search

originrose / cortex

Machine learning in Clojure
Eclipse Public License 1.0
1.27k stars 111 forks source link

Resnet50 profiling work #219

Closed cnuernber closed 7 years ago

cnuernber commented 7 years ago

Major profiling wins in this work:

  1. Generalized select in the tensor opts is expensive. Special cased out submatrix.
  2. Found and fixed a holding-onto-head issue reducing overall memory usage.
  3. The stream-augmentation system was broken up making loading data a lot faster.