はじめに
AIを勉強して学んだ技術で簡単なアプリを作ってみようと思い、
以下のようなものを作ってみました。
何故アンパンマンにしたかというと、簡単に書ける漫画キャラという事で選びました。
モデルはGANを扱ってみたかったのと、正常画像のみの学習で異常検知ができるという事で
ANOGANにしようと思いましたが、調べてみるとANOGANの高速版でEfficientGANという
ものがあるらしいので、それにしました。
また、シンプルにするためアンパンマンの顔だけを判別する前提で作成しました。
流れ
1. AIにアンパンマンの正常画像を学習させる。
2. 正常画像群と異常画像群をAIに入力し、それぞれのスコアの平均の中間値を
異常画像を判別する閾値に設定。
3. 実際に手書き画像を入力し、AIにその絵が異常かどうか判別させる。
やったこと
データセット作成
・学習用画像
ネットからスクレイピングでアンパンマン画像を集め、画像を加工して顔だけ切り取りました。
また後述の通り携帯撮影画像の背景がグレーになっていたので、背景も学習させるため
下記関数を用いて、グレーのグラデーションのデータ拡張を行いました。
from PIL import Image, ImageOps
import numpy as np
def make_gray_gradation(img, gradation_range=(230, 255)):
"""
入力画像の背景をランダムな度合いのグレイグラデーションに変換する
Input : 画像ファイル(カラーでも可)
Output : 画像ファイル(背景がグレイグラデーション変換された画像)
Pramater
img : 入力画像
gradation_range : グラデーションするRGB値の範囲
"""
gra_range = np.random.randint(*gradation_range)
gray = ImageOps.grayscale(img)
output = ImageOps.colorize(gray, black=(0, 0, 0), white=(gra_range, gra_range, gra_range))
return output
・正常画像群と異常画像群
正常画像は、イラスト画像を透かして画用紙に書いたアンパンマンと、
私と学友が書いたアンパンマンを携帯で撮影した画像を用いました。
異常画像は、バイキンマンやドキンちゃん等のイラスト画像を上記同様に携帯で撮影した画像と
ネットでスクレイピングした下手なアンパンマン画像を用いました。
学習後、正常画像群と異常画像群の入力、生成画像比較、スコア分布
携帯で撮影した入力画像の背景がグレーになっていたので、学習用画像にグレーのグラデーションも
入れたのですが、うまく再現できませんでした。
スコアもアンパンマンの絵がうまく書けているかよりも、背景がうまく生成できているかどうかで
決まっているようで、異常画像を判別するための閾値がうまく決まりませんでした。
対策 2値化
対策として、すべての画像について背景は全て真っ白にして、絵の輪郭がアンパンマンに似ているかだけで
異常な絵かどうか判別できるようにしました。以下は入力画像を2値化する関数です。
import os
import scipy.stats as stats
from PIL import Image
def image_binarization(path_in, path_out, th_zero_num=1400, width=100, height=100):
"""
入力画像の輪郭を白黒で2値化して出力する。
Input : 画像ファイルが保存されているフォルダパス(終わりは/) (フォルダには画像以外入れない)
Output : 2値化後の画像を指定フォルダに保存。 2値化後の0(輪郭線)のドット数を出力。
Pramater
path_in : 入力画像群が入ったディレクトリパス
path_out : 出力ディレクトリパス
th_zero_num : 画像の0(輪郭線)のドット数のMIN値(輪郭が濃過ぎる時は小さくして調整)
width : 画像の横幅サイズ
height : 画像の縦幅サイズ
"""
list_in = os.listdir(path_in)
im_np_out = np.empty((0, width*height))
for img in list_in:
path_name = path_in+img
x_img = cv2.imread(path_name)
x_img = cv2.resize(x_img, (width, height))
x_img= cv2.cvtColor(x_img, cv2.COLOR_BGR2GRAY)
x_img = np.array(x_img)
x_img = x_img / 255.0
x_img = x_img.reshape((1, width, height))
x_img = x_img.reshape(1, width*height)
m = stats.mode(x_img)
max_hindo = m.mode[0][0]
for c in reversed(range(50)):
th = (c+1)*0.01
th_0_1 = max_hindo-th
x_img_ = np.where(x_img>th_0_1, 1, 0)
if (np.count_nonzero(x_img_ == 0))>th_zero_num:
break
display(np.count_nonzero(x_img_ == 0))
x_img = x_img_.reshape(width, height)
x_img = (x_img * 2.0) - 1.0
img_np_255 = (x_img + 1.0) * 127.5
img_np_255_mod1 = np.maximum(img_np_255, 0)
img_np_255_mod1 = np.minimum(img_np_255_mod1, 255)
img_np_uint8 = img_np_255_mod1.astype(np.uint8)
image = Image.fromarray(img_np_uint8)
image.save(path_out+img, quality=95)
対策後、正常画像群と異常画像群の入力、生成画像比較、スコア分布
正解画像と異常画像で、ある程度スコアの分布が分かれましたので、とりあえずざっくりと
正常、異常を分ける閾値を決められそうです。(閾値は0.40372に決められていました)
入力したアンパンマン画像が正常な絵か、異常な絵か判別
入力画像
5枚だけイラスト画像を透かして書いたアンパンマン画像を入力し、
他はネットでスクレイピングした下手なアンパンマンを14枚を入力しました。
(上記の2値化関数を用いて2値化してから入力)
判別結果
判別結果はスコア表示枠の背景の色で表されます。背景が青なら正常画像、赤なら異常画像です。
結果は、イラストを透かしたアンパンマンは4/5が正常判定。
他のスクレイピングした下手なアンパンマンは8/14が異常判定で
正解率12/19＝63.15%でした。もっと精度上げないといけないですね。
学んだこと
・真っ白な画用紙に描いた絵を撮影したにもかかわらず、実際の画像は背景がグレーだったので
光の影響の大きさを感じたとともに、2値化等やり方を工夫して、より学習しやすい条件に限定する事で
精度を向上出来る事を学びました。
・GANの精度を向上させるために、GrobalAvaragePoolingやLeakyReLuやレイヤーにノイズを入れる等
いろいろ精度向上手法を試せたのが良かったと思います。(結果はあまり改善しませんでしたが)
今後
GANの精度向上策をもっといろいろ調べて試していきたいですね。
また、GoogleColabやAWSのEC2を使用していたのですが、今後AWSのSageMakerやGCP等いろいろな
クラウドを使用して勉強していきたいと思っています。
コード
train_BiGAN.py
train_BiGAN.py
import numpy as np
import os
import tensorflow as tf
import utility as Utility
import argparse
import matplotlib.pyplot as plt
from model_BiGAN import BiGAN as Model
from make_datasets_TRAIN import Make_datasets_TRAIN as Make_datasets
def parser():
parser = argparse.ArgumentParser(description='train LSGAN')
parser.add_argument('--batch_size', '-b', type=int, default=300, help='Number of images in each mini-batch')
parser.add_argument('--log_file_name', '-lf', type=str, default='anpanman', help='log file name')
parser.add_argument('--epoch', '-e', type=int, default=1001, help='epoch')
parser.add_argument('--file_train_data', '-ftd', type=str, default='../Train_Data/191103/', help='train data')
parser.add_argument('--test_true_data', '-ttd', type=str, default='../Valid_True_Data/191103/', help='test of true_data')
parser.add_argument('--test_false_data', '-tfd', type=str, default='../Valid_False_Data/191103/', help='test of false_data')
parser.add_argument('--valid_span', '-vs', type=int, default=100, help='validation span')
return parser.parse_args()
args = parser()
# global variants
BATCH_SIZE = args.batch_size
LOGFILE_NAME = args.log_file_name
EPOCH = args.epoch
FILE_NAME = args.file_train_data
TRUE_DATA = args.test_true_data
FALSE_DATA = args.test_false_data
IMG_WIDTH = 100
IMG_HEIGHT = 100
IMG_CHANNEL = 1
BASE_CHANNEL = 32
NOISE_UNIT_NUM = 200
NOISE_MEAN = 0.0
NOISE_STDDEV = 1.0
TEST_DATA_SAMPLE = 5 * 5
L2_NORM = 0.001
KEEP_PROB_RATE = 0.5
SEED = 1234
SCORE_ALPHA = 0.9 # using for cost function
VALID_SPAN = args.valid_span
np.random.seed(seed=SEED)
BOARD_DIR_NAME = './tensorboard/' + LOGFILE_NAME
OUT_IMG_DIR = './out_images_BiGAN' #output image file
out_model_dir = './out_models_BiGAN/' #output model_ckpt file
# Load_model_dir = '../model_ckpt/' #Load model_ckpt file
OUT_HIST_DIR = './out_score_hist_BiGAN' #output histogram file
CYCLE_LAMBDA = 1.0
try:
os.mkdir('log')
os.mkdir('out_graph')
os.mkdir(OUT_IMG_DIR)
os.mkdir(out_model_dir)
os.mkdir(OUT_HIST_DIR)
os.mkdir('./out_images_Debug') #for debug
except:
pass
make_datasets = Make_datasets(FILE_NAME, TRUE_DATA, FALSE_DATA, IMG_WIDTH, IMG_HEIGHT, SEED)
model = Model(NOISE_UNIT_NUM, IMG_CHANNEL, SEED, BASE_CHANNEL, KEEP_PROB_RATE)
z_ = tf.placeholder(tf.float32, [None, NOISE_UNIT_NUM], name='z_') #noise to generator
x_ = tf.placeholder(tf.float32, [None, IMG_HEIGHT, IMG_WIDTH, IMG_CHANNEL], name='x_') #image to classifier
d_dis_f_ = tf.placeholder(tf.float32, [None, 1], name='d_dis_g_') #target of discriminator related to generator
d_dis_r_ = tf.placeholder(tf.float32, [None, 1], name='d_dis_r_') #target of discriminator related to real image
is_training_ = tf.placeholder(tf.bool, name = 'is_training')
with tf.variable_scope('encoder_model'):
z_enc = model.encoder(x_, reuse=False, is_training=is_training_)
with tf.variable_scope('decoder_model'):
x_dec = model.decoder(z_, reuse=False, is_training=is_training_)
x_z_x = model.decoder(z_enc, reuse=True, is_training=is_training_) # for cycle consistency
with tf.variable_scope('discriminator_model'):
#stream around discriminator
drop3_r, logits_r = model.discriminator(x_, z_enc, reuse=False, is_training=is_training_) #real pair
drop3_f, logits_f = model.discriminator(x_dec, z_, reuse=True, is_training=is_training_) #real pair
drop3_re, logits_re = model.discriminator(x_z_x, z_enc, reuse=True, is_training=is_training_) #fake pair
with tf.name_scope("loss"):
loss_dis_f = tf.reduce_mean(tf.square(logits_f - d_dis_f_), name='Loss_dis_gen') #loss related to generator
loss_dis_r = tf.reduce_mean(tf.square(logits_r - d_dis_r_), name='Loss_dis_rea') #loss related to real image
#total loss
loss_dis_total = loss_dis_f + loss_dis_r
loss_dec_total = loss_dis_f
loss_enc_total = loss_dis_r
with tf.name_scope("score"):
l_g = tf.reduce_mean(tf.abs(x_ - x_z_x), axis=(1,2,3))
l_FM = tf.reduce_mean(tf.abs(drop3_r - drop3_re), axis=1)
score_A = SCORE_ALPHA * l_g + (1.0 - SCORE_ALPHA) * l_FM
with tf.name_scope("optional_loss"):
loss_dec_opt = loss_dec_total + CYCLE_LAMBDA * l_g
loss_enc_opt = loss_enc_total + CYCLE_LAMBDA * l_g
tf.summary.scalar('loss_dis_total', loss_dis_total)
tf.summary.scalar('loss_dec_total', loss_dec_total)
tf.summary.scalar('loss_enc_total', loss_enc_total)
merged = tf.summary.merge_all()
# t_vars = tf.trainable_variables()
dec_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="decoder")
enc_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="encoder")
dis_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="discriminator")
with tf.name_scope("train"):
train_dis = tf.train.AdamOptimizer(learning_rate=0.0001, beta1=0.5).minimize(loss_dis_total, var_list=dis_vars
, name='Adam_dis')
train_dec = tf.train.AdamOptimizer(learning_rate=0.01, beta1=0.5).minimize(loss_dec_total, var_list=dec_vars
, name='Adam_dec')
train_enc = tf.train.AdamOptimizer(learning_rate=0.005, beta1=0.5).minimize(loss_enc_total, var_list=enc_vars
, name='Adam_enc')
train_dec_opt = tf.train.AdamOptimizer(learning_rate=0.005, beta1=0.5).minimize(loss_dec_opt, var_list=dec_vars
, name='Adam_dec')
train_enc_opt = tf.train.AdamOptimizer(learning_rate=0.005, beta1=0.5).minimize(loss_enc_opt, var_list=enc_vars
, name='Adam_enc')
sess = tf.Session()
ckpt = tf.train.get_checkpoint_state(out_model_dir)
saver = tf.train.Saver()
if ckpt: # checkpointがある場合
last_model = ckpt.model_checkpoint_path # 最後に保存したmodelへのパス
saver.restore(sess, last_model) # 変数データの読み込み
print("load " + last_model)
else: # 保存データがない場合
#init = tf.initialize_all_variables()
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(BOARD_DIR_NAME, sess.graph)
log_list = []
log_list.append(['epoch', 'AUC'])
# training loop
for epoch in range(0, EPOCH):
sum_loss_dis_f = np.float32(0)
sum_loss_dis_r = np.float32(0)
sum_loss_dis_total = np.float32(0)
sum_loss_dec_total = np.float32(0)
sum_loss_enc_total = np.float32(0)
len_data = make_datasets.make_data_for_1_epoch()
for i in range(0, len_data, BATCH_SIZE):
img_batch = make_datasets.get_data_for_1_batch(i, BATCH_SIZE)
z = make_datasets.make_random_z_with_norm(NOISE_MEAN, NOISE_STDDEV, len(img_batch), NOISE_UNIT_NUM)
tar_g_1 = make_datasets.make_target_1_0(1.0, len(img_batch)) #1 -> real
tar_g_0 = make_datasets.make_target_1_0(0.0, len(img_batch)) #0 -> fake
#train discriminator
sess.run(train_dis, feed_dict={z_:z, x_: img_batch, d_dis_f_: tar_g_0, d_dis_r_: tar_g_1, is_training_:True})
#train decoder
sess.run(train_dec, feed_dict={z_:z, d_dis_f_: tar_g_1, is_training_:True})
# sess.run(train_dec_opt, feed_dict={z_:z, x_: img_batch, d_dis_f_: tar_g_1, is_training_:True})
#train encoder
sess.run(train_enc, feed_dict={x_:img_batch, d_dis_r_: tar_g_0, is_training_:True})
# sess.run(train_enc_opt, feed_dict={x_:img_batch, d_dis_r_: tar_g_0, is_training_:True})
# loss for discriminator
loss_dis_total_, loss_dis_r_, loss_dis_f_ = sess.run([loss_dis_total, loss_dis_r, loss_dis_f],
feed_dict={z_: z, x_: img_batch, d_dis_f_: tar_g_0,
d_dis_r_: tar_g_1, is_training_:False})
#loss for decoder
loss_dec_total_ = sess.run(loss_dec_total, feed_dict={z_: z, d_dis_f_: tar_g_1, is_training_:False})
#loss for encoder
loss_enc_total_ = sess.run(loss_enc_total, feed_dict={x_: img_batch, d_dis_r_: tar_g_0, is_training_:False})
#for tensorboard
merged_ = sess.run(merged, feed_dict={z_:z, x_: img_batch, d_dis_f_: tar_g_0, d_dis_r_: tar_g_1, is_training_:False})
summary_writer.add_summary(merged_, epoch)
sum_loss_dis_f += loss_dis_f_
sum_loss_dis_r += loss_dis_r_
sum_loss_dis_total += loss_dis_total_
sum_loss_dec_total += loss_dec_total_
sum_loss_enc_total += loss_enc_total_
print("----------------------------------------------------------------------")
print("epoch = {:}, Encoder Total Loss = {:.4f}, Decoder Total Loss = {:.4f}, Discriminator Total Loss = {:.4f}".format(
epoch, sum_loss_enc_total / len_data, sum_loss_dec_total / len_data, sum_loss_dis_total / len_data))
print("Discriminator Real Loss = {:.4f}, Discriminator Generated Loss = {:.4f}".format(
sum_loss_dis_r / len_data, sum_loss_dis_r / len_data))
if epoch % VALID_SPAN == 0:
# score_A_list = []
score_A_np = np.zeros((0, 2), dtype=np.float32)
val_data_num = len(make_datasets.valid_data)
