microsoft / tensorflow-directml

Fork of TensorFlow accelerated by DirectML
Apache License 2.0
457 stars 32 forks source link

Add unsorted segment ops for DML (emulated on the CPU) #383

Closed PatriceVignola closed 2 years ago

PatriceVignola commented 2 years ago

The DML implementation of segment ops was using way too much memory to be practically usable, so we'll use an emulated version for now. So far, it seems like this op isn't used that many times per epoch and is generally paired with Random ops that are used to initialize weights.