DISK paper code

Additional scripts to reproduce analyses from the DISK paper. DISK code is available at: https://github.com/bozeklab/DISK

List of figures and related code

Figure	Title	Script files	Comments
Fig. 1	The missing data problem	-	Measure % missing per dataset
Fig. 2 a & f	DISK and other imputation methods’ performance across datasets	fig2_barplots.py	outputs of DISK main code (test_fillmissing script); input files in fig2a_result_files folder
Fig. 2 b - e & g	DISK and other imputation methods’ performance across datasets	final_comparison_methods_short_segment.py	input files in comparison_methods_files
Fig. 2 h & i	DISK and other imputation methods’ performance across datasets	final_comparison_methods_full_file.py	input files in comparison_methods_files
Fig. 3	Estimated imputation error	-	outputs of DISK main code (test_fillmissing script); input files in fig3_files folder
Fig. 4 a	Inference of multiple simultaneously missing keypoints	fig_4a_nmissing.py	input files in fig4_files folder
Fig. 4 c	Inference of multiple simultaneously missing keypoints	fig_4b.py	input files in fig4_files folder
Fig. 4 d-f	Inference of multiple simultaneously missing keypoints	-	outputs of DISK main code (test_fillmissing script)
Fig. 5	DISK learns meaningful representations of sequences from Human and 2-Mice-2D datasets	-	outputs of DISK main code (embedding_umap script); checkpoints and other files in fig5_files folder
Fig. 6	DISK allows to detect more steps and emphasizes differences in step kinematics between different treatments	folder fig6_step_detection
Ext. Fig. 1 a - b	Comparison of different architectures [...] and of different methods according to the MPJPE and PCK@0.01 metrics.	fig2_barplots.py
Ext. Fig. 1 c - d	Comparison of different architectures [...] and of different methods according to the MPJPE and PCK@0.01 metrics.	final_comparison_methods_short_segment.py
Ext. Fig. 2	PCK@0.01, RMSE, and MPJPE per keypoint and averaged across keypoints (”all”) in datasets	fig2_barplots.py
Ext. Fig. 3 (upper row)	Test RMSE with respect to the gap length for the other datasets.	-	outputs of DISK main code (test_fillmissing script)
Ext. Fig. 3 (lower row)	Test RMSE with respect to the gap length for the other datasets.	final_comparison_methods_short_segment.py
Ext. Fig. 4	DISK performance dependence on dataset size and input length (2-Fish dataset)	DISK_peformance_dataset_size_input_length.py
Ext. Fig. 5 - 6	Imputation of switched keypoints.	analyse_swap_keypoints.py
Ext. Fig. 7	Observed missing proportions for each keypoint in mouse FL2 and CLB datasets	-	outputs of DISK main code (create_proba_missing script)
Supp Data	Analysis of DISK performance with respect to action classes, quantity of movement, periodicity, and gap length	DISK_perforamcne_vs_action.py
Supp Data	Estimated error correlation plots [...] for DISK-proba and GRU-proba for all tested datasets.	-	outputs of DISK main code (test_fillmissing script)
Supp Data	RMSE wrt the distance between the two fish and the number and scheme of missing keypoints for short gaps and long gaps	fig_4b.py
Supp Data	DISK learns meaningful representations of 1 sec-long sequences of the Mouse FL2 dataset	-	outputs of DISK main code (embedding_umap script)
Supp Data	Number of parameters for each network.	-	-
Supp Data	Comparison [...] 2-fish dataset, all keypoints taken together or only 1 fish	fig_4b.py	-

Comparison with other published methods

See Note file for the description of the comparison and step by step pipeline.

Links to external github repositories:

• Optipose
• Keypoint-moseq
• MarkerBasedImputation

Switch data augmentation module

• Available in the branch swap of DISK (we used "switch" in the manuscript and "swap" in the code, but they mean the same thing. Our apologies)