LoRA Fine Tuning

LoRA Fine-Tuning

FlashTrack supports 6 LoRA variants for parameter-efficient fine-tuning.

Available Variants

Variant	Description	Best For
standard	Classic LoRA (Hu et al., 2022)	General fine-tuning
dora	Weight-decomposed LoRA (Liu et al., 2024)	Higher quality
lora_plus	Asymmetric LR for A/B matrices	Faster convergence
adalora	Adaptive rank via SVD pruning	Automatic rank
ortho	Orthogonal regularization	Stable training
lora_fa	Frozen A, trainable B only	Minimal memory

Usage

Python API

from flashtrack import Trainer

trainer = Trainer(
    model_size="m",
    lora=True,
    lora_rank=8,
    lora_variant="standard",
    train_data="data/MOT17/train",
)
trainer.train()

Direct Application

from flashtrack.models.tracker import FlashTracker
from flashtrack.models.lora import apply_lora

model = FlashTracker(backbone_size="1.0x", reid_dim=128, encoder_channels=256)
model = apply_lora(model, rank=8, variant="dora", target_modules=["backbone", "encoder"])

QLoRA (Quantized LoRA)

trainer = Trainer(
    qlora=True,
    qlora_dtype="int8",
    lora_rank=8,
    ...
)

YAML Config

train:
  use_lora: true
  lora_rank: 8
  lora_alpha: 16.0
  lora_dropout: 0.05
  lora_target_modules: ["backbone", "encoder"]

Merging LoRA Weights

After training, merge LoRA into base weights for zero-overhead inference:

from flashtrack.models.lora import merge_lora_weights
model = merge_lora_weights(model)