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 Turing / Ampere / Ada_Lovelace 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	51/35 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

12.1.0 8.9.0 8.6.1.6 2.1.2
Extract Apex-CV-YOLO-v8-Aim-Assist-Bot-main.zip to C:\TEMP\Ape-xCV
Install Visual Studio 2019 Build Tools.
- Download from: OneDrive or 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 12.1.0 from: NVIDIA website.
- ✅ I understand, and wish to continue the installation regardless.
Install cuDNN 8.9.0.
- Register for the NVIDIA developer program.
  - Go to the cuDNN download site:cuDNN download archive.
  - Click Download cuDNN v8.9.0 (April 11th, 2023), for CUDA 12.x.
  - Download Local Installer for Windows (Zip).
- Unzip cudnn-windows-x86_64-8.9.0.131_cuda12-archive.zip.
- Copy all three folders (bin,include,lib) and paste <overwrite> them into C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.1
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 python requirements.

CUDA	cuDNN	TensorRT	PyTorch
12.1.0	8.9.0	8.6.1.6	2.1.2

cd /D C:\TEMP\Ape-xCV
pip install numpy==1.23.1
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
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. Change your mouse DPI instead.
SHIFT
- Hold to lock on target.
LEFT_LOCK
- Enabled when pressing '1' or '2'. Disabled when pressing 'G'.
- Use to lock on target while firing automatic weapons.
RIGHT_LOCK
- You need to change your ADS from toggle to hold.
- Use CURSOR_RIGHT to toggle lock on target while scoping.
AUTO_FIRE
- Use CURSOR_UP to toggle auto fire non-automatic weapons while locking on target.
HOME
- 💀 Terminate script.
Load the Firing Range and give this script a go!
- 🐵 Run Ape-xCV.bat.

2.2 Best pratices

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

3.1 TensorRT (.engine)

Install TensorRT.
- Go to the TensorRT download site: NVIDIA TensorRT 8.x Download.
- Download TensorRT 8.6 GA for Windows 10 and CUDA 12.0 and 12.1 ZIP Package from: NVIDIA website.
- Extract TensorRT-8.6.1.6.Windows10.x86_64.cuda-12.0.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-8.6.1.6\lib
```
- TensorRT was added to PATH. Close that Command Prompt and open a new one. Then input:
```
cd /D C:\TEMP\TensorRT-8.6.1.6\python
pip install tensorrt-8.6.1-cp310-none-win_amd64.whl
```

To export best_8s.pt to best_8s.engine:

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=engine opset=12 workspace=7

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.engine
Save args_.py.
- 🐵 Run Ape-xCV.bat.

3.2 TensorRT (.trt)

If C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.1\bin\zlibwapi.dll is missing.
- Copy C:\Program Files\NVIDIA Corporation\Nsight Systems 2023.1.2\host-windows-x64\zlib.dll to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.1\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-8.6.1.6\bin\trtexec.exe --onnx=best_8s.onnx --saveEngine=best_8s.trt --buildOnly --workspace=7168 --fp16

If RTX 30 Series (FP32):

C:\TEMP\TensorRT-8.6.1.6\bin\trtexec.exe --onnx=best_8s.onnx --saveEngine=best_8s.trt --buildOnly --workspace=7168

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")

In --model change best_8s.pt to best_8s.trt
Save args_.py.
- 🐵 Run Ape-xCV.bat.

3.3 TensorRT with GPU NMS (.trt)

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.pt", help="model path")
    args.add_argument("--end2end", type=bool,
                    default=False, help="use TensorRT efficientNMSPlugin")

In --model change best_8s.pt 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.

Ape-xCV/Apex-CV-YOLO-v8-Aim-Assist-Bot

Apex CV YOLO v8 Aim Assist Bot

Introduction

Features and Requirements

Benchmarks

0. Disclaimer

1. Environment set up in Windows

2.1 Usage

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

2.2 Best pratices

3.1 TensorRT (.engine)

3.2 TensorRT (.trt)

3.3 TensorRT with GPU NMS (.trt)

4. args_.py

❤️ Sponsor

5. Train your own model