Implementation for Task2Sim

Installation

For installing all python packages required, run pip install -r requirements.txt. Additionally, for generating data using Three-D-World(TDW), you might need to follow additional instructions based on your use-case following this.

Downloading Downstream Task Data: Please follow the instructions here for downloading data for the seen set of tasks. For unseen tasks, please follow the citations in the paper.

As a general rule, all commands below should be run from the base directory of this repository.

Generating Three-D-World Data

The file generator/generator.py contains the definition of a Generator class, an object of which is the main interface for generating Three-D-World data. The GenParams class in generator/generator_params.py defines the different simulation parameters, and an object of Generator uses and object of GenParams to generate a synthetic image set. Each GenParams object has corresponding tuple and string representations, which are just a concatenation of all its parameters with a '_' separator in case of a string.

For generating the 256 sets of size 40k images each, run python generator/generate_256_sets.py --root_dir /path/to/output/dir --num_nodes=1. The options num_nodes and idx (where idx specifies node index), can be used if generation of the 256 datasets is being split across multiple nodes.

Note that currently, the script assumes that the TDW build is launched manually, if not the launch_build argument in defining a Generator needs to be set to True.

Pre-training

NOTE: Our scripts used slurm for launching jobs and for it we used the submitit python utility. All jobs used for pre-training and downstream evaluation can be appropriately modified for machines that do not use slurm. Additionally, we used wandb for experiment tracking and logging. If not used, it can be disabled using the --debug flag in all of the following scripts.

For pre-training a Resnet-50 backbone on an image dataset, run the command

python run_with_submitit.py --debug --ngpus 4 --cfg-yml configs/pt_default.yaml \
--cfg-override SAVE_DIR path/to/output_dir DATA_DIR path/to/pre-training/image/data

Equivalently, on a machine that does not use slurm for job management (and has at least 4 gpus), run

bash distributed_train.sh 4 --cfg-yml configs/pt_default.yaml \
--cfg-override SAVE_DIR path/to/output_dir DATA_DIR path/to/pre-training/image/data

Downstream Evaluation

In the following, ChestX has been used as an example downstream task, which is to be replaced appropriately. Also, the hyperparameters in the override arguments were substituted appropriately from configs/lineval_best_hps.yaml or configs/finetune_best_hps.yaml.

Linear Probing

python run_with_submitit.py --debug --cfg-yml configs/lineval_defaults.yaml \
--cfg-override DOWNSTREAM_EVAL lineval DATASET ChestX \
BACKBONE_PATH path/to/pre-trained/model SAVE_DIR path/to/output_dir
LR 0.001 WD 0. BATCH_SIZE 128

Full Network Fine-tuning

python run_with_submitit.py --debug --cfg-yml configs/finetune_defaults.yaml \
--cfg-override DOWNSTREAM_EVAL finetune DATASET ChestX \
BACKBONE_PATH path/to/pre-trained/model SAVE_DIR path/to/output_dir
LR 0.001 WD 1.e-5 BATCH_SIZE 128

K-nearest neighbors

python run_with_submitit.py --debug --cfg-yml configs/knn_defaults.yaml \
--cfg-override DOWNSTREAM_EVAL knn DATASET ChestX \
BACKBONE_PATH path/to/pre-trained/model SAVE_DIR path/to/output_dir

Training Task2Sim

Using all the above, we pre-generated 256 synthetic image datasets, pre-trained a Resnet-50 on each and ran downstream evaluation using a 5 nearest neighbors classifier on all 20 downstream tasks. These accuracies, which for the seen tasks, are used as the rewards for training Task2Sim, are stored in controller_db/reward_db.json. Additionally, the Task2Vec vectors, which can be computed using the script controller/task_db.py are available in task_db/resnet18_imagenet_12_seen_tasks.py and task_db/resnet18_imagenet_8_unseen_tasks.py.

For training the Task2Sim model, simply run

python controller/train_controller.py --debug --cfg-override SAVE_DIR path/to/output/dir

Citation

If you use code from this repository in your work, please cite :

@inproceedings{mishra2022task2sim,
  title={Task2Sim: Towards Effective Pre-training and Transfer from Synthetic Data},
  author={Mishra, Samarth and Panda, Rameswar and Phoo, Cheng Perng and Chen, Chun-Fu Richard and Karlinsky, Leonid and Saenko, Kate and Saligrama, Venkatesh and Feris, Rogerio S},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9194--9204},
  year={2022}
}