3 ms·
Indeed, TensorFlow likes pushing everything to the GPU by default whereas many PyTorch DL implementations encourage feeding data from the CPU to the GPU as need
by minimaxir 2y ago
Indeed, 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