Training¶

Training in objdet is handled via the CLI, which acts as a wrapper around LightningCLI.

The `fit` Command¶

The primary command for training is objdet fit.

objdet fit --config configs/experiment/faster_rcnn_coco.yaml

Configuration Structure¶

Our configurations follow a composed structure:

Model: Defines architecture, optimizer, and scheduler.
Data: Defines dataset, batch size, and transforms.
Trainer: Defines PyTorch Lightning trainer flags (epochs, GPUs, callbacks).

Experiment Configs¶

Experiment configs in configs/experiment combine these sections for reproducible runs.

Example faster_rcnn_coco.yaml:

# @package _global_

defaults:
  - /model: faster_rcnn
  - /data: coco
  - /trainer: default

trainer:
  max_epochs: 12
  accelerator: gpu
  devices: 1

data:
  batch_size: 4

Running Experiments¶

To run an experiment:

# Basic run
objdet fit --config configs/experiment/yolov8_coco.yaml

# Override parameters
objdet fit --config configs/experiment/yolov8_coco.yaml \
    --trainer.max_epochs 50 \
    --data.batch_size 16 \
    --model.init_args.learning_rate 0.005

Multi-GPU Training¶

objdet supports distributed training out of the box.

# Train on 2 GPUs using DDP
objdet fit --config configs/experiment/faster_rcnn_coco.yaml \
    --trainer.devices 2 \
    --trainer.strategy ddp

Resume Training¶

To resume from a checkpoint:

objdet fit --config configs/experiment/faster_rcnn_coco.yaml \
    --ckpt_path training_logs/version_0/checkpoints/last.ckpt