Training

Training

Dataset Format

FlashTrack expects MOT17/MOT20-format datasets:

data/MOT17/train/
  MOT17-02-DPM/
    img1/
      000001.jpg
      000002.jpg
      ...
    gt/
      gt.txt
  MOT17-04-DPM/
    ...

Each line in gt.txt:

frame, id, bb_left, bb_top, bb_width, bb_height, conf, class, visibility

Basic Training

from flashtrack import Trainer

trainer = Trainer(
    model_size="m",
    epochs=120,
    batch_size=64,
    lr=0.0003,
    train_data="data/MOT17/train",
    val_data="data/MOT17/val",
    amp=True,
    save_dir="workspace/reid_training",
)
results = trainer.train()

YAML Config Training

flashtrack train --config configs/flashtrack_m.yaml

Training Outputs

workspace/reid_training/
  checkpoint_best.pth        # Best checkpoint (full state)
  checkpoint_last.pth        # Latest checkpoint
  model_best_inference.pth   # Best model (inference weights)
  model_final_inference.pth  # Final model (FP32)
  model_final_fp16.pth       # Final model (FP16)
  train_*.log                # Training log

Loss Functions

• Triplet loss — Hard positive/negative mining for embedding learning
• Classification loss — Cross-entropy for identity prediction
• Combined: total = triplet_weight * triplet + cls_weight * classification

Mixed Precision

Enable AMP for faster training with lower memory:

trainer = Trainer(amp=True, ...)

Resume Training

trainer = Trainer(resume="workspace/checkpoint_last.pth", ...)