You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
* Improve CMake build script
* Builds or ignore CUDA objects based on availibility of CUDA tools.
* Builds C++ binaries of tfrnnlm if `-DTENSORFLOW_DIR=...` is provided.
* Select Apple BLAS by default when building on MacOS.
* Enable CTest for tests (run `ctest` or `make test`).
Typical build on Linux with CUDA+MKL:
```
export MKLROOT
cmake -DMATHLIB=MKL <cmakefile>
make
ctest
```
* Update model training script to TensorFlow v2 flavor
* Enable rescore tool to load from SavedModel format
* Update the training script which uses custom loss function
Still tried to mimic original behavior without whole pipeline ready.
* Fix numerical stability of the fast softmax model
0 commit comments