search

IBM / USD

Code for the NeurIPS 2020 paper "Unbalanced Sobolev Descent"
Apache License 2.0
7 stars 4 forks source link

Regarding wot_comparison.ipynb #3

Open Piyushi-0 opened 3 years ago

Piyushi-0 commented 3 years ago

I was trying the USD code on the Embryoid Body dataset used in the TrajectoryNet paper. The line mid_step = find_intersec(mmd_to_ref(collQ, dataQ0), mmd_to_ref(collQ, dataP))[0] inside the function run_interval(...) of wot_comparison.ipynb returns error due to the intersection being empty. How do we compute mmd & emd for USD in this case?

matrig commented 3 years ago

Hi @Piyushi-0, thanks for your interest in our paper! Sorry, I am not able to reproduce the error. Can you please find out if the code runs correctly until the line dataQ, collQ, coll_mmd = train_unbalanced_descent(D, dataQ0, dataP, None, opt) in run_interval()? Could you for instance insert a break point after that line and verify that dataQ is a FloatTensor of size 4556 x 30?

Piyushi-0 commented 3 years ago

Hi @matrig, thanks for the reply. The code ran perfectly to reproduce results of the experiment Developmental Trajectories of Single Cells, thanks for making the code available. However, I had problems running it for a similar experiment given in the TrajectoryNet paper. I was trying to run the USD code on Embryoid Body sc-RNA time course data to get results like Table 3 of the TrajectoryNet paper. I loaded the data as they show in their github repository. I am able to run USD for the Table 3 experiment for timesteps 2 & 3 but while running the code for timestep 1, I get the error that I mentioned in my first post. Before the line mid_step = find_intersec(mmd_to_ref(collQ, dataQ0), mmd_to_ref(collQ, dataP))[0] the code runs perfectly. dataQ was a FloatTensor of size num_samples x num_features.