Skip to content

Latest commit

 

History

History
97 lines (61 loc) · 4.25 KB

TRAINING.md

File metadata and controls

97 lines (61 loc) · 4.25 KB

Training ALDI

Training ALDI involves two steps:

  1. Baseline training / burn-in
  2. Domain adaptive training

This assumes you have:

1. Baseline training / burn-in

The first step is to train a standard object detector on your source dataset (AKA "source-only training"). This has two purposes:

  • Baseline. For evaluation, this model tells you the best you can do before domain adaptation, using only source-domain data.
  • Burn-in. Source-only training is also an effective method for network initialization, AKA "burn-in", before performing domain adaptation. See our paper for more details.

Set up configuration file for baseline training / burn-in

You will need a custom Detectron2 configuration file for every training run. We recommend you start with configs/cityscapes/Base-RCNN-FPN-Cityscapes_strongaug_ema.yaml as a guide for performing baseline training / burn-in. Note that compared to standard Detectron2 training configurations, this config file uses:

  • Stronger augmentations
  • Exponential moving average (EMA)

We have found these settings to both increase overall performance (in terms of mAP) as well as reduce training time during the domain adaptation step. See our paper for more details.

Modifications you must make To use with your dataset, make a copy of configs/cityscapes/Base-RCNN-FPN-Cityscapes_strongaug_ema.yaml and change the following values:

DATASETS:
  TRAIN: ("your_training_dataset_name",) # needs to be a tuple, and can contain multiple datasets if you want
  TEST: ("your_test_dataset_name",)  # needs to be a tuple, and can contain multiple datasets if you want

Run training

To train, run:

python tools/train_net.py --config path/to/your/burn_in_config.yaml

Other training options include --num-gpus to run distributed training; see tools/train_net.py and the Detectron2 training docs for more details.

Where is your final model?

A BestCheckpointer will be used by default to save the best model checkpoint based on validation performance on each DATASETS.TEST; this model will be saved according to the OUTPUT_DIR in your config file, and will end in _best.pth.

2. Domain adaptive training

Now you're ready to use ALDI for domain adaptation. Again this involves creating a configuration file and running tools/train_net.py.

Configuration file setup

See configs/cityscapes/ALDI-Best-Cityscapes.yaml for an example. You must change two things: the dataset and the path to initial model weights.

Dataset To use with a different dataset, simply change DATASETS.UNLABELED to point to your unsupervised target-domain data (like DATASETS.TRAIN AND DATASETS.TEST, this must also be registered with Detectron2 as described in the dataset setup guide):

DATASETS:
  UNLABELED: ("your_unlabeled_dataset_name",)

Model weights To use your burned-in model from step 1, change:

MODEL:
  WEIGHTS: "path/to/your/burned_in_model.pth"

Training

To train, run:

python tools/train_net.py --config path/to/your/aldi_config.yaml

Domain adaptive training is also compatible with standard Detectron2 training options such as --num-gpus.

Evaluation

After training, to run evaluation with your model:

python tools/train_net.py --eval-only --config-file path/to/your/aldi_config.yaml MODEL.WEIGHTS path/to/your/model_best.pth

Feature visualization

To visualize a trained model's feature space, run:

python tools/visualize_featurespace.py --config-file path/to/your/aldi_config.yaml MODEL.WEIGHTS path/to/your/model.pth

The plots will be written into that config's OUTPUT_DIR.