Reproduce Keras training results in Jupyter Notebook

Question

I am using keras and Jupyter notebook and want to make my results reproducible every time I ran it. This is the tutorial I used https://machinelearningmastery.com/time-series-prediction-lstm-recurrent-neural-networks-python-keras/. I copied his codes in Stacked LSTMs with Memory Between Batches part.

This is my cell1 in Jupyter Notebook, I only used CPU to avoid randomness brought by GPU, making sure the same results can be reproduced every time.

import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"   # see issue #152
os.environ["CUDA_VISIBLE_DEVICES"] = ""
from tensorflow.python.client import device_lib
device_lib.list_local_devices()

This is cell2, stricting following the suggestions from this question https://stackoverflow.com/questions/32419510/how-to-get-reproducible-results-in-keras

# Seed value
# Apparently you may use different seed values at each stage
seed_value= 0

# 1. Set the `PYTHONHASHSEED` environment variable at a fixed value
import os
os.environ['PYTHONHASHSEED']=str(seed_value)

# 2. Set the `python` built-in pseudo-random generator at a fixed value
import random
random.seed(seed_value)

# 3. Set the `numpy` pseudo-random generator at a fixed value
import numpy as np

np.random.seed(seed_value)

import tensorflow as tf

tf.compat.v1.set_random_seed(seed_value)

tf.random.set_seed(seed_value)

This is cell3 from his codes, (changed it a little for example, from keras to tensorflow.keras)

import numpy
import matplotlib.pyplot as plt
from pandas import read_csv
import math
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Dense, LSTM

from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_squared_error

# convert an array of values into a dataset matrix
def create_dataset(dataset, look_back=1):
    dataX, dataY = [], []
    for i in range(len(dataset)-look_back-1):
        a = dataset[i:(i+look_back), 0]
        dataX.append(a)
        dataY.append(dataset[i + look_back, 0])
    return numpy.array(dataX), numpy.array(dataY)

# load the dataset
dataframe = read_csv('airline-passengers.csv', usecols=[1], engine='python')
dataset = dataframe.values
dataset = dataset.astype('float32')
# normalize the dataset
scaler = MinMaxScaler(feature_range=(0, 1))
dataset = scaler.fit_transform(dataset)
# split into train and test sets
train_size = int(len(dataset) * 0.67)
test_size = len(dataset) - train_size
train, test = dataset[0:train_size,:], dataset[train_size:len(dataset),:]

# reshape into X=t and Y=t+1
look_back = 3
trainX, trainY = create_dataset(train, look_back)
testX, testY = create_dataset(test, look_back)
# reshape input to be [samples, time steps, features]
trainX = numpy.reshape(trainX, (trainX.shape[0], trainX.shape[1], 1))
testX = numpy.reshape(testX, (testX.shape[0], testX.shape[1], 1))

And this is cell4,

# create and fit the LSTM network
batch_size = 1
model = Sequential()
model.add(LSTM(4, batch_input_shape=(batch_size, look_back, 1), stateful=True, return_sequences=True))
model.add(LSTM(4, batch_input_shape=(batch_size, look_back, 1), stateful=True))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
for i in range(100):
    model.fit(trainX, trainY, epochs=1, batch_size=batch_size, verbose=2, shuffle=False)
    model.reset_states()
# make predictions
trainPredict = model.predict(trainX, batch_size=batch_size)
model.reset_states()
testPredict = model.predict(testX, batch_size=batch_size)
# invert predictions
trainPredict = scaler.inverse_transform(trainPredict)
trainY = scaler.inverse_transform([trainY])
testPredict = scaler.inverse_transform(testPredict)
testY = scaler.inverse_transform([testY])
# calculate root mean squared error
trainScore = math.sqrt(mean_squared_error(trainY[0], trainPredict[:,0]))
print('Train Score: %.2f RMSE' % (trainScore))
testScore = math.sqrt(mean_squared_error(testY[0], testPredict[:,0]))
print('Test Score: %.2f RMSE' % (testScore))
# shift train predictions for plotting
trainPredictPlot = numpy.empty_like(dataset)
trainPredictPlot[:, :] = numpy.nan
trainPredictPlot[look_back:len(trainPredict)+look_back, :] = trainPredict
# shift test predictions for plotting
testPredictPlot = numpy.empty_like(dataset)
testPredictPlot[:, :] = numpy.nan
testPredictPlot[len(trainPredict)+(look_back*2)+1:len(dataset)-1, :] = testPredict
# plot baseline and predictions
plt.plot(scaler.inverse_transform(dataset))
plt.plot(trainPredictPlot)
plt.plot(testPredictPlot)
plt.show()

However, I still found that the training loss is different every time I ran cell4.

But I found that, as long as I add cell2's contents to cell4, I can get the same training loss curve every time I ran cell4.

So my question is, to reproduce my results, why should I set the random seed every time I run my model in the cell(cell4), instead of just setting it in the beginning of my jupyter notebook once and for all?

Answer 1

Did you set the same random seed at each step?

The seed works well for the first function, but then it is lost in the next ones because NumPy applies a global seed reset automatically.

For example, you can do:

def reset_seed(seed_value):
    np.random.RandomState(seed_value)
    tf.compat.v1.set_random_seed(seed_value)
    tf.random.set_seed(seed_value)

for i in range(100):
    reset_seed(seed_value)
    model.fit(trainX, trainY, epochs=1, batch_size=batch_size, verbose=2, shuffle=False)
    model.reset_states()

Otherwise, you can also use a function with a random number generator: https://albertcthomas.github.io/good-practices-random-number-generators/