val_true_data_num = len(make_datasets.valid_true_np)
val_false_data_num = len(make_datasets.valid_false_np)
img_batch_1, _ = make_datasets.get_valid_data_for_1_batch(0, val_true_data_num)
img_batch_0, _ = make_datasets.get_valid_data_for_1_batch(val_data_num - val_false_data_num, val_true_data_num)
x_z_x_1 = sess.run(x_z_x, feed_dict={x_:img_batch_1, is_training_:False})
x_z_x_0 = sess.run(x_z_x, feed_dict={x_:img_batch_0, is_training_:False})
score_A_1 = sess.run(score_A, feed_dict={x_:img_batch_1, is_training_:False})
score_A_0 = sess.run(score_A, feed_dict={x_:img_batch_0, is_training_:False})
score_A_re_1 = np.reshape(score_A_1, (-1, 1))
score_A_re_0 = np.reshape(score_A_0, (-1, 1))
tars_batch_1 = np.ones(val_true_data_num)
tars_batch_0 = np.zeros(val_false_data_num)
tars_batch_re_1 = np.reshape(tars_batch_1, (-1, 1))
tars_batch_re_0 = np.reshape(tars_batch_0, (-1, 1))
score_A_np_1_tmp = np.concatenate((score_A_re_1, tars_batch_re_1), axis=1)
score_A_np_0_tmp = np.concatenate((score_A_re_0, tars_batch_re_0), axis=1)
score_A_np = np.concatenate((score_A_np_1_tmp, score_A_np_0_tmp), axis=0)
#print(score_A_np)
tp, fp, tn, fn, precision, recall = Utility.compute_precision_recall(score_A_np)
auc = Utility.make_ROC_graph(score_A_np, 'out_graph/' + LOGFILE_NAME, epoch)
print("tp:{}, fp:{}, tn:{}, fn:{}, precision:{:.4f}, recall:{:.4f}, AUC:{:.4f}".format(tp, fp, tn, fn, precision, recall, auc))
log_list.append([epoch, auc])
Utility.make_score_hist(score_A_1, score_A_0, epoch, LOGFILE_NAME, OUT_HIST_DIR)
Utility.make_output_img(img_batch_1, img_batch_0, x_z_x_1, x_z_x_0, score_A_0, score_A_1, epoch, LOGFILE_NAME, OUT_IMG_DIR)
#after learning
Utility.save_list_to_csv(log_list, 'log/' + LOGFILE_NAME + '_auc.csv')
# saver2 = tf.train.Saver()
save_path = saver.save(sess, out_model_dir + 'anpanman_weight.ckpt')
print("Model saved in file: ", save_path)
model_BiGAN.py
model_BiGAN.py
import numpy as np
# import os
import tensorflow as tf
# from PIL import Image
# import utility as Utility
# import argparse
class BiGAN():
def __init__(self, noise_unit_num, img_channel, seed, base_channel, keep_prob):
self.NOISE_UNIT_NUM = noise_unit_num # 200
self.IMG_CHANNEL = img_channel # 1
self.SEED = seed
np.random.seed(seed=self.SEED)
self.BASE_CHANNEL = base_channel # 32
self.KEEP_PROB = keep_prob
def leaky_relu(self, x, alpha):
return tf.nn.relu(x) - alpha * tf.nn.relu(-x)
def gaussian_noise(self, input, std): #used at discriminator
noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=std, dtype=tf.float32, seed=self.SEED)
return input + noise
def conv2d(self, input, in_channel, out_channel, k_size, stride, seed):
w = tf.get_variable('w', [k_size, k_size, in_channel, out_channel],
initializer=tf.random_normal_initializer
(mean=0.0, stddev=0.02, seed=seed), dtype=tf.float32)
b = tf.get_variable('b', [out_channel], initializer=tf.constant_initializer(0.0))
conv = tf.nn.conv2d(input, w, strides=[1, stride, stride, 1], padding="SAME", name='conv') + b
return conv
def conv2d_transpose(self, input, in_channel, out_channel, k_size, stride, seed):
w = tf.get_variable('w', [k_size, k_size, out_channel, in_channel],
initializer=tf.random_normal_initializer
(mean=0.0, stddev=0.02, seed=seed), dtype=tf.float32)
b = tf.get_variable('b', [out_channel], initializer=tf.constant_initializer(0.0))
out_shape = tf.stack(
[tf.shape(input)[0], tf.shape(input)[1] * 2, tf.shape(input)[2] * 2, tf.constant(out_channel)])
deconv = tf.nn.conv2d_transpose(input, w, output_shape=out_shape, strides=[1, stride, stride, 1],
padding="SAME") + b
return deconv
def batch_norm(self, input):
shape = input.get_shape().as_list()
n_out = shape[-1]
scale = tf.get_variable('scale', [n_out], initializer=tf.constant_initializer(1.0))
beta = tf.get_variable('beta', [n_out], initializer=tf.constant_initializer(0.0))
batch_mean, batch_var = tf.nn.moments(input, [0])
