Last active
February 25, 2019 13:07
-
-
Save eladshabi/205787e471d5ef162630a2766202abc6 to your computer and use it in GitHub Desktop.
Optimizers
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
# source: https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html | |
with tf.device('/gpu:0'), \ | |
tf.variable_scope('fp32_storage',custom_getter=float32_variable_storage_getter): | |
data, target, loss = create_simple_model(nbatch, nin, nout, dtype) | |
variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) | |
# Traning variables | |
lr = 0.0002 | |
beta = 0.5 | |
variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) | |
model_opt = tf.train.AdamOptimizer(learning_rate, beta) # Adam optimizer | |
# Mixed precision | |
scale_size = 128 # There is no one scale size | |
loss_scale_manager = FixedLossScaleManager(scale) | |
loss_scale_optimizer = LossScaleOptimizer(model_opt, loss_scale_manager) | |
# Calculate the gradients with scaled loss and return the unscaled gradients | |
grads_variables = loss_scale_optimizer.compute_gradients(loss, variables) | |
""" | |
Doing some gradient manipulation (if needed) | |
only example! | |
grads_variables = [(g,v) for (g,v) in grads_variables if g is not None] | |
""" | |
training_opt = loss_scale_optimizer.apply_gradients(grads_variables) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment