diff options
author | 2018-10-01 15:25:27 -0700 | |
---|---|---|
committer | 2018-10-01 15:29:11 -0700 | |
commit | cca204f12a5838f0ffdd4a80c27d451cf61d3636 (patch) | |
tree | 0338708443f5acb4d3e1caa9111ffb1316a861e7 /tensorflow | |
parent | c7237e6070dbf4acd1ade5a40dc676418cbd889b (diff) |
Added option (off by default) to enable a higher-performance variant of the
Adam optimizer's variable update formula.
PiperOrigin-RevId: 215290881
Diffstat (limited to 'tensorflow')
-rw-r--r-- | tensorflow/contrib/tpu/proto/optimization_parameters.proto | 6 |
1 files changed, 5 insertions, 1 deletions
diff --git a/tensorflow/contrib/tpu/proto/optimization_parameters.proto b/tensorflow/contrib/tpu/proto/optimization_parameters.proto index a43f45554f..8529b48c15 100644 --- a/tensorflow/contrib/tpu/proto/optimization_parameters.proto +++ b/tensorflow/contrib/tpu/proto/optimization_parameters.proto @@ -62,7 +62,10 @@ message FtrlParameters { // (https://www.tensorflow.org/api_docs/python/tf/contrib/opt/AdamOptimizer). If // use_non_lazy_adam is enabled, use_gradient_accumulation is also required in // order to get correct results; a warning will be printed otherwise (which may -// change to an error in the future). +// change to an error in the future). If use_max_with_epsilon is set, the Adam +// variable update formula will be changed from m / (sqrt(v) + epsilon) to +// m / max(sqrt(v), abs(epsilon)); this option improves the performance of TPU +// training and is not expected to harm model quality. message AdamParameters { float beta1 = 3; float beta2 = 4; @@ -70,6 +73,7 @@ message AdamParameters { float initial_m = 6; float initial_v = 7; bool use_non_lazy_adam = 8; + bool use_max_with_epsilon = 9; } // https://www.tensorflow.org/api_docs/python/tf/train/MomentumOptimizer |