How to remove randomness? about keras-tcn HOT 7 CLOSED

@sishue good point. In tensorflow, it's very difficult to have reproducible results (especially on a GPU).

from keras-tcn.

Just a heads-up that reproducible results are possible on (NVIDIA) GPUs using the tensorflow-determinism library. I've been getting reproducible results consistently for CNNs, and I also just tested with keras-tcn and got reproducible results, so it is possible.

from keras-tcn.

@lingdoc yeah you're right. Can you detail the steps to make it reproducible? That would be highly appreciated!

from keras-tcn.

Is it just adding those lines at the beginning of the script that you run?

import tensorflow as tf
from tfdeterminism import patch
patch()
os.environ['PYTHONHASHSEED']=str(SEED)
random.seed(SEED)
np.random.seed(SEED)
tf.set_random_seed(SEED)
# use tf as normal

from keras-tcn.

yes, after installing the library and setting SEED to some value, i.e. SEED = 54. I'm running on TF 1.15, btw - haven't tried it on TF 2.0

This should ensure that any sources of randomization, like shuffle=True in the Keras model.fit() call or the various split functions in sklearn will be deterministic, provided they are using one of these random seeds. One caveat is that if you change your hardware or model architecture it seems the results will be different, but as long as all the hardware is the same I get the same outcome from running the same model.

from keras-tcn.

That sounds already very promising! Thank you very much for the feedback. I'm going to add it to the README

from keras-tcn.

Done! @sishue try the tensorflow-determinism library. It should work!

from keras-tcn.