Help with replicating SoccerNet Action Spotting results
Paulkie99 opened this issue · comments
Hi, I am trying to replicate your results for SoccerNet action spotting. However, the large number of parameters associated with the embedding size is causing me problems. Below is a tensorflow.keras model summary of my model:
However I get an OOM error, suspecting somewhere in the multi-headed attention:
It makes me curious how you managed to train this network since it was not mentioned in the paper that any dimensionality reduction was used. I am using an RTX 3080 with 10 GB VRAM, please advise.
Update: the model seems to train fine on Google Colab TPUs. Seems my GPU just doesn't have enough memory. Sorry for bothering with this. I still think it's strange that you managed to train on a Titan X which has only 2 GB more memory than my 3080, but I guess it was enough to make a difference.