Apex CV YOLO v8 Aim Assist Bot

• Due to the widespread adoption of controller aim assist on PC gaming...

Introduction

• It is the aim of this project to provide a quality object detection model.
• Object detection is a technique used in computer vision for the identification and localization of objects within an image or a video.
• This project is based on: Franklin-Zhang0/Yolo-v8-Apex-Aim-assist.
• It is safe to use if you follow the best practices.
• This project is regularly maintained. Better models are made available as more training is created. It is a slowly but surely endeavour.

Features and Requirements

Features:

• Faster screen capture with dxshot
• Faster CPU NMS (Non Maximum Supression) with NumPy
• Optional GPU NMS (Non Maximum Supression) with TensorRT efficientNMSPlugin
• Class targeting Ally, Enemy, Tag
• Humanized mouse control with PID (Proportional-Integral-Derivative)

Requirements:

• NVIDIA RTX Series GPU

Benchmarks

Test system:

- OS: Windows 10 Enterprise 1803 (OS build 17134)
- CPU: Intel Core i7 3770K @ 4.0 GHz
- GPU: NVIDIA GeForce RTX 2070_8G / 3080_12G
- RAM: 16G DDR3 @ 2133 MHz
- Monitor resolution: 1920 x 1080
- In-game resolution: 1920 x 1080
- In-game FPS: RTSS async locked @ 72 FPS

GPU	imgsz	apex_8s.pt	apex_8s.trt	Precision
RTX 2070_8G	640/ 1080p	51/33 FPS	72/50 FPS	FP16/32
RTX 3080_12G	640/ 1080p	53 FPS	72 FPS	FP32

Video settings:

- Aspect Ratio               16:9
- Resolution                 1920 x 1080
- Brightness                 50%
- Field of View (FOV)        90
- FOV Ability Scaling        Enabled
- Sprint View Shake          Normal
- V-Sync                     Disabled
- NVidia Reflex              Enabled+Boost
- Adaptive Resolution FPS    60
- Adaptive Supersampling     Disabled
- Anti-aliasing              TSAA
- Texture Streaming Budget   Ultra (8GB VRAM)
- Texture Filtering          Anisotropic 16X
- Ambient Occlusion Quality  High
- Sun Shadow Coverage        Low
- Sun Shadow Detail          Low
- Spot Shadow Detail         Low
- Volumetric Lighting        Enabled
- Dynamic Spot Shadows       Enabled
- Model Detail               High
- Effects Detail             High
- Impact Marks               Disabled
- Ragdolls                   Low

0. Disclaimer

• This guide has been tested twice, and each time on a fresh install of Windows.
  • Every detail matters. If you are having issues, you are not following the guide.

1. Environment set up in Windows

• Version checklist:

  CUDA	cuDNN	TensorRT	PyTorch
  11.8.0	8.6.0	8.5.3.1	2.0.1

• Extract Apex-CV-YOLO-v8-Aim-Assist-Bot-main.zip to C:\TEMP\Ape-xCV

• Install Visual Studio 2019 Build Tools.

  • Download from: Microsoft website.
  • On Individual components tab:
    • ✅ MSVC v142 - VS 2019 C++ x64/x86 build tools (Latest)
    • ✅ C++ CMake tools for Windows
    • ✅ Windows 10 SDK (10.0.19041.0)
  • ➡️ Install

• Install CUDA 11.8.0 from: NVIDIA website.

  • The minimum Visual Studio components were installed in previous step.
    • ✅ I understand, and wish to continue the installation regardless.

• Install cuDNN 8.6.0.

  • Download from: OneDrive.
  • OR
  • Register for the NVIDIA developer program.
    • Go to the cuDNN download site:cuDNN download archive.
    • Click Download cuDNN v8.6.0 (October 3rd, 2022), for CUDA 11.x.
    • Download Local Installer for Windows (Zip).
  • Unzip cudnn-windows-x86_64-8.6.0.163_cuda11-archive.zip.
  • Copy all three folders (bin,include,lib) and paste <overwrite> them into C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8

• Install Python 3.10.0 (64-bit) from: Python website.

  • ❌ Install launcher for all users
  • ✅ Add Python 3.10 to PATH
  • ➡️ Install Now
    • 🔰 Disable path length limit

• Install TensorRT.

  • Go to the TensorRT download site: NVIDIA TensorRT 8.x Download.
  • Download TensorRT 8.5 GA Update 2 for Windows 10 and CUDA 11.0, 11.1, 11.2, 11.3, 11.4, 11.5, 11.6, 11.7 and 11.8 ZIP Package from: NVIDIA website.
  • Extract TensorRT-8.5.3.1.Windows10.x86_64.cuda-11.8.cudnn8.6.zip to C:\TEMP
  • Press [Win+R] and enter cmd to open a Command Prompt. Then input:

  cd /D C:\TEMP\Ape-xCV
  addenv C:\TEMP\TensorRT-8.5.3.1\lib

  • TensorRT was added to PATH. Close that Command Prompt and open a new one. Then input:

  cd /D C:\TEMP\TensorRT-8.5.3.1\python
  pip install tensorrt-8.5.3.1-cp310-none-win_amd64.whl

• Install python requirements.

cd /D C:\TEMP\Ape-xCV
pip install numpy==1.23.1
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

2.1 Usage

• Lock your in-game FPS. Give your GPU some slack. It is needed for object detection.

  • Do not use V-Sync to lock your FPS. V-Sync introduces input lag.
    • Use NVIDIA Control Panel.
    • OR
    • Use RTSS.

• Set in-game mouse sensitivity to 3.50.

  • The PID control (Kp, Ki, Kd) values in args_.py come already fine-tuned.
  • If the mouse moves too fast, EAC will flag your account and you will be banned on the next ban wave.
    • So, don't mess with the PID. Set in-game mouse sensitivity to 3.50. Change your mouse DPI instead.

• SHIFT

  • Hold to aim and fire automatically non-automatic weapons such as: Hemlok, Nemesis, Prowler, G7, 3xTake, 30-30, Mastiff, P2020, Wingman. Do not use with automatic weapons.

• LEFT_LOCK

  • Enabled when pressing '1' or '2'. Disabled when pressing 'G'.
  • OR
  • Use CURSOR_LEFT to toggle aim assist while firing automatic weapons.

• RIGHT_LOCK

  • This is not recommended, and you also need to change your ADS from toggle to hold.
  • Use CURSOR_RIGHT to toggle aim and fire while scoping. You will have to quickly follow up with MOUSE1 when using automatic weapons, or else your firing pattern will be a dead giveaway.

• HOME

  • 💀 Terminate script.

• Load the Firing Range and give this script a go!

  • 🐵 Run Ape-xCV.bat.

2.2 Best pratices

• To summarize:
  • ✅ Set in-game mouse sensitivity to 3.50. Use default PID.
  • ❌ No SHIFT with automatic weapons.
  • ❌ No RIGHT_LOCK.

3.1 TensorRT

• If C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin\zlibwapi.dll is missing.

  • Copy C:\Program Files\NVIDIA Corporation\Nsight Systems 2022.4.2\host-windows-x64\zlib.dll to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin and then rename it to zlibwapi.dll

• To export best_8s.pt to best_8s.trt:

  • Press [Win+R] and enter cmd to open a Command Prompt. Then input:

  set CUDA_MODULE_LOADING=LAZY
  cd /D C:\TEMP\Ape-xCV\MODEL
  yolo export model=best_8s.pt format=onnx opset=12

  • If RTX 20 Series (FP16):

  C:\TEMP\TensorRT-8.5.3.1\bin\trtexec.exe --onnx=best_8s.onnx --saveEngine=best_8s.trt --buildOnly --workspace=7168 --fp16

  • If RTX 30 Series (FP32):

  C:\TEMP\TensorRT-8.5.3.1\bin\trtexec.exe --onnx=best_8s.onnx --saveEngine=best_8s.trt --buildOnly --workspace=7168

• Install Notepad++ from: Notepad++ website.

• Open C:\TEMP\Ape-xCV\args_.py with Notepad++.

def arg_init(args):
    ...
    args.add_argument("--model", type=str,
                    default="/best_8s.pt", help="model path")

• Do not change the identation! In --model change best_8s.pt to best_8s.trt
• Save args_.py.
  • 🐵 Run Ape-xCV.bat.

3.2 TensorRT with GPU NMS

• Cons:

  • ❌ No speed increase.
  • ❌ IoU (Intersection over Union) threshold is hardcoded into engine, ignoring args_.py.

• Download: Linaom1214/TensorRT-For-YOLO-Series.

