I will show you how to fix the random number seed with Tensorflow 2.x
( tf.keras
).
The code used for the test can be found here (https://github.com/tokusumi/tf-keras-random-seed).
In the development of machine learning, there are demands such as "I want to make learning reproducible" and "I want to fix the initial value of the model for testing". Since the difference in the initial value of the weight affects the learning result, it seems that fixing the initial value will help solve these problems.
Random numbers are used to generate the initial value of the weight. Random numbers are generated based on random number seeds. By default, TensorFlow has a variable random number seed. Therefore, a model with a different initial value will be generated each time. Therefore, this time, we aim to improve the reproducibility by fixing the random number seed.
In addition to TensorFlow, we also fix the seeds for NumPy and Python built-in functions. In summary, the following random number fixed functions can be implemented.
import tensorflow as tf
import numpy as np
import random
import os
def set_seed(seed=200):
tf.random.set_seed(seed)
# optional
# for numpy.random
np.random.seed(seed)
# for built-in random
random.seed(seed)
# for hash seed
os.environ["PYTHONHASHSEED"] = str(seed)
It is used as follows. However, if TensorFlow random seeding is sufficient, replace set_seed
with tf.random.set_seed
.
set_seed(0)
toy_model = tf.keras.Sequential(
tf.keras.layers.Dense(2, input_shape=(10,))
)
#Some processing...
#Reproduce the model
set_seed(0)
reproduced_toy_model = tf.keras.Sequential(
tf.keras.layers.Dense(2, input_shape=(10,))
)
reproduced_toy_model
has the same initial value (weight) as the previously generated model toy_model
. In other words, it has been reproduced.
If you do not use set_seed
, reproducible_toy_model
and toy_model
will have completely different initial values, resulting in poor reproducibility.
In addition to tf.keras.Sequential
, you can also use the Functional API and SubClass.
Let's sort out the method of fixing the random number seed (set_seed
) a little more.
tf.random.set_seed
The behavior of tf.random.set_seed
needs a little attention.
First, after using tf.random.set_seed
, try using a function that uses random numbers (tf.random.uniform
: sampling values randomly from a uniform distribution) several times.
tf.random.set_seed(0)
tf.random.uniform([1]) # => [0.29197514]
tf.random.uniform([1]) # => [0.5554141] (Different values!)
tf.random.uniform([1]) # => [0.1952138] (Different values!!)
tf.random.uniform([1]) # => [0.17513537](Different values!!!)
Different values were output for each. It seems that reproducibility will not be possible as it is.
However, use tf.random.set_seed
again as follows.
tf.random.set_seed(0)
tf.random.uniform([1]) # => [0.29197514](A)
tf.random.uniform([1]) # => [0.5554141] (B)
tf.random.set_seed(0)
tf.random.uniform([1]) # => [0.29197514](Reproduction of A)
tf.random.uniform([1]) # => [0.5554141] (Reproduction of B)
In this way, the output is reproduced starting from the place where tf.random.set_seed
is called (even though tf.random.uniform
is a function that outputs a random value).
So, for example, if you call tf.random.set_seed
just before creating a model instance (using Sequential, functional API or SubClass), the generated model will have the same initial value every time.
TensorFlow has layers and functions that allow you to pass seed as an argument.
However, I think that explicitly specifying the initializer argument to be passed to layer or layer is not a very realistic method as the model grows.
In addition, there are some that do not work well unless the tf.random.set_seed
introduced this time is used together.
So, even if you don't have many places to fix, try tf.random.set_seed
first.
In TensorFlow 2.x (tf.keras) you can use tf.random.set_seed
to fix the random seed.
In particular, it will be possible to generate a model with the same initial weight value each time, so improvement in reproducibility can be expected.
Ref