? CUDA可用")
        print(f"CUDA版本: {torch.version.cuda}")
        print(f"GPU數(shù)量: {torch.cuda.device_count()}")
        for i in range(torch.cuda.device_count()):
            print(f"GPU {i}: {torch.cuda.get_device_name(i)}")
        # 測(cè)試GPU計(jì)算
        try:
            x = torch.randn(1000, 1000).cuda()
            y = torch.randn(1000, 1000).cuda()
            z = torch.mm(x, y)
            print("? GPU計(jì)算測(cè)試通過(guò)")
        except Exception as e:
            print(f"? GPU計(jì)算測(cè)試失敗: {e}")
    else:
        print("? CUDA不可用")
def test_mps():
    """Metal Performance Shaders測(cè)試 (Apple Silicon)"""
    print("\n" + "=" * 50)
    print("Metal Performance Shaders (MPS) 信息:")
    if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
        print("? MPS可用")
        device = torch.device("mps")
        x = torch.randn(100, 100, device=device)
        y = torch.randn(100, 100, device=device)
        z = torch.mm(x, y)
        print("? MPS計(jì)算測(cè)試通過(guò)")
    else:
        print("? MPS不可用")
def test_cpu():
    """CPU測(cè)試"""
    print("\n" + "=" * 50)
    print("CPU信息:")
    print(f"CPU線(xiàn)程數(shù): {torch.get_num_threads()}")
    # 測(cè)試CPU計(jì)算
    x = torch.randn(1000, 1000)
    y = torch.randn(1000, 1000)
    z = torch.mm(x, y)
    print("? CPU計(jì)算測(cè)試通過(guò)")
def test_versions():
    """版本兼容性檢查"""
    print("\n" + "=" * 50)
    print("版本兼容性檢查:")
    # 檢查基本兼容性
    required_versions = {
        'torch': '2.0.0',
        'torchvision': '0.15.0',
        'torchaudio': '2.0.0'
    }
    current_versions = {
        'torch': torch.__version__,
        'torchvision': torchvision.__version__,
        'torchaudio': torchaudio.__version__
    }
    for package, required in required_versions.items():
        current = current_versions[package]
        if current >= required:
            print(f"? {package}: {current} (>= {required})")
        else:
            print(f"??  {package}: {current} (< {required})")
if __name__ == "__main__":
    test_basic()
    test_cuda()
    test_mps()
    test_cpu()
    test_versions()
    print("\n" + "=" * 50)
    print("環(huán)境驗(yàn)證完成!")

conda activate pytorch
python test_pytorch.py

# OpenCV
conda install opencv -c conda-forge

# PIL/Pillow
conda install pillow -c conda-forge

# Albumentations (數(shù)據(jù)增強(qiáng))
pip install albumentations

# Detectron2 (Facebook檢測(cè)框架)
pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu118/torch2.0/index.html

# Transformers (Hugging Face)
pip install transformers

# Tokenizers
pip install tokenizers

# spaCy
conda install spacy -c conda-forge
python -m spacy download en_core_web_sm

# NLTK
conda install nltk -c conda-forge

# Jupyter Lab
conda install jupyterlab -c conda-forge

# Matplotlib (可視化)
conda install matplotlib -c conda-forge

# Seaborn (統(tǒng)計(jì)可視化)
conda install seaborn -c conda-forge

# Scikit-learn (機(jī)器學(xué)習(xí))
conda install scikit-learn -c conda-forge

# Pandas (數(shù)據(jù)處理)
conda install pandas -c conda-forge

# Lightning (簡(jiǎn)化訓(xùn)練流程)
pip install pytorch-lightning

# Ignite (訓(xùn)練循環(huán))
pip install pytorch-ignite

# Captum (可解釋性)
pip install captum

# 在PyTorch環(huán)境中安裝
conda activate pytorch
conda install jupyter notebook jupyterlab -c conda-forge -y

# 安裝ipykernel
conda install ipykernel -y

# 創(chuàng)建內(nèi)核
python -m ipykernel install --user --name pytorch --display-name "Python (PyTorch)"

# 啟動(dòng)Jupyter Lab（推薦）
jupyter lab

# 或啟動(dòng)經(jīng)典N(xiāo)otebook
jupyter notebook

import torch
print(f"PyTorch版本: {torch.__version__}")
print(f"設(shè)備: {torch.device('cuda' if torch.cuda.is_available() else 'cpu')}")

# 配置Jupyter默認(rèn)目錄
jupyter notebook --generate-config
# 編輯 ~/.jupyter/jupyter_notebook_config.py
# c.NotebookApp.notebook_dir = '/path/to/your/notebooks'

