Turn the randomness off in order to reproduce results

[BogdanM1]

Is there a way to exclude all the stochastics within SciANN so that we get results that can be easily reproduced?

So far I tried
import tensorflow as tf
from numpy.random import seed
import random

_seed = 0
seed(_seed)
random.seed(_seed)
tf.random.set_seed(_seed)
tf.compat.v1.set_random_seed(_seed)

but I am getting different results each time I run the training.

[ehsanhaghighat]

Use sn.set_random_seed(). Also to get exactly the same results, you need to turn off TF multithreading otherwise order of operations may perturb the outcome.

…

On Nov 6, 2021, at 9:25 AM, Bogdan ***@***.***> wrote:  Is there a way to exclude all the stochastics within SciANN so that we get results that can be easily reproduced? So far I tried import tensorflow as tf from numpy.random import seed import random _seed = 0 seed(_seed) random.seed(_seed) tf.random.set_seed(_seed) tf.compat.v1.set_random_seed(_seed) but I am getting different results each time I run the training. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[BogdanM1]

Thank you for your fast response!

This is what I did:
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
import tensorflow as tf
from numpy.random import seed
import random
import sciann as sn

      _seed = 0
      
      seed(_seed)
      random.seed(_seed)
      tf.random.set_seed(_seed)
      tf.compat.v1.set_random_seed(_seed)
      
      sn.set_random_seed(_seed)
      
      session_conf = tf.compat.v1.ConfigProto(intra_op_parallelism_threads=1,inter_op_parallelism_threads=1)
      sess = tf.compat.v1.Session(graph=tf.compat.v1.get_default_graph(), config=session_conf)
      K.set_session(sess)

Unfortunatelly, I am still getting minor differences in my results.

[pradhyumna85]

Hey @BogdanM1,
Just have a look at this issue which I raised - "https://github.com/sciann/sciann/issues/53", and you can also find the code changes in the associated pull request "https://github.com/sciann/sciann/pull/57" which I raised.

This tweak worked for me for reproducibility. Let me know it that helps.