tf.config.experimental.list_logical_devices
Return a list of logical devices created by runtime.
tf.config.experimental.list_logical_devices( device_type=None )
Logical devices may correspond to physical devices or remote devices in the cluster. Operations and tensors may be placed on these devices by using the name
of the LogicalDevice.
For example:
logical_devices = tf.config.experimental.list_logical_devices('GPU') # Allocate on GPU:0 with tf.device(logical_devices[0].name): one = tf.constant(1) # Allocate on GPU:1 with tf.device(logical_devices[1].name): two = tf.constant(2)
Args | |
---|---|
device_type | (optional) Device type to filter by such as "CPU" or "GPU" |
Returns | |
---|---|
List of LogicalDevice objects |
© 2020 The TensorFlow Authors. All rights reserved.
Licensed under the Creative Commons Attribution License 3.0.
Code samples licensed under the Apache 2.0 License.
https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/config/experimental/list_logical_devices