YOLO Local Testing and Server Training

This guide separates the workflow into two parts: local validation and formal server-side training. That split keeps the process stable and makes debugging much easier.

Recommended Workflow

1. Verify the local environment.
2. Confirm that CUDA and PyTorch work.
3. Test the training command on a small scale.
4. Move the full run to a server.
5. Monitor logs with TensorBoard.
6. Validate, predict, and export the final results.

What To Do Locally

• Check GPU visibility with nvidia-smi.
• Create and activate a clean environment.
• Install PyTorch, Ultralytics, and TensorBoard.
• Run a short test job before using a long training schedule.

What To Do On The Server

• Recreate the environment cleanly.
• Pull or upload the project.
• Confirm dataset paths and model weights.
• Run the full training command with the intended batch, imgsz, and epochs.

Why This Split Works

Local testing catches environment and command issues early. Server training keeps the heavy GPU workload in the right place. This avoids mixing quick debugging work with long-running production training.

Final Checks

• Run validation on the trained weights.
• Use prediction mode on sample images.
• Copy the runs/ directory back to your local machine for review and archiving.