Closed davidchd closed 11 months ago
I am very sorry for the long reply.
The issue is the data_loaders
variable was not updated from the training regime (there is a frames parameter there, which cuts random N frames, where N=60 is the tutorial). To make a prediction for the whole sequence you can set to_cut=False
(True
by default) during creating a dataloader like (data_loaders = mouse_video_loader(... to_cut=False)
(frames
parameter would be ignored in this case and generate_submission
should cut the redundant frames).
Hope this helps!
The generated submission file appears to be broken when I tried to run model_demo.ipynb. I used the first model mentioned in the notebook, and I barely changed anything. After some examination of the generated file, I figured that each cell in prediction is an array of size (7800+, 10) instead of the required size for submission (xxxx, 250). This is a screenshot of the error message when I submit my file. I'm using the Main [new] track.
Edit: In addition to the issue above, I wonder if there is an api that we can use to submit our result. Uploading through web is painful, especially for people who train models on a remote server. One would have to download the data file from the remote host and upload locally. It can take an hour to just uploading a single submission.