This is the code for the paper: Comparing Reconstruction- and Contrastive-based Models for Visual Task Plannin
Visit the dedicated website: here for more information.
If you use this code in your work, please cite it as follows:
Avalible after Puplication
pip install -r requirements.txt
prepare datasets:
cd datasets
chmod +x get_datasets.sh
./get_datasets.sh
prepare datasets:
python preprocess_dataset/augment_dataset.py --pkl_dataset='datasets/2500_shelf_stacking' --ac_rand_mul=2
python preprocess_dataset/augment_dataset.py --pkl_dataset='datasets/2500_shelf_stacking_all_distractors' --ac_rand_mul=2
python preprocess_dataset/augment_dataset.py --pkl_dataset='datasets/box_manipulation_view1_train' --ac_rand_mul=2
python preprocess_dataset/augment_dataset.py --pkl_dataset='datasets/box_manipulation_view2_train' --ac_rand_mul=2
python preprocess_dataset/augment_dataset.py --pkl_dataset='datasets/box_manipulation_view1_view2_train' --ac_rand_mul=2
Box stacking task:
python run_box_stacking_pca_siam_simntx.py
Shelf arranging task:
python run_shelf_arrangment_pca_siam_simntx.py
Box manipulation task:
python run_box_manipulation_pca_siam_simntx.py
Builds on code work from: Latent Space Roadmap website
python prepare_recon_dataset_for_cluster.py
chmod +x run_train_recon_models.sh
./run_train_recon_models.sh
To get the encodins and plots run:
python produce_recon_models_encodings.py
The representations are saved in the model folder together with the t-SNE plots:
