如何向图像数据添加 Keras- 高斯噪声

Question

Importing the modules:

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras.layers import GaussianNoise
from tensorflow.keras.datasets import mnist

(X_train, y_train), (X_test, y_test) = mnist.load_data()

Re-scaling the data

X_train = X_train/255
X_test = X_test/255
plt.imshow(X_train[0])

Adding Gaussian Noise with std dev=0.2

sample = GaussianNoise(0.2)
noisey = sample(X_test[0:2],training=True)   #plt.imshow(noisey[0])

Getting Error:

ValueError: Tensor conversion requested dtype float64 for Tensor with dtype float32: 'Tensor("gaussian_noise_4_1/random_normal:0", shape=(2, 28, 28), dtype=float32)'

Answer 1

Type casting is costly, and so Tensorflow doesn't do automatic type casting. As a default, Tensorflow's dtype is float32 , and the dataset you imported has a dtype float64 . You will just have to pass the optional dtype argument to GaussianNoise :

sample = GaussianNoise(0.2, dtype=tf.float64)

Or cast the array:

noisey = sample(X_test[0:2].astype(np.float32),training=True)

I suggest the second one.

Answer 2

I tried this in my localhost for Jupyter Notebook and the following was the result with a warning.

From the warning it's clear that the problem of Type casting which is very costly. You can improve it as:

X_train = X_train.astype('float32') / 255

X_test = X_test.astype('float32') / 255