Python: Fixing the Random Seed for Reproducibility

This post shows how to fix the random seed to get reproducible results with Keras.

Setting random seeds fixed

For an illustration, a simple DNN model is implemented in the Jupyter Notebook using a function for fixing the random seed. The function seed_everything() is borrowed from https://dacon.io/codeshare/2363.

Every time we run this code as a whole, the result is always 96.562645 and is not changed.

# Import library and define functions
 
# In[1]:
_________________________________________________________________
from keras.models import Sequential
from keras.layers import Dense
 
import tensorflow as tf
import numpy as np
import random
import os
 
def seed_everything(seed: int = 42):
    random.seed(seed)
    np.random.seed(seed)
    os.environ["PYTHONHASHSEED"] = str(seed)
    tf.random.set_seed(seed)
_________________________________________________________________
 
# Simple DNN model with a fixed random seed 
# In[2]:
_________________________________________________________________
x = np.array([[10, 20, 30], [20, 30, 40], [30, 40, 50], [40, 50, 60]])
y = np.array([40, 50, 60, 70])
 
seed_everything(1) # fix the random seed
 
model = Sequential()
model.add(Dense(50, activation='relu', input_dim=3))
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse')
model.fit(x, y, epochs=1000, verbose=0)
 
x_input = np.array([60, 70, 80])
x_input = x_input.reshape((1,3))
pred = model.predict(x_input, verbose=0)
print(pred)
_________________________________________________________________
 
[[96.562645]]

Robustness checks

However, when we reuse and rerun some parts of the above code in the same file, we can encounter some unexpected results as follows.

# rerun model fit
# In[3]:
_________________________________________________________________
model.fit(x, y, epochs=1000, verbose=0)
x_input = np.array([60, 70, 80])
x_input = x_input.reshape((1,3))
pred = model.predict(x_input, verbose=0)
print(pred)
_________________________________________________________________
[[90.000145]]
 
# rerun model.fit with a fixed random seed
# In[4]:
_________________________________________________________________
seed_everything(1)
 
model.fit(x, y, epochs=1000, verbose=0)
x_input = np.array([60, 70, 80])
x_input = x_input.reshape((1,3))
pred = model.predict(x_input, verbose=0)
print(pred)
_________________________________________________________________
[[90.00001]]
 
 
# rerun model.compile & model.fit with a fixed random seed
# In[5]:
_________________________________________________________________
seed_everything(1)
 
model.compile(optimizer='adam', loss='mse')
model.fit(x, y, epochs=1000, verbose=0)
x_input = np.array([60, 70, 80])
x_input = x_input.reshape((1,3))
pred = model.predict(x_input, verbose=0)
print(pred)
_________________________________________________________________
[[89.99999]]
 
# call seed_everything(1) model = Sequential()
# In[6]:
_________________________________________________________________
seed_everything(1)
 
model = Sequential()
model.add(Dense(50, activation='relu', input_dim=3))
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse')
model.fit(x, y, epochs=1000, verbose=0)
x_input = np.array([60, 70, 80])
x_input = x_input.reshape((1,3))
pred = model.predict(x_input, verbose=0)
print(pred)
_________________________________________________________________
[[96.562645]]

The bottom line is that to reproduce always the same result of the same model several times in one file, seed_everything() function needs to be located before model = Sequential().

Originally posted on SH Fintech Modeling Blog.