tensorflow真是一个我绕不开的坑（苍天饶过谁.jpg）

其实tensorflow1和2的差别挺大的，暂时从1入坑，2的话之后简单过一下。

tf2中更改的函数（供参考）：https://docs.google.com/spreadsheets/d/1FLFJLzg7WNP6JHODX5q8BDgptKafq_slHpnHVbJIteQ/edit#gid=0

本文仅记录我的踩坑历程。

参考文献：https://www.datacamp.com/community/tutorials/tensorflow-tutorial

数据来源：https://btsd.ethz.ch/shareddata/

---

基础知识部分另外编写，这里只记录操作和结果。

import skimage
import tensorflow as tf
from skimage import io # [MUST] for skimage.io.imread
import os
import matplotlib.pyplot as plt # draw distribution graph
from skimage import transform
from skimage.color import rgb2gray # convert img to grayscale
import numpy as np

def first_try():
# initialize constant
x1 = tf.constant([1,2,3,4])
x2 = tf.constant([5,6,7,8])
# multiply
result = tf.multiply(x1, x2)
# only return a tensor, not real-value
# that means: tf does not calculate. only deprive a graph
print(result) # Tensor("Mul:0", shape=(4,), dtype=int32)
# run result and print. 'with' will close automatically
#sess = tf.Session()
#print(sess.run(result))
#sess.close()
with tf.Session() as sess:
output = sess.run(result)
print(output)

def load_data(data_dir):
dirs = [d for d in os.listdir(data_dir)
if os.path.isdir(os.path.join(data_dir, d))]
labels = []
images = []
# each type of sign
for d in dirs:
# .ppm 's file name
label_dir = os.path.join(data_dir, d)
# real path of .ppm
file_names = [os.path.join(label_dir, f)
for f in os.listdir(label_dir)
if f.endswith(".ppm")]
for f in file_names:
# load image
images.append(skimage.io.imread(f))
labels.append(int(d))
return images, labels

def random_show(images, name, cmap=None):
for i in range(len(name)):
plt.subplot(1, len(name), i+1)
plt.axis('off')
# add cmap for gray-scaled pic, which set cmap='gray'
# or u'll get wrong color
plt.imshow(images[name[i]], cmap)
plt.subplots_adjust(wspace=0.5)
print("shape: {0}, min: {1}, max: {2}".format(images[name[i]].shape,
images[name[i]].min(),
images[name[i]].max()))
plt.show()

def show_each_label_pic(labels):
uniq_labels = set(labels)
# initialize the figure
plt.figure(figsize=(15, 15))
i = 1
for label in uniq_labels:
# pick the 1st image for each label
image = images[labels.index(label)]
# 8X8, ith
plt.subplot(8, 8, i)
plt.axis('off')
plt.title("Label {0} ({1})".format(label, labels.count(label)))
i += 1
plt.imshow(image) # plot single picture
plt.show()

def transform_img(images, rows, cols):
return [transform.resize(image, (rows, cols)) for image in images]

def to_gray(images):
# need array
return rgb2gray(np.array(images))

if __name__=="__main__":
ROOT_PATH = r"G:/share/testTF"
train_data_dir = ROOT_PATH + "/Training"
images, labels = load_data(train_data_dir)
#print(len(set(labels))) # 62. coz 62 type of traffic signs
#print(len(images)) # 4575
#plt.hist(labels, 63) # draw a bar-graph.
#plt.show()
#random_show(images, [300, 2250, 3650, 4000])
#print(type(images[0])) #
#show_each_label_pic(labels)
images28 = transform_img(images, 28, 28)
#random_show(images28, [300, 2250, 3650, 4000])
gray_images28 = to_gray(images28)
random_show(gray_images28, [300, 2250, 3650, 4000], cmap="gray")

图像：

条形图：

随机查看的四个图：

统计一下每个label有多少个图：

而且这个resize之后数据其实进行了归一化，进到(0,1)了

灰度图怎么样：这里转化成灰度图是因为作者说，当前问题中，颜色在分类时不起作用。这一点我随后会再验证。