tensorflow random rotation that follow normal distributions

目錄：　　

以下tensorflow版本爲1.12.0

• 先看個例子

• 再自己實現，利用讀取tfrecord階段進行數據增強

 

先看個例子，來自stackoverflow

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import random

# load mnist
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot = True)

# Tensorflow random angle rotation
input_size = mnist.train.images.shape[1]
side_size = int(np.sqrt(input_size))

dataset = tf.placeholder(tf.float32, [None, input_size])
images = tf.reshape(dataset,(-1, side_size, side_size, 1))

random_angles = tf.random.uniform(shape=(tf.shape(images)[0],), minval=-np.pi/4, maxval=np.pi/4)

rotated_images = tf.contrib.image.transform(
    images,
    tf.contrib.image.angles_to_projective_transforms(
        random_angles, tf.cast(tf.shape(images)[1], tf.float32), tf.cast(tf.shape(images)[2], tf.float32)
    ))

# Run and Print
sess = tf.Session()
result = sess.run(rotated_images, feed_dict = {
    dataset: mnist.train.images,
})

original = np.reshape(mnist.train.images * 255, (-1, side_size, side_size)).astype(
    np.uint8)
rotated = np.reshape(result * 255, (-1, side_size, side_size)).astype(np.uint8)


# Print 10 random samples
fig, axes = plt.subplots(2, 10, figsize = (15, 4.5))
choice = np.random.choice(range(len(mnist.test.labels)), 10)
for k in range(10):
    axes[0][k].set_axis_off()
    axes[0][k].imshow(original[choice[k, ]], interpolation = 'nearest', cmap = 'gray')
    axes[1][k].set_axis_off()
    axes[1][k].imshow(rotated[choice[k, ]], interpolation = 'nearest', cmap = 'gray')

plt.show()

效果如下

 

再自己實現:

#!/usr/bin/env python
# -*- coding: utf-8 -*-
import os
import numpy as np
import tensorflow as tf
import random
import matplotlib.pyplot as plt

def read_tfrecord_use_queue_runner(filename, batch_size=32, image_shape=(224, 224, 3)):
    filequeue = tf.train.string_input_producer([filename])
    reader = tf.TFRecordReader()
    _, example_tensor = reader.read(filequeue)
    example_features = tf.parse_single_example(
        example_tensor,
        features={
            'image/label': tf.FixedLenFeature([], dtype=tf.string),
            'image/height': tf.FixedLenFeature([], dtype=tf.int64),
            'image/width': tf.FixedLenFeature([], dtype=tf.int64),
            'image/encoded': tf.FixedLenFeature([], dtype=tf.string)
        }
    )
    height = tf.cast(example_features['image/height'], tf.int32)
    width = tf.cast(example_features['image/width'], tf.int32)
    label = tf.cast(example_features['image/label'], tf.string)
    image = tf.image.decode_jpeg(example_features['image/encoded'], channels=image_shape[2])
    image = tf.reshape(image, tf.stack(image_shape))

    min_after_dequeue = 2000
    capacity = min_after_dequeue + 3 * batch_size
    image_batch, label_batch = tf.train.shuffle_batch(
        [image, label],
        batch_size=batch_size,
        min_after_dequeue=min_after_dequeue,
        capacity=capacity,
        num_threads=8
    )

    #### 以下爲數據增強部分

    # 隨機旋轉   
    random_angles = tf.random.uniform(shape=(tf.shape(image_batch)[0],), minval=-np.pi/6, maxval=np.pi/6)
    rotated_images = tf.contrib.image.transform(
        image_batch,
        tf.contrib.image.angles_to_projective_transforms(
            random_angles, tf.cast(tf.shape(image_batch)[1], tf.float32), tf.cast(tf.shape(image_batch)[2], tf.float32)
        ))

    # 在[lower, upper]的範圍隨機調整圖的對比度。
    # rotated_images = tf.image.random_contrast(rotated_images, 0.1, 0.6)

    # 在[lower, upper]的範圍隨機調整圖的飽和度。
    # rotated_images = tf.image.random_saturation(rotated_images, 0, 5)

    # 在[-max_delta, max_delta)的範圍隨機調整圖片的亮度。
    # rotated_images = tf.image.random_brightness(rotated_images, max_delta=0.5)

    return rotated_images, label_batch

def main():

    tfrecord_file = 'validation.tfrecord'
    batch_tensor_dict = read_tfrecord_use_queue_runner(tfrecord_file, batch_size=10, image_shape=(96, 200, 3))

    sess = tf.Session()
    init_op = tf.group(tf.local_variables_initializer(), tf.global_variables_initializer())
    sess.run(init_op)
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)
    max_time_steps = 1
    try:
        for step in range(max_time_steps):
            if coord.should_stop():
                break
            images, labels = sess.run(batch_tensor_dict)
            for i, (image, label) in enumerate(zip(images, labels)):
                # image = np.squeeze(image, axis=2)
                plt.subplot(5, 2, i + 1)
                plt.imshow(image)
                print(label)
            plt.show()

    except tf.errors.OutOfRangeError():
        print('Done training')
    finally:
        coord.request_stop()  # send stop message
        coord.join(threads)  # wait for all

    sess.close()
    exit()


if __name__ == "__main__":
    main()

 

增強的過程中，可以使得隨機的角度在某個範圍以內，並服從正態分佈

例如batch爲128，可以隨機角度在(-30°-30°)之間，並集中旋轉在0度附近服從正態分佈，如：

如何生成正態分佈代碼可以參照如下:

import tensorflow as tf
import matplotlib.pyplot as plt
w1 = tf.Variable(tf.random_normal([128,], mean=0, stddev=10, seed=1))
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    sw1 = sess.run(w1)
    plt.hist(sw1, bins=100,  normed=True)
    plt.show()