3 ms·
From a quick glance, it appears that it's because the implementation pushes the entire graph (all edges) to the GPU. Sampling of edges during training could all
by bravura 2y ago
From a quick glance, it appears that it's because the implementation pushes the entire graph (all edges) to the GPU. Sampling of edges during training could alleviate this.
- minimaxir 2y agoIndeed, TensorFlow likes pushing everything to the GPU by default whereas many PyTorch DL implementations encourage feeding data from the CPU to the GPU as needed with a DataLoader. There have been attempts at a PyTorch port of Parametric UMAP (https://github.com/lmcinnes/umap/issues/580 https://github.com/lmcinnes/umap/issues/580) but nothing as good.
- bravura 2y agoLooks like there is a little motion on this topic: https://github.com/lmcinnes/umap/pull/1103 https://github.com/lmcinnes/umap/pull/1103