Deep Learning Environment Setup and Project Configuration

Version Checking


# Check CUDA version (Command Prompt)
nvcc -V   or   nvcc --version

# Check Python version (Command Prompt)
python

# Check available CUDA versions (Command Prompt)
nvidia-smi  # CUDA Version is displayed after this text

Installation Process

1. Visual Studio Installation

Version Selection: For CUDA 11.8.0 (can be higher than system requirements), visit the CUDA Toolkit Archive at https://developer.nvidia.com/cuda-toolkit-archive to find compatible Visual Studio versions.
Download Visual Studio: Get the 2022 Community edition from https://visualstudio.microsoft.com/zh-hans/downloads/
Installation: After launching the installer, wait for initial loading to complete.
Path Configuration: Modify installation paths for the three main components to a non-system drive (e.g., D:) to save space.
Component Selection: Choose necessary components - for deep learning work, select "Desktop development with C++" and "Python developmant".
Installation: Click the install button to begin the process.

2. CUDA Installation

Download: Get the appropriate version from https://developer.nvidia.com/cuda-toolkit-archive
Installation: Run the downloaded .exe file. The temporary file path can be modified but shouldn't match the final CUDA installation path.
Environment Configuraton: Verify system variables are automatically created (NV11_0 will vary based on installed version). If missing, manually add:
- NVCUDASAMPLES_ROOT
- NVCUDASAMPLES11_0_ROOT
Verification: In Command Prompt, use: ``` nvcc --version
set cuda
```
 to check CUDA version and environment variables.
```

3. cuDNN Installation

Download: Get the compatible version from https://developer.nvidia.com/rdp/cudnn-archive
Installation: Extract the downloaded archive and copy all folders to the CUDA installation directory (default: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8)
Environment Configuration: Add these paths to the system PATH variable: ``` C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\include C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\lib C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\libnvvp
Verification: In Command Prompt, navigate to extras\demo_suite and run: ``` bandwidthTest.exe
deviceQuery.exe

4. Conda Environment Setup

# List environments
conda env list

# Create environment (name: MC, Python: 3.8)
conda create --name MC python=3.8

# Create environment with custom path (name: MC, path: D:/env, Python: 3.8)
conda create --name MC --prefix D:/env python=3.8

# Activate environment
conda activate MC
source conda activate MC  # For Linux

# Remove environment
conda remove --name MC --all

5. PyCharm Installation

Download: Get the Community edition from https://www.jetbrains.com/pycharm/download/?section=windows
Installation: Run the .exe file
Path Configuration: Can use default path
Component Selection: Generally select all components

6. Python Installation

Download: From https://www.python.org/downloads/, select appropriate version (not latest recommended) and system architecture
Installation : Run the installer. The "Add Path" option can be checked during first setup 4. Path Configuration : Remember installation directory for environment variables 6. Verification : In Command Prompt, check with: ``` python

pip


10. **Environment Variables**
: Add Python installation path to system variables 12. **Final Check**
: Restart computer and verify Python and pip functionality 
### 7. PyTorch Installation

1. **Download**
: From https://pytorch.org/get-started/previous-versions/, select appropriate version 3. **Installation**
: Use the following command in PyCharm terminal: ```
# CUDA 11.8
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

Troubleshooting : If download fails, configure alternative mirrors or download .whl files for offline installation Package Management

Mirror Configuration

# List current mirrors
pip config list

# Configure mirrors
# Configure Tsinghua mirror
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
# Trust the mirror
pip config set install.trusted-host pypi.tuna.tsinghua.edu.cn

# Common mirrors
# Tsinghua: https://pypi.tuna.tsinghua.edu.cn/simple
# Douban: https://pypi.douban.com/simple/
# Aliyun: https://mirrors.aliyun.com/pypi/simple/

Pip Commands

# List installed packages
pip list

# Install package
pip install package_name

# Install specific version
pip install package_name==1.1.1

# Uninstall package
pip uninstall package_name

# Update package
pip install --upgrade package_name

# Show package information
pip show package_name

# Install local .whl file
cd D:\package
pip install package_name.whl

Project Setup and Execution

Dataset Organization

RTTS Dataset Structure


\---RTTS
    +---images
    |   +---train
            +---image_001.png
            \---image_002.png
    |   \---val
    \---labels
        +---train
            +---image_001.txt
            \---image_002.txt
        \---val
        \---rtts.json
        \---rtts_100_val.json

