Autoencoder_bits/autoencoder_tf2.0.py at main · ngmsoftware/Autoencoder_bits · GitHub

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
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Fri Jul 20 13:03:16 2018

@author: ninguem
"""

import sys
import os

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

tf.executing_eagerly()

BATCH_SIZE = 8
MAX_EPOCH = 2000


lossFnc = tf.keras.losses.mean_squared_error

optimizer = optimizer = tf.keras.optimizers.RMSprop(learning_rate=0.03,  momentum=0.1)
#optimizer = optimizer = tf.keras.optimizers.Adadelta(learning_rate=20.1) # Adam(learning_rate=0.01)
#optimizer = optimizer = tf.keras.optimizers.Adam(learning_rate=0.02)


folderName = './digits/imgs2/'


X_train = np.ndarray((0,28,28,1))

files = os.listdir(folderName)
for file in files:
    name, ext = file.split('.')
    digit, _ = name.split('-')

    if digit == '0' or digit == '1' or digit == '2' or digit == '3' or digit == '4' or digit == '5' or digit == '6' or digit == '7':

        if (ext == 'png'):

            _I = 1.0 - plt.imread(folderName+file).reshape((1,28,28,1))

            X_train = np.concatenate((X_train, _I), 0)


NTrain = X_train.shape[0]

initializer = tf.keras.initializers.RandomNormal(mean=0.0, stddev=2.0)

# 7. build the model
modelEncoder = tf.keras.models.Sequential()
modelEncoder.add( tf.keras.layers.Input(shape=(28, 28, 1)) )
modelEncoder.add(tf.keras.layers.Flatten())
modelEncoder.add(tf.keras.layers.Dense(10, activation='sigmoid', kernel_initializer = initializer ) )
modelEncoder.add(tf.keras.layers.Dense(6, activation='sigmoid', kernel_initializer = initializer ) )


# FEATURES!!!


modelDecoder = tf.keras.models.Sequential()
modelDecoder.add( tf.keras.layers.Input(6))
modelDecoder.add(tf.keras.layers.Dense(28*28*1, activation='linear'))
modelDecoder.add(tf.keras.layers.Reshape((28, 28, 1)))
modelDecoder.add( tf.keras.layers.Conv2D(40, kernel_size=(3,3), activation='linear', padding='same') )
modelDecoder.add(tf.keras.layers.LeakyReLU())
modelDecoder.add( tf.keras.layers.Conv2D(1, kernel_size=(1,1), activation='sigmoid', padding='same') )


# load last
# modelDecoder = tf.keras.models.load_model('decoder')
# modelEncoder = tf.keras.models.load_model('encoder')


factor = np.exp(np.log(50)/MAX_EPOCH)
expo = 0.0006

_loss = []

for epoch in range(MAX_EPOCH):
    print('epoch : %d of %d'%(epoch, MAX_EPOCH))

    p = np.random.permutation(NTrain)

    X_train = X_train[p]

    X_train_dataset = tf.data.Dataset.from_tensor_slices(X_train).batch(batch_size=BATCH_SIZE)

    i = 0

    expo = expo*factor

    for X in X_train_dataset:

        with tf.GradientTape(persistent=True) as tape:


            features = modelEncoder(X)


            #newFeatures = tf.nn.sigmoid(expo*features)
            newFeatures = features


            XHat = modelDecoder(newFeatures)


            #regularizer1 = tf.reduce_mean( (np.transpose(newFeatures-0.5)@(newFeatures-0.5))**2  )
            regularizer1 = tf.reduce_mean( tf.exp(-6*(features-0.5)**2  ) )


            loss = tf.reduce_mean(lossFnc(X, XHat)) + regularizer1*expo
            #loss = tf.reduce_mean(lossFnc(X, XHat))


            #loss = tf.reduce_mean(lossFnc(X, XHat))

        gradsEncoder = tape.gradient(loss, modelEncoder.trainable_variables)
        gradsDecoder = tape.gradient(loss, modelDecoder.trainable_variables)

        optimizer.apply_gradients(zip(gradsEncoder, modelEncoder.trainable_variables))
        optimizer.apply_gradients(zip(gradsDecoder, modelDecoder.trainable_variables))


        _loss.append(loss)

        print('batch : %d of %d (loss = %.4f)'%(i,int(NTrain/BATCH_SIZE),loss))

        i = i + 1


        # if len(_loss)%100 == 0:
        #     plt.cla()
        #     plt.plot(_loss)
        #     plt.plot([0, len(_loss)],[0.001, 0.001],color='k')
        #     plt.axis([0, len(_loss), 0.0, 0.002])
        #     plt.pause(0.000001)


plt.close('all')
for X in X_train:
    features = modelEncoder.predict(X.reshape((1,28,28,1)))

    print(features)

    #newFeatures = tf.nn.sigmoid(expo*features)
    newFeatures = features


    IHat = modelDecoder.predict(newFeatures).reshape((28,28))

    plt.figure()
    plt.subplot(2,2,1);
    plt.imshow(X.reshape((28,28)))
    plt.subplot(2,2,2);
    plt.imshow(IHat)
    plt.subplot(2,1,2);
    plt.imshow(newFeatures)

    plt.pause(0.0001)