-
Notifications
You must be signed in to change notification settings - Fork 4
/
facial_keypoints.py
109 lines (88 loc) · 3.76 KB
/
facial_keypoints.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
# -*- coding: utf-8 -*-
"""
Created on Wed Oct 18 14:29:53 2017
@author: KAI
"""
import pandas as pd
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
TRAIN_FILE = './facial data/training.csv'
TEST_FILE = './facial data/test.csv'
SAVE_PATH = './facial data/save_CNN.ckpt'
VALIDATION_SIZE = 12000 # 验证集大小
EPOCHS = 1000 # 迭代次数
BATCH_SIZE = 100 # 每个batch大小，稍微大一点的batch会更稳定
BEST_LOSS = 10
Examples_to_show = 6
def input_data(test=False):
file_name = TEST_FILE if test else TRAIN_FILE
df = pd.read_csv(file_name, nrows=VALIDATION_SIZE)
cols = df.columns[:-1]
# dropna()是丢弃有缺失数据的样本，这样最后7000多个样本只剩2140个可用的。
df = df.dropna()
df['Image'] = df['Image'].apply(lambda img: np.fromstring(img, sep=' ') / 255.0)
X = np.vstack(df['Image'])
X = X.reshape((-1, 96, 96, 1)) # (batch height,width,channel)
if test:
y = None
else:
y = df[cols].values / 96.0 # 将y值缩放到[0,1]区间 30维
return X, y
def CNN(tf_x, tf_y):
cnn1 = tf.layers.conv2d(
inputs=tf_x,
filters=16, # shape(96,96,16)
kernel_size=5,
strides=1,
padding='same',
activation=tf.nn.relu)
pool1 = tf.layers.max_pooling2d(
cnn1,
pool_size=2,
strides=2) # 池化=压缩->shape(48,48,16)
cnn2 = tf.layers.conv2d(pool1, 32, 5, 1, 'same', activation=tf.nn.relu) # (48,48,32)
pool2 = tf.layers.max_pooling2d(cnn2, 2, 2) # (24,24,32)
cnn3 = tf.layers.conv2d(pool2, 64, 3, 1, 'same', activation=tf.nn.relu) # (24,24,64)
pool3 = tf.layers.max_pooling2d(cnn3, 2, 2) # (12,12,64)
cnn4 = tf.layers.conv2d(pool3, 128, 3, 1, 'same', activation=tf.nn.relu) # (12,12,128)
pool4 = tf.layers.max_pooling2d(cnn4, 2, 2) # (6,6,128)
output_previous = tf.layers.dense(tf.reshape(pool4, [-1, 6 * 6 * 128]), 500)
# output_second=tf.layers.dense(output_previous,500)
output = tf.layers.dense(output_previous, 30)
loss = tf.losses.mean_squared_error(tf_y, output)
# loss=tf.losses.sparse_softmax_cross_entropy(labels=tf_y,logits=output)
# accuracy=tf.metrics.accuracy(labels=tf_y,predictions=tf.argmax(output,axis=1),)[1]
return output, loss
if __name__ == '__main__':
X, y = input_data()
tf_x = tf.placeholder(tf.float32, [None, 96, 96, 1])
tf_y = tf.placeholder(tf.float32, [None, 30])
output_train, loss_train = CNN(tf_x, tf_y)
train_op = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(loss_train)
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for step in range(EPOCHS):
xs, ys = X[step:step + BATCH_SIZE], y[step:step + BATCH_SIZE]
_, r_loss, pred = sess.run([train_op, loss_train, output_train], feed_dict={tf_x: xs, tf_y: ys})
if r_loss < BEST_LOSS:
saver.save(sess=sess, save_path=SAVE_PATH)
BEST_LOSS = r_loss
if step % 50 == 0:
print('loss is %.4f' % r_loss)
print('BEST_LOSS is %.4f' % BEST_LOSS)
# 最后生成提交结果的时候要用到
# TEST
# X_test,y_test=input_data(True)
# saver.restore(sess=sess,save_path=SAVE_PATH)
# pred_test=sess.run(output_train,feed_dict={tf_x:X_test})
#
# i=np.random.choice(pred_test.shape[0],size=Examples_to_show)
# x_img=pred_test[i,::2]*96.0
# y_img=pred_test[i,1::2]*96.0
# _, a = plt.subplots(1, Examples_to_show, figsize=(Examples_to_show, 1))
# for j in range(Examples_to_show):
# a[j].imshow(np.reshape(X_test[i][:][j], (96, 96)))
# a[j].scatter(x_img[j],y_img[j],c='r')
# plt.show()