DINO Framework Configuration

Library Installation

panopticapi

git clone https://github.com/cocodataset/panopticapi.git
cd .\panopticapi\
python setup.py build_ext --inplace  
python setup.py build_ext install

MultiScaleDeformableAttention

cd .\models\dino\ops\
python setup.py build install

If installation fails with "unsupported Microsoft Visual Studio version", modify:

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\include\crt\host_config.h

Change the number after "_MSC_VER >=" to 2000, then retry installation.

ResNet50 Backbone Modifications

Dataset Configuration

Edit DINO\datasets\coco.py (around line 615):

# PATHS = {
#     "train": (root / "train2017", root / "annotations" / f'{mode}_train2017.json'),
#     "train_reg": (root / "train2017", root / "annotations" / f'{mode}_train2017.json'),
#     "val": (root / "val2017", root / "annotations" / f'{mode}_val2017.json'),
#     "eval_debug": (root / "val2017", root / "annotations" / f'{mode}_val2017.json'),
#     "test": (root / "test2017", root / "annotations" / 'image_info_test-dev2017.json' ),
# }
# modified paths
PATHS = {
    "train": (root / "images/train", root / "labels/rtts.json"),
    "train_reg": (root / "images/train", root / "labels/rtts.json"),
    "val": (root / "images/val", root / "labels/rtts_100_val.json"),
    "eval_debug": (root / "images/val", root / "labels/rtts_100_val.json"),
    "test": (root / "images/val", root / "labels/rtts_100_val.json"),
};

Training Hyperparameters

Edit DINO\config\DINO\DINO_4scale_mc.py:

num_classes=5
batch_size = 2
epochs = 12

Training Script Modifications

Edit DINO\main.py:

parser.add_argument('--config_file', '-c', type=str, default='config/DINO/DINO_4scale_mc.py')
parser.add_argument('--output_dir', default='logs/DINO/R50-MS4', help='path where to save, empty for no saving')
parser.add_argument('--coco_path', type=str, default='D:/project/dataset/RTTS/')

Edit DINO\config\DINO\DINO_4scale.py:

embed_init_tgt = True
use_ema = False
dn_box_noise_scale = 1.0
dn_scalar=100
dn_label_coef=1.0
dn_bbox_coef=1.0

Swin Backbone Modifications

Dataset Configuration

Same as ResNet50 configuration above

Training Hyperparameters

Edit DINO\config\DINO\DINO_4scale_swin_mc.py:

num_classes=5
batch_size = 2
epochs = 12

Training Script Modifications

Edit DINO\main.py:

parser.add_argument('--config_file', '-c', type=str, default='config/DINO/DINO_4scale_swin_mc.py')

Common Issues and Solutions

Shared Folder Error

Error: "No shared folder available dino"

Solution: Create a folder named "comp_robot" in the project directory with an "experiments" subfolder. In run_with_submitit.py, modify the get_shared_folder() function to use your custom folder path.

Resume Training Issues

When resuming training after interruption, only setting the resume parameter can cause issues with checkpoint selection.

Solution: Specify the best checkpoint manually:

parser.add_argument('--resume', default='logs/DINO/checkpoint0010.pth', help='resume from checkpoint')
parser.add_argument('--start_epoch', default=11, type=int)

Library Version Conflicts

Error: "FormatCode() got an unexpected keyword argument 'verify'"

Solution: Install compatible versions:

pip install yapf==0.40.0
pip install mmcv==2.1.0
pip install mmengine==0.10.3
pip install mmdet==3.3.0

Encoding Issues

Error: "UnicodeDecodeError: 'gbk' codec can't decode byte"

Solution: Modify util/slconfig.py:

# with open(filename) as f:
with open(filename,encoding='UTF-8') as f

# with open(filename, 'r') as f:
with open(filename, 'r',encoding='UTF-8') as f

DINOv2 Framework

A self-supervised learning model that doesn't rely on manually labeled data, instead using the inherent structure of data to generate pseudo-labels for training.

