-
Notifications
You must be signed in to change notification settings - Fork 140
Change Log
We now support rapids.ai-accelerated UMAP dimensionality reduction, greatly reducing the time necessary to embed the 2,048-dimensional space into two dimensions. Here is a comparison of regular (umap-learn) and cuda-accelerated (rapids.ai) UMAP on the same dataset:
-----
oslo no hyperparameters regular umap:
2021-07-20 16:18:44.520148: Creating UMAP layout
2021-07-20 16:19:08.822013: Creating umap pointgrid
= 24 seconds
-----
oslo no hyperparameters rapidsai umap:
2021-07-20 17:21:41.096082: Creating UMAP layout
2021-07-20 17:21:43.098537: Creating umap pointgrid
= 2 seconds
Due to the complexities of rapids.ai, we recommend setting up a conda environment as detailed on their setup page. Pixplot, which now uses TensorFlow 2.x (see below) can be set up in this conda environment and will automatically use gpu acceleration if the rapids.ai libraries are present and importable.
Pixplot now uses TensorFlow 2.x instead of 1.x. This will allow us to take advantage of newer GPUs, such as NVIDIA's 30x0 series, as well as gpu acceleration at other stages in our pipeline.
This shift does mean we have deprecated the TF1.x-based OpenPose support. We're looking at building a more robust framework for multiple kinds of specialized neural networks -- poses, faces, car makes, etc -- and hope to have some wiki pages about how to integrate these networks soon.