# 設(shè)置密碼
jupyter notebook password

# 激活環(huán)境
conda activate pytorch

# 導(dǎo)出完整環(huán)境（包含構(gòu)建信息）
conda env export > pytorch_environment.yml

# 導(dǎo)出僅包名和版本（更便攜）
conda env export --no-builds > pytorch_environment_no_builds.yml

# 導(dǎo)出pip風(fēng)格的requirements.txt
pip freeze > requirements.txt

name: pytorch
channels:
  - pytorch
  - nvidia
  - conda-forge
  - defaults
dependencies:
  - python=3.11.8
  - pytorch=2.2.0
  - torchvision=0.17.0
  - torchaudio=2.2.0
  - pytorch-cuda=12.1
  - jupyterlab=4.0.11
  - matplotlib=3.8.3
  - pandas=2.2.1
  - pip
  - pip:
    - transformers==4.38.2
    - albumentations==1.4.3
prefix: /home/user/anaconda3/envs/pytorch

# 方法1: 使用environment.yml
conda env create -f pytorch_environment.yml

# 方法2: 手動(dòng)創(chuàng)建并安裝
conda create -n pytorch python=3.11 -y
conda activate pytorch
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia -y
pip install -r requirements.txt

# 重新安裝，確保版本匹配
conda activate pytorch
conda uninstall pytorch torchvision torchaudio
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia -y

# 檢查驅(qū)動(dòng)和CUDA版本
nvidia-smi

# 根據(jù)輸出選擇正確的PyTorch CUDA版本
# 例如: nvidia-smi顯示CUDA 12.2 → 安裝CUDA 12.1的PyTorch（向下兼容）

# 在代碼開(kāi)頭添加內(nèi)存清理
import torch
torch.cuda.empty_cache()
# 或減少batch size
# 或使用混合精度訓(xùn)練
from torch.cuda.amp import autocast
with autocast():
    # your model forward pass

# 添加清華鏡像源
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/

# 重新安裝ipykernel
conda activate pytorch
pip install --force-reinstall ipykernel
python -m ipykernel install --user --name pytorch --display-name "Python (PyTorch)"

# 確保使用MPS設(shè)備
device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
# 對(duì)于某些操作，CPU可能更快
# 可以比較MPS和CPU性能
import time
# MPS測(cè)試
start = time.time()
x = torch.randn(10000, 10000, device='mps')
y = torch.randn(10000, 10000, device='mps')
z = torch.mm(x, y)
mps_time = time.time() - start
# CPU測(cè)試
start = time.time()
x = torch.randn(10000, 10000)
y = torch.randn(10000, 10000)
z = torch.mm(x, y)
cpu_time = time.time() - start
print(f"MPS時(shí)間: {mps_time:.4f}s, CPU時(shí)間: {cpu_time:.4f}s")

# 使用float32而不是float64（默認(rèn)）
tensor = torch.randn(1000, 1000, dtype=torch.float32)

# 對(duì)于整數(shù)標(biāo)簽，使用int64
labels = torch.randint(0, 10, (1000,), dtype=torch.long)

# 在訓(xùn)練開(kāi)始前添加
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.deterministic = False

from torch.utils.data import DataLoader
dataloader = DataLoader(
    dataset,
    batch_size=32,
    shuffle=True,
    num_workers=4,  # 根據(jù)CPU核心數(shù)調(diào)整
    pin_memory=True,  # GPU訓(xùn)練時(shí)啟用
    prefetch_factor=2  # 預(yù)取數(shù)據(jù)
)

from torch.cuda.amp import GradScaler, autocast
scaler = GradScaler()
for data, target in dataloader:
    optimizer.zero_grad()
    with autocast():
        output = model(data)
        loss = criterion(output, target)
    scaler.scale(loss).backward()
    scaler.step(optimizer)
    scaler.update()

# 及時(shí)刪除不需要的變量
del large_tensor
torch.cuda.empty_cache()
# 使用上下文管理器
with torch.no_grad():
    # 推理時(shí)禁用梯度計(jì)算
    output = model(input)

# 在~/.bashrc中添加
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
export NUMEXPR_NUM_THREADS=1

# 減少多線(xiàn)程沖突

# 創(chuàng)建環(huán)境
conda create -n pytorch python=3.11 -y

# 激活/停用環(huán)境
conda activate pytorch
conda deactivate

# 刪除環(huán)境
conda env remove -n pytorch

# 列出所有環(huán)境
conda env list

# CPU版本
conda install pytorch torchvision torchaudio cpuonly -c pytorch -y

# CUDA 12.1版本
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia -y

# CUDA 11.8版本  
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia -y

# 導(dǎo)出環(huán)境
conda env export > environment.yml

# 重建環(huán)境
conda env create -f environment.yml

# 快速驗(yàn)證
python -c "import torch; print(torch.__version__, torch.cuda.is_available())"