Cannot Reproduce Fraxtil Step Placement Result Using Pretrained Model

[chaonan99]

Hi Chris,

Thanks for this great work. I was trying to reproduce the reported result with the pretrained model provided in the infer directory. I used Docker so I suppose there is no inconsistency in the environment. I run the model on Fraxtil test set and saved the placed_times for each chart, then used the predicted times and ground truth times to compute F1-c. But I can only get ~0.587 out of that, which is far from the 0.681 reported in your paper, and even worse than LogReg baseline. Is this model actually trained on ITG dataset, or did I do something wrong?

By the way, the link to download ITG dataset is broken. Could you please point me the way to download the dataset?

Thanks!