TogetherNet Framework

Environment Setup

pip install h5py==2.10.0
pip install tensorboard

Dataset Structure


\---RTTS
    +---images
    |   +---train
            +---image_001.png
            \---image_002.png
    |   +---train_dehaze
            +---image_001_dehaze.png
            \---image_002_dehaze.png
    |   +---val
            +---0.png
            \---1.png
    |   +---val_dehaze
            +---0_MSBDN.png
            \---1_MSBDN.png
    \---labels
        +---train
            +---image_001.txt
            \---image_002.txt
        \---val
        \---rtts.json
        \---rtts_100_val.json

Configuration Files

Additional configuration files are needed: train.txt, train_dehaze.txt, val.txt, val_dehaze.txt

Code Modifications

train.py

classes_path    = 'model_data/rtts_classes.txt'
# model_path      = 'model_data/yolox_s.pth'

#batch and epoch settings
Freeze_batch_size   = 16
UnFreeze_Epoch      = 100
Unfreeze_batch_size = 16

# Dataset paths
train_annotation_path   = '2007_train_fog.txt'
val_annotation_path     = '2007_val_fog.txt'
clear_annotation_path = '2007_train.txt'
val_clear_annotation_path = '2007_val.txt'

utils/dataloader.py

def get_random_data(self, annotation_line, clearimage_line, input_shape, jitter=.3, hue=.1, sat=0.7, val=0.4, random=True):
        # Extract image names
        pic_name=line[0]
        pic_clear_name=clearline[0]
        
        # Construct image paths
        line=os.path.join("D:/project/dataset/RTTS/images/" ,pic_name+ ".png")
        clearline = os.path.join("D:/project/dataset/RTTS/images/", pic_clear_name + ".png")

        # Load images
        image   = Image.open(line)
        image   = cvtColor(image)

        clearimg = Image.open(clearline)
        clearimg = cvtColor(clearimg)

        # Load label data
        train_label_path=os.path.join("D:/project/dataset/RTTS/labels/" ,pic_name+ ".txt")
        with open(train_label_path, encoding='utf-8') as f:
            train_labels_lines = f.readlines()
        
        box     = np.array([np.array(list(map(float,box_element.split( )[1:]))) for box_element   in train_labels_lines])

nets/yolo.py

def forward(self, input):
    if self.training:
        # Split input into haze and clear images
        input, clear_x = input.split((4, 4), dim=0)

Training Process

Run train.py to start training. Results are stored in the logs folder. With 4300+ images and batch size 16, each epoch takes approximately 3 minutes.

RT-DETR-V2 Framework

Environment Setup

conda create --name rt python=3.8
conda install pytorch2.0.1 torchvision0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia

Dataset Notes

JSON class numbering should start from 1.

Tags: cuda pytorch deep-learning environment-setup dino

Posted on Mon, 08 Jun 2026 17:32:29 +0000 by warren

Freaks City

Deep Learning Environment Setup and Project Configuration

Version Checking

Installation Process

1. Visual Studio Installation

2. CUDA Installation

3. cuDNN Installation

4. Conda Environment Setup

5. PyCharm Installation

6. Python Installation

Mirror Configuration

Pip Commands

Project Setup and Execution

Dataset Organization

RTTS Dataset Structure

DINO Framework Configuration

Library Installation

panopticapi

MultiScaleDeformableAttention

ResNet50 Backbone Modifications

Dataset Configuration

Training Hyperparameters

Training Script Modifications

Swin Backbone Modifications

Dataset Configuration

Training Hyperparameters

Training Script Modifications

Common Issues and Solutions

Shared Folder Error

Resume Training Issues

Library Version Conflicts

Encoding Issues

DINOv2 Framework

TogetherNet Framework

Environment Setup

Dataset Structure

Configuration Files

Code Modifications

train.py

utils/dataloader.py

nets/yolo.py

Training Process

RT-DETR-V2 Framework

Environment Setup

Dataset Notes

Hot Tags