diff options
author | A. Unique TensorFlower <gardener@tensorflow.org> | 2018-01-11 10:14:43 -0800 |
---|---|---|
committer | TensorFlower Gardener <gardener@tensorflow.org> | 2018-01-11 10:18:45 -0800 |
commit | 3abd6e8809ca9c3072e3d846e8313d854e63ed65 (patch) | |
tree | da96c44378929c9eeff117f3b25ed167121831e9 /tensorflow/contrib/kfac | |
parent | 0578ee1f652bcad18d5dc7089eab12b6fff60333 (diff) |
Changing the default value of the momentum constant to 0.9.
Changing the default value of colocate_gradients_with_ops to True.
PiperOrigin-RevId: 181624864
Diffstat (limited to 'tensorflow/contrib/kfac')
-rw-r--r-- | tensorflow/contrib/kfac/python/ops/estimator.py | 4 | ||||
-rw-r--r-- | tensorflow/contrib/kfac/python/ops/optimizer.py | 9 |
2 files changed, 7 insertions, 6 deletions
diff --git a/tensorflow/contrib/kfac/python/ops/estimator.py b/tensorflow/contrib/kfac/python/ops/estimator.py index 5e1680967c..02b0677824 100644 --- a/tensorflow/contrib/kfac/python/ops/estimator.py +++ b/tensorflow/contrib/kfac/python/ops/estimator.py @@ -74,7 +74,7 @@ class FisherEstimator(object): damping, layer_collection, estimation_mode="gradients", - colocate_gradients_with_ops=False, + colocate_gradients_with_ops=True, cov_devices=None, inv_devices=None): """Create a FisherEstimator object. @@ -110,7 +110,7 @@ class FisherEstimator(object): is more expensive to compute than the other three options by a factor equal to the output dimension, roughly speaking. colocate_gradients_with_ops: Whether we should request gradients be - colocated with their respective ops. + colocated with their respective ops. (Default: True) cov_devices: Iterable of device strings (e.g. '/gpu:0'). Covariance computations will be placed on these devices in a round-robin fashion. Can be None, which means that no devices are specified. diff --git a/tensorflow/contrib/kfac/python/ops/optimizer.py b/tensorflow/contrib/kfac/python/ops/optimizer.py index ecf7f3e4e5..0c9444241f 100644 --- a/tensorflow/contrib/kfac/python/ops/optimizer.py +++ b/tensorflow/contrib/kfac/python/ops/optimizer.py @@ -41,12 +41,12 @@ class KfacOptimizer(gradient_descent.GradientDescentOptimizer): damping, layer_collection, var_list=None, - momentum=0., + momentum=0.9, momentum_type="regular", norm_constraint=None, name="KFAC", estimation_mode="gradients", - colocate_gradients_with_ops=False, + colocate_gradients_with_ops=True, cov_devices=None, inv_devices=None): """Initializes the KFAC optimizer with the given settings. @@ -70,8 +70,8 @@ class KfacOptimizer(gradient_descent.GradientDescentOptimizer): var_list: Optional list or tuple of variables to train. Defaults to the list of variables collected in the graph under the key `GraphKeys.TRAINABLE_VARIABLES`. - momentum: The momentum value for this optimizer. Only applies when - momentum_type is 'regular' or 'adam'. (Default: 0) + momentum: The momentum decay constant to use. Only applies when + momentum_type is 'regular' or 'adam'. (Default: 0.9) momentum_type: The type of momentum to use in this optimizer, one of 'regular', 'adam', or 'qmodel'. (Default: 'regular') norm_constraint: float or Tensor. If specified, the update is scaled down @@ -85,6 +85,7 @@ class KfacOptimizer(gradient_descent.GradientDescentOptimizer): more a more detailed description of these options. colocate_gradients_with_ops: Whether we should request gradients we compute in the estimator be colocated with their respective ops. + (Default: True) cov_devices: Iterable of device strings (e.g. '/gpu:0'). Covariance computations will be placed on these devices in a round-robin fashion. Can be None, which means that no devices are specified. |