bn = tf.nn.batch_normalization(input, batch_mean, batch_var, beta, scale, 0.0001, name='batch_norm')
return bn
def fully_connect(self, input, in_num, out_num, seed):
w = tf.get_variable('w', [in_num, out_num], initializer=tf.random_normal_initializer
(mean=0.0, stddev=0.02, seed=seed), dtype=tf.float32)
b = tf.get_variable('b', [out_num], initializer=tf.constant_initializer(0.0))
fc = tf.matmul(input, w, name='fc') + b
return fc
def encoder(self, x, reuse=False, is_training=False): #x is expected [n, 28, 28, 1]
with tf.variable_scope('encoder', reuse=reuse):
with tf.variable_scope("layer1"): # layer1 conv nx28x28x1 -> nx14x14x32
conv1 = self.conv2d(x, self.IMG_CHANNEL, self.BASE_CHANNEL, 3, 2, self.SEED)
with tf.variable_scope("layer2"): # layer2 conv nx14x14x32 -> nx7x7x64
conv2 = self.conv2d(conv1, self.BASE_CHANNEL, self.BASE_CHANNEL*2, 3, 2, self.SEED)
bn2 = self.batch_norm(conv2)
lr2 = self.leaky_relu(bn2, alpha=0.1)
with tf.variable_scope("layer3"): # layer3 conv nx7x7x64 -> nx4x4x128
conv3 = self.conv2d(lr2, self.BASE_CHANNEL*2, self.BASE_CHANNEL*4, 3, 2, self.SEED)
bn3 = self.batch_norm(conv3)
lr3 = self.leaky_relu(bn3, alpha=0.1)
with tf.variable_scope("layer4"): # layer4 fc nx4x4x128 -> nx200
shape = tf.shape(lr3)
print(shape[1])
reshape4 = tf.reshape(lr3, [shape[0], shape[1]*shape[2]*shape[3]])
fc4 = self.fully_connect(reshape4, 21632, self.NOISE_UNIT_NUM, self.SEED)
return fc4
def decoder(self, z, reuse=False, is_training=False): # z is expected [n, 200]
with tf.variable_scope('decoder', reuse=reuse):
with tf.variable_scope("layer1"): # layer1 fc nx200 -> nx1024
fc1 = self.fully_connect(z, self.NOISE_UNIT_NUM, 1024, self.SEED)
bn1 = self.batch_norm(fc1)
rl1 = tf.nn.relu(bn1)
with tf.variable_scope("layer2"): # layer2 fc nx1024 -> nx6272
fc2 = self.fully_connect(rl1, 1024, 25*25*self.BASE_CHANNEL*4, self.SEED)
bn2 = self.batch_norm(fc2)
rl2 = tf.nn.relu(bn2)
with tf.variable_scope("layer3"): # layer3 deconv nx6272 -> nx7x7x128 -> nx14x14x64
shape = tf.shape(rl2)
reshape3 = tf.reshape(rl2, [shape[0], 25, 25, 128])
deconv3 = self.conv2d_transpose(reshape3, self.BASE_CHANNEL*4, self.BASE_CHANNEL*2, 4, 2, self.SEED)
bn3 = self.batch_norm(deconv3)
rl3 = tf.nn.relu(bn3)
with tf.variable_scope("layer4"): # layer3 deconv nx14x14x64 -> nx28x28x1
deconv4 = self.conv2d_transpose(rl3, self.BASE_CHANNEL*2, self.IMG_CHANNEL, 4, 2, self.SEED)
tanh4 = tf.tanh(deconv4)
return tanh4
def discriminator(self, x, z, reuse=False, is_training=True): #z[n, 200], x[n, 28, 28, 1]
with tf.variable_scope('discriminator', reuse=reuse):
with tf.variable_scope("x_layer1"): # layer x1 conv [n, 28, 28, 1] -> [n, 14, 14, 64]
convx1 = self.conv2d(x, self.IMG_CHANNEL, self.BASE_CHANNEL*2, 4, 2, self.SEED)
lrx1 = self.leaky_relu(convx1, alpha=0.1)
dropx1 = tf.layers.dropout(lrx1, rate=1.0 - self.KEEP_PROB, name='dropout', training=is_training)
with tf.variable_scope("x_layer2"): # layer x2 conv [n, 14, 14, 64] -> [n, 7, 7, 64] -> [n, 3136]
convx2 = self.conv2d(dropx1, self.BASE_CHANNEL*2, self.BASE_CHANNEL*2, 4, 2, self.SEED)
bnx2 = self.batch_norm(convx2)
lrx2 = self.leaky_relu(bnx2, alpha=0.1)
dropx2 = tf.layers.dropout(lrx2, rate=1.0 - self.KEEP_PROB, name='dropout', training=is_training)
shapex2 = tf.shape(dropx2)
reshape3 = tf.reshape(dropx2, [shapex2[0], shapex2[1]*shapex2[2]*shapex2[3]])
with tf.variable_scope("z_layer1"): # layer1 fc [n, 200] -> [n, 512]
fcz1 = self.fully_connect(z, self.NOISE_UNIT_NUM, 512, self.SEED)
lrz1 = self.leaky_relu(fcz1, alpha=0.1)
dropz1 = tf.layers.dropout(lrz1, rate=1.0 - self.KEEP_PROB, name='dropout', training=is_training)
with tf.variable_scope("y_layer3"): # layer1 fc [n, 6272], [n, 1024]
con3 = tf.concat([reshape3, dropz1], axis=1)
fc3 = self.fully_connect(con3, 40000+512, 1024, self.SEED)
lr3 = self.leaky_relu(fc3, alpha=0.1)
self.drop3 = tf.layers.dropout(lr3, rate=1.0 - self.KEEP_PROB, name='dropout', training=is_training)
