Tensorflow model.fit() reproducibility

Question

import tensorflow as tf

RANDOM_SEED_CONSTANT = 42  # FOR_REPRODUCIBILITY
tf.random.set_seed(RANDOM_SEED_CONSTANT)

# Prevent NHWC errors https://www.nuomiphp.com/eplan/en/50125.html
from tensorflow.keras import backend as K
K.set_image_data_format("channels_last")

from tensorflow import keras
from tensorflow.keras import datasets, layers, models

(train_images, train_labels), (test_images, test_labels) = datasets.cifar10.load_data()
train_images, test_images = train_images / 255.0, test_images / 255.0 # Normalize pixel values to be between 0 and 1

# Create a simple CNN
model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(32, 32, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.Flatten())
model.add(layers.Dense(64, 
                       activation='relu', 
                       kernel_initializer=tf.keras.initializers.HeNormal(seed=RANDOM_SEED_CONSTANT)))
model.add(layers.Dense(10, 
                       kernel_initializer=tf.keras.initializers.HeNormal(seed=RANDOM_SEED_CONSTANT)))

print(model.summary())

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

model.save_weights('myweights.h5')

# Run1
history = model.fit(train_images, train_labels, epochs=1, 
                    shuffle=False,
                    validation_data=(test_images, test_labels))

# Run2
model.load_weights('myweights.h5')
history = model.fit(train_images, train_labels, epochs=1, 
                    shuffle=False,
                    validation_data=(test_images, test_labels))

# Run3
model.load_weights('myweights.h5')
history = model.fit(train_images, train_labels, epochs=1, 
                    shuffle=False,
                    validation_data=(test_images, test_labels))

The above 3 model.fit() calls gives me the following results:

1563/1563 [==============================] - 7s 4ms/step - loss: 1.4939 - accuracy: 0.4543 - val_loss: 1.2516 - val_accuracy: 0.5567

1563/1563 [==============================] - 6s 4ms/step - loss: 1.6071 - accuracy: 0.4092 - val_loss: 1.3857 - val_accuracy: 0.4951

1563/1563 [==============================] - 7s 4ms/step - loss: 1.5538 - accuracy: 0.4325 - val_loss: 1.3187 - val_accuracy: 0.5294

What is the reason for this difference? I am trying to understand sources which might impede reproducing results from models. Apart from random seed, dense layers initialization, what else am I missing?

Answer 1

The way you are testing the reproducibility is not correct. You need to close the program and rerun it to see if the results are the same. Otherwise, the run 2 depends on the events that happened during the run 1, and the run 3 depends on the events that happened during the run 1 and 2.

The reason is that Tensorflow maintains an internal counter for random generation, as stated in the documentation of tf.random.set_seed (emphasis is mine) :

print(tf.random.uniform([1], seed=1))  # generates 'A1'
print(tf.random.uniform([1], seed=1))  # generates 'A2'

The reason we get 'A2' instead 'A1' on the second call of tf.random.uniform above is because the same tf.random.uniform kernel (i.e. internal representation) is used by TensorFlow for all calls of it with the same arguments, and the kernel maintains an internal counter which is incremented every time it is executed, generating different results.

If I run only the first run of your program twice, closing the program between each run (in IPython in that case), I get:

In [1]: run program.py
1563/1563 [==============================] - 13s 8ms/step - loss: 1.4997 - accuracy: 0.4540 - val_loss: 1.2528 - val_accuracy: 0.5494
{'loss': [1.4996991157531738], 'accuracy': [0.4540199935436249], 'val_loss': [1.2527965307235718], 'val_accuracy': [0.5493999719619751]}
In [2]: run program.py
1563/1563 [==============================] - 12s 8ms/step - loss: 1.4997 - accuracy: 0.4540 - val_loss: 1.2528 - val_accuracy: 0.5494
{'loss': [1.4996991157531738], 'accuracy': [0.4540199935436249], 'val_loss': [1.2527965307235718], 'val_accuracy': [0.5493999719619751]}

Minus the time taken to perform the computation, that can vary a bit depending on the load on the machine, the results are completely identical.