• Extract TensorRT-For-YOLO-Series-main.zip to C:\TEMP

• Rename C:\TEMP\TensorRT-For-YOLO-Series-main to C:\TEMP\Linaom1214

• To export best_8s.onnx to best_8s_e2e.trt:

  • Press [Win+R] and enter cmd to open a Command Prompt. Then input:

  set CUDA_MODULE_LOADING=LAZY
  cd /D C:\TEMP\Linaom1214

  • If RTX 20 Series (FP16):

  python export.py -o C:/TEMP/Ape-xCV/MODEL/best_8s.onnx -e C:/TEMP/Ape-xCV/MODEL/best_8s_e2e.trt -p fp16 -w 7 --end2end --conf_thres 0.6 --iou_thres 0.8 --v8

  • If RTX 30 Series (FP32):

  python export.py -o C:/TEMP/Ape-xCV/MODEL/best_8s.onnx -e C:/TEMP/Ape-xCV/MODEL/best_8s_e2e.trt -p fp32 -w 7 --end2end --conf_thres 0.6 --iou_thres 0.8 --v8

• Open C:\TEMP\Ape-xCV\args_.py with Notepad++.

def arg_init(args):
    ...
    args.add_argument("--model", type=str,
                    default="/best_8s.trt", help="model path")
    args.add_argument("--end2end", type=bool,
                    default=False, help="use TensorRT efficientNMSPlugin")

• In --model change best_8s.trt to best_8s_e2e.trt
• In --end2end change False to True
• Save args_.py.
  • 🐵 Run Ape-xCV.bat.

4. args_.py

• Open C:\TEMP\Ape-xCV\args_.py with Notepad++.

def arg_init(args):
    ...
    args.add_argument("--classes", type=int,
                    default=[1,2], help="classes to be detected TensorRT(.trt); can be expanded but needs to be an array. "
                    "0 represents 'Ally', "
                    "1 represents 'Enemy', "
                    "2 represents 'Tag'... "
                    "Change default accordingly if your dataset changes")
    args.add_argument("--target_index", type=int,
                    default=1, help="class to be targeted PyTorch(.pt)")
    args.add_argument("--half", type=bool,
                    default=True, help="use FP16 to predict PyTorch(.pt)")
    args.add_argument("--iou", type=float,
                    default=0.8, help="predict intersection over union")  # 0.8 is recommended
    args.add_argument("--conf", type=float,
                    default=0.6, help="predict confidence")  # 0.6+ is recommended
    screen_height = win32api.GetSystemMetrics(1)
    args.add_argument("--crop_size", type=float,
                    default=640/screen_height, help="the portion to detect from the screen. 1/3 for 1440P or 1/2 for 1080P, imgsz/screen_height=direct")
    args.add_argument("--wait", type=float, default=0, help="wait time")
    args.add_argument("--verbose", type=bool, default=False, help="predict verbose")
    args.add_argument("--draw_boxes", type=bool,
                    default=False, help="outline detected target, borderless window")

• Until you understand how NMS works, do not change --iou.
• --crop_size "the portion to detect from the screen". Will be scaled down to 640x640 for input. The default 640/screen_height is the best value.
• --draw_boxes "outline detected target, borderless window". Set to True in the Firing Range only.

❤️ Sponsor

PayPal	Litecoin (LTC)
	LKLYRUadyHitp5B7aB56yfBYSuG2UnuLEz

5. Train your own model

• Download starter dataset from: OneDrive.
• Extract apex.zip to C:\TEMP\Ape-xCV\datasets\apex
• Press [Win+R] and enter cmd to open a Command Prompt. Then input:

cd /D C:\TEMP\Ape-xCV
python train8s40.py

• This will train your YOLO v8 small model for 40 epochs with images and labels from C:\TEMP\Ape-xCV\datasets\apex and save it to C:\TEMP\Ape-xCV\runs\detect\train\weights\best.pt.

• You can add your own images (640x640); and create the labels with: developer0hye/Yolo_Label.

  • Copy those into C:\TEMP\Ape-xCV\SPLIT\input
  • Press [Win+R] and enter cmd to open a Command Prompt. Then input:

  cd /D C:\TEMP\Ape-xCV\SPLIT
  python split.py

  • Your images and labels are now split into train and valid.
  • Browse C:\TEMP\Ape-xCV\datasets\apex and distribute them accordingly.