🏋️ 1.1 Install basic system packages

sudo apt update && sudo apt upgrade -y
sudo apt install -y build-essential git wget curl unzip zip python3 python3-pip python3-venv libglib2.0-0

These are needed for Python, OpenCV, etc.

1.2 Install NVIDIA Driver

You confirmed you have:

Driver version: 575.64.03

CUDA version: 12.9

GPU: GeForce GTX 1060 3GB

✅ Everything looks good.

1.3 Install Conda

Download & install Miniconda:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
source ~/.bashrc

Restart terminal after install.

1.4 Create Python environment

conda create -n yolov8 python=3.10 -y
conda activate yolov8

1.5 Install PyTorch (CUDA 12.1 — compatible with your 12.9)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Test it:

python -c "import torch; print(torch.cuda.is_available())"

✅ It should print True.

1.6 Install YOLOv8

pip install ultralytics

Test CLI:

yolo help

1.7 Edit data.yaml

Use full absolute paths. Example:

train: /home/afraj/yolov8-dataset/train/images
val: /home/afraj/yolov8-dataset/val/images

nc: 3  # change this to your number of classes
names: ['class_0', 'class_1', 'class_2']  # replace with real class names

To get full paths:

realpath yolov8-dataset/train/images
realpath yolov8-dataset/val/images

1.8 Train YOLOv8-nano

Run this command (adjust batch if you get OOM):

yolo task=detect mode=train model=yolov8n.pt \
  data=/home/afraj/yolov8-dataset/data.yaml \
  epochs=50 imgsz=640 batch=4 name=yolov8n_run

Training outputs saved to:

📁 runs/detect/yolov8n_run/   ↳ best.pt ← this is the final model

✅ Training complete

📤 2. Export to TFLite (for Android)

First try direct export:

yolo export model=runs/detect/yolov8n_run/weights/best.pt format=tflite

It will produce:

📄 best.tflite

🚨 3. Fallback Export (if direct TFLite fails)

3.1 Export to ONNX

yolo export model=runs/detect/yolov8n_run/weights/best.pt format=onnx

3.2 Convert ONNX → TensorFlow SavedModel

Install converter:

pip install onnx2tf

Run:

onnx2tf -i best.onnx -o saved_model

3.3 Convert to TFLite

import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_saved_model("saved_model")
converter.optimizations = [tf.lite.Optimize.DEFAULT]  # optional
tflite_model = converter.convert()

with open("model.tflite", "wb") as f:
    f.write(tflite_model)

Now you have model.tflite 🎉

✅ Step-by-Step Fix for Conda

✅ Step 1: Accept the Terms of Service

Run these two commands:

conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r

This tells Conda: “Yes, I accept the license for the default package sources.”