# Distributed training with low level Tensorflow API

I am using low level Tensorflow API's for my model training. When I say low level it means I'm defining the tf.Session() object of the graph and evaluate graph with in this session.

I would like to distribute the model training using tf.distribute.MirroredStrategy().

I am able to use mirroredstrategy() on tensorflow sequential API's using the example shared by tensorflow in their document.

But I am facing difficulty in executing tf low level code using mirror strategy.

I tried to use distribute.MirrorStrategy() and below are the results of resource utilization:

[0] GeForce RTX 2080 Ti | 48'C, 40 % | 10771 / 11019 MB | vipin(10763M) gdm(4M)
[1] GeForce RTX 2080 Ti | 37'C, 0 % | 10376 / 11014 MB | vipin(10327M) gdm(36M) gdm(8M)

Even though model used the memory of both the GPU's, but still GPU1 utilization is 0.

I am not sure about the issue. Even not sure if tensorflow support this.

Please clear my doubts and if possible share the sample code as well.

• Can you share some code that leads up to the above? In this way w emay be able to help more. Jun 17 at 11:05

1- Use distribute.MirrorStrategy() at the beginning of your code (just after library declaration) to activate correctly all the related functions.

2- Set env variable CUDA_VISIBLE_DEVICES=1 (https://stackoverflow.com/questions/37893755/tensorflow-set-cuda-visible-devices-within-jupyter)

3- Ensure that mirrored variables are set correctly thanks to:

with strategy.scope():
x = tf.Variable(..)

• Thanks for your response. Rather I want to utlize my both the available GPU's using Tensorflow Mirror Strategy API's. As I mentioned, I'm suing Low Level Tensorflow API's and defining the session() object and then using sess.run(), executing the model. Even I have defined my whole graph under startegy.scope(). But still no luck. Jun 23 at 3:32
• Did you force strategy to GPU 1 only to see if it can be activated? It could be a hardware issue. You can test if the GPU 1 is KO thanks to gpu-burn: github.com/wilicc/gpu-burn Jun 23 at 5:52
• Thinks are working fine, if Im using Tensorflow Sequential API. Not sure if Mirror Strategy works for low level API's? Jun 23 at 7:06
• What is the batch size? Did you use experimental_distribute_datasets_from_function inside with strategy.scope() to split manually your data? Without reading the code, it's more difficult to see what is going wrong. Here is the most complete guide I've found about mirrored strategy: youtube.com/watch?v=jKV53r9-H14 Jun 23 at 7:42
• Thanks for the link. Meanwhile if you have 2 gpus machine, you can try some basic code . For me its difficult to share the code due to some restrictions. Jun 24 at 4:06