with tf.variable_scope("y_fc_logits"):
self.logits = self.fully_connect(self.drop3, 1024, 1, self.SEED)
return self.drop3, self.logits
make_datasets_TRAIN.py
make_datasets_TRAIN.py
import numpy as np
import os
import glob
import re
import random
# import cv2
from PIL import Image
from keras.preprocessing import image
class Make_datasets_TRAIN():
def __init__(self, filename, true_data, false_data, img_width, img_height, seed):
self.filename = filename
self.true_data = true_data
self.false_data = false_data
self.img_width = img_width
self.img_height = img_height
self.seed = seed
x_train, x_valid_true, x_valid_false, y_train, y_valid_true, y_valid_false = self.read_DATASET(self.filename, self.true_data, self.false_data)
self.train_np = np.concatenate((y_train.reshape(-1,1), x_train), axis=1).astype(np.float32)
self.valid_true_np = np.concatenate((y_valid_true.reshape(-1,1), x_valid_true), axis=1).astype(np.float32)
self.valid_false_np = np.concatenate((y_valid_false.reshape(-1,1), x_valid_false), axis=1).astype(np.float32)
print("self.train_np.shape, ", self.train_np.shape)
print("self.valid_true_np.shape, ", self.valid_true_np.shape)
print("self.valid_false_np.shape, ", self.valid_false_np.shape)
print("np.max(x_train), ", np.max(x_train))
print("np.min(x_train), ", np.min(x_train))
self.valid_data = np.concatenate((self.valid_true_np, self.valid_false_np))
random.seed(self.seed)
np.random.seed(self.seed)
def read_DATASET(self, train_path, true_path, false_path):
train_list = os.listdir(train_path)
y_train = np.ones(len(train_list))
x_train = np.empty((0, self.img_width*self.img_height))
for img in train_list:
path_name = train_path+img
x_img = Image.open(path_name)
# サイズを揃える
x_img = x_img.resize((self.img_width, self.img_height))
# 3chを1chに変換
x_img= x_img.convert('L')
# PIL.Image.Imageからnumpy配列へ
x_img = np.array(x_img)
# 正規化
x_img = x_img / 255.0
# axisの追加
x_img = x_img.reshape((1,self.img_width, self.img_height))
# flatten
x_img = x_img.reshape(1, self.img_width*self.img_height)
x_train = np.concatenate([x_train, x_img], axis = 0)
print("x_train.shape, ", x_train.shape)
print("y_train.shape, ", y_train.shape)
test_true_list = os.listdir(true_path)
y_test_true = np.ones(len(test_true_list))
x_test_true = np.empty((0, self.img_width*self.img_height))
for img in test_true_list:
path_name = true_path+img
x_img = Image.open(path_name)
x_img = x_img.resize((self.img_width, self.img_height))
x_img= x_img.convert('L')
x_img = np.array(x_img)
x_img = x_img / 255.0
x_img = x_img.reshape((1,self.img_width, self.img_height))
x_img = x_img.reshape(1, self.img_width*self.img_height)
x_test_true = np.concatenate([x_test_true, x_img], axis = 0)
print("x_test_true.shape, ", x_test_true.shape)
print("y_test_true.shape, ", y_test_true.shape)
test_false_list = os.listdir(false_path)
y_test_false = np.zeros(len(test_false_list))
x_test_false = np.empty((0, self.img_width*self.img_height))
for img in test_false_list:
path_name = false_path+img
x_img = Image.open(path_name)
x_img = x_img.resize((self.img_width, self.img_height))
x_img= x_img.convert('L')
x_img = np.array(x_img)
x_img = x_img / 255.0
x_img = x_img.reshape((1,self.img_width, self.img_height))
x_img = x_img.reshape(1, self.img_width*self.img_height)
x_test_false = np.concatenate([x_test_false, x_img], axis = 0)
print("x_test_false.shape, ", x_test_false.shape)
print("y_test_false.shape, ", y_test_false.shape)
return x_train, x_test_true, x_test_false, y_train, y_test_true, y_test_false
def get_file_names(self, dir_name):
target_files = []
for root, dirs, files in os.walk(dir_name):
targets = [os.path.join(root, f) for f in files]
target_files.extend(targets)
return target_files
def read_data(self, d_y_np, width, height):
tars = []
images = []
for num, d_y_1 in enumerate(d_y_np):
image = d_y_1[1:].reshape(width, height, 1)
tar = d_y_1[0]
images.append(image)
tars.append(tar)
return np.asarray(images), np.asarray(tars)
def normalize_data(self, data):
# data0_2 = data / 127.5
# data_norm = data0_2 - 1.0
data_norm = (data * 2.0) - 1.0 #applied for tanh
return data_norm
def make_data_for_1_epoch(self):
self.filename_1_epoch = np.random.permutation(self.train_np)
return len(self.filename_1_epoch)
def get_data_for_1_batch(self, i, batchsize):
filename_batch = self.filename_1_epoch[i:i + batchsize]
images, _ = self.read_data(filename_batch, self.img_width, self.img_height)
images_n = self.normalize_data(images)
return images_n
def get_valid_data_for_1_batch(self, i, batchsize):
filename_batch = self.valid_data[i:i + batchsize]
images, tars = self.read_data(filename_batch, self.img_width, self.img_height)
images_n = self.normalize_data(images)
return images_n, tars
def make_random_z_with_norm(self, mean, stddev, data_num, unit_num):
norms = np.random.normal(mean, stddev, (data_num, unit_num))
# tars = np.zeros((data_num, 1), dtype=np.float32)
return norms
def make_target_1_0(self, value, data_num):
if value == 0.0:
target = np.zeros((data_num, 1), dtype=np.float32)
elif value == 1.0:
target = np.ones((data_num, 1), dtype=np.float32)
else:
print("target value error")
return target
utility.py
utility.py
import numpy as np
# import os
from PIL import Image
import matplotlib.pyplot as plt
import sklearn.metrics as sm
import csv
import seaborn as sns
def compute_precision_recall(score_A_np, ):
array_1 = np.where(score_A_np[:, 1] == 1.0)
array_0 = np.where(score_A_np[:, 1] == 0.0)
mean_1 = np.mean((score_A_np[array_1])[:, 0])
mean_0 = np.mean((score_A_np[array_0])[:, 0])
medium = (mean_1 + mean_0) / 2.0
print("mean_positive_score, ", mean_1)
print("mean_negative_score, ", mean_0)
print("score_threshold(pos_neg middle), ", medium)
np.save('./score_threshold.npy', medium)
array_upper = np.where(score_A_np[:, 0] >= medium)[0]
array_lower = np.where(score_A_np[:, 0] < medium)[0]
#print(array_upper)
print("negative_predict_num, ", array_upper.shape)
print("positive_predict_num, ", array_lower.shape)
array_1_tf = np.where(score_A_np[:, 1] == 1.0)[0]
array_0_tf = np.where(score_A_np[:, 1] == 0.0)[0]
#print(array_1_tf)
print("negative_fact_num, ", array_0_tf.shape)
print("positive_fact_num, ", array_1_tf.shape)
tn = len(set(array_lower)&set(array_1_tf))
tp = len(set(array_upper)&set(array_0_tf))
fp = len(set(array_lower)&set(array_0_tf))
fn = len(set(array_upper)&set(array_1_tf))
precision = tp / (tp + fp + 0.00001)
recall = tp / (tp + fn + 0.00001)
return tp, fp, tn, fn, precision, recall
def score_divide(score_A_np):
array_1 = np.where(score_A_np[:, 1] == 1.0)[0]
array_0 = np.where(score_A_np[:, 1] == 0.0)[0]
print("positive_predict_num, ", array_1.shape)
print("negative_predict_num, ", array_0.shape)
array_1_np = score_A_np[array_1][:, 0]
array_0_np = score_A_np[array_0][:, 0]
#print(array_1_np)
#print(array_0_np)
return array_1_np, array_0_np
def save_graph(x, y, filename, epoch):
plt.figure(figsize=(7, 5))
plt.plot(x, y)
plt.title('ROC curve ' + filename + ' epoch:' + str(epoch))
# x axis label
plt.xlabel("FP / (FP + TN)")
# y axis label
plt.ylabel("TP / (TP + FN)")
# save
plt.savefig(filename + '_ROC_curve_epoch' + str(epoch) +'.png')
plt.close()
def make_ROC_graph(score_A_np, filename, epoch):
argsort = np.argsort(score_A_np, axis=0)[:, 0]
value_1_0 = score_A_np[argsort][::-1].astype(np.float32)
#value_1_0 = (np.where(score_A_np_sort[:, 1] == 7., 1., 0.)).astype(np.float32)
# score_A_np_sort_0_1 = np.concatenate((score_A_np_sort, value_1_0), axis=1)
sum_1 = np.sum(value_1_0)
len_s = len(score_A_np)
sum_0 = len_s - sum_1
tp = np.cumsum(value_1_0[:, 1]).astype(np.float32)
index = np.arange(1, len_s + 1, 1).astype(np.float32)
fp = index - tp
fn = sum_1 - tp
tn = sum_0 - fp
tp_ratio = tp / (tp + fn + 0.00001)
fp_ratio = fp / (fp + tn + 0.00001)
save_graph(fp_ratio, tp_ratio, filename, epoch)
auc = sm.auc(fp_ratio, tp_ratio)
return auc
def unnorm_img(img_np):
img_np_255 = (img_np + 1.0) * 127.5
img_np_255_mod1 = np.maximum(img_np_255, 0)
img_np_255_mod1 = np.minimum(img_np_255_mod1, 255)
img_np_uint8 = img_np_255_mod1.astype(np.uint8)
return img_np_uint8
def convert_np2pil(images_255):
list_images_PIL = []
for num, images_255_1 in enumerate(images_255):
# img_255_tile = np.tile(images_255_1, (1, 1, 3))
image_1_PIL = Image.fromarray(images_255_1)
list_images_PIL.append(image_1_PIL)
return list_images_PIL
def make_score_hist(score_a_1, score_a_0, epoch, LOGFILE_NAME, OUT_HIST_DIR):
list_1 = score_a_1.tolist()
list_0 = score_a_0.tolist()
#print(list_1)
#print(list_0)
plt.figure(figsize=(7, 5))
plt.title("Histgram of Score")
plt.xlabel("Score")
plt.ylabel("freq")
plt.hist(list_1, bins=40, alpha=0.3, histtype='stepfilled', color='r', label="1")
plt.hist(list_0, bins=40, alpha=0.3, histtype='stepfilled', color='b', label='0')
plt.legend(loc=1)
plt.savefig(OUT_HIST_DIR + "/resultScoreHist_"+ LOGFILE_NAME + '_' + str(epoch) + ".png")
plt.show()
def make_score_hist_test(score_a_1, score_a_0, score_th, LOGFILE_NAME, OUT_HIST_DIR):
list_1 = score_a_1.tolist()
list_0 = score_a_0.tolist()
#print(list_1)
#print(list_0)
plt.figure(figsize=(7, 5))
plt.title("Histgram of Score")
plt.xlabel("Score")
plt.ylabel("freq")
plt.hist(list_1, bins=40, alpha=0.3, histtype='stepfilled', color='r', label="1")
plt.hist(list_0, bins=40, alpha=0.3, histtype='stepfilled', color='b', label='0')
plt.legend(loc=1)
plt.savefig(OUT_HIST_DIR + "/resultScoreHist_"+ LOGFILE_NAME + "_test.png")
plt.show()
def make_score_bar(score_a):
score_a = score_a.tolist()
list_images_PIL = []
for score in score_a:
x="score"
plt.bar(x,score,label=score)
fig, ax = plt.subplots(figsize=(1, 1))
ax.bar(x,score,label=round(score,3))
ax.legend(loc='center', fontsize=12)
fig.canvas.draw()
#im = np.array(fig.canvas.renderer.buffer_rgba()) # matplotlibが3.1より以降の場合
im = np.array(fig.canvas.renderer._renderer)
image_1_PIL = Image.fromarray(im)
list_images_PIL.append(image_1_PIL)
return list_images_PIL
def make_score_bar_predict(score_A_np_tmp):
score_a = score_A_np_tmp.tolist()
list_images_PIL = []
for score in score_a:
x="score"
#plt.bar(x,score[0],label=score)
fig, ax = plt.subplots(figsize=(1, 1))
if score[1]==0:
ax.bar(x,score[0], color='red',label=round(score[0],3))
else:
ax.bar(x,score[0], color='blue',label=round(score[0],3))
ax.legend(loc='center', fontsize=12)
fig.canvas.draw()
#im = np.array(fig.canvas.renderer.buffer_rgba()) # matplotlibが3.1より以降の場合
im = np.array(fig.canvas.renderer._renderer)
image_1_PIL = Image.fromarray(im)
list_images_PIL.append(image_1_PIL)
return list_images_PIL
def make_output_img(img_batch_1, img_batch_0, x_z_x_1, x_z_x_0, score_a_0, score_a_1, epoch, log_file_name, out_img_dir):
(data_num, img1_h, img1_w, _) = img_batch_1.shape
img_batch_1_unn = np.tile(unnorm_img(img_batch_1), (1, 1, 3))
img_batch_0_unn = np.tile(unnorm_img(img_batch_0), (1, 1, 3))
x_z_x_1_unn = np.tile(unnorm_img(x_z_x_1), (1, 1, 3))
x_z_x_0_unn = np.tile(unnorm_img(x_z_x_0), (1, 1, 3))
diff_1 = img_batch_1 - x_z_x_1
diff_1_r = (2.0 * np.maximum(diff_1, 0.0)) - 1.0 #(0.0, 1.0) -> (-1.0, 1.0)
diff_1_b = (2.0 * np.abs(np.minimum(diff_1, 0.0))) - 1.0 #(-1.0, 0.0) -> (1.0, 0.0) -> (1.0, -1.0)
diff_1_g = diff_1_b * 0.0 - 1.0
diff_1_r_unnorm = unnorm_img(diff_1_r)
diff_1_b_unnorm = unnorm_img(diff_1_b)
diff_1_g_unnorm = unnorm_img(diff_1_g)
diff_1_np = np.concatenate((diff_1_r_unnorm, diff_1_g_unnorm, diff_1_b_unnorm), axis=3)
diff_0 = img_batch_0 - x_z_x_0
diff_0_r = (2.0 * np.maximum(diff_0, 0.0)) - 1.0 #(0.0, 1.0) -> (-1.0, 1.0)
diff_0_b = (2.0 * np.abs(np.minimum(diff_0, 0.0))) - 1.0 #(-1.0, 0.0) -> (1.0, 0.0) -> (1.0, -1.0)
diff_0_g = diff_0_b * 0.0 - 1.0
diff_0_r_unnorm = unnorm_img(diff_0_r)
diff_0_b_unnorm = unnorm_img(diff_0_b)
diff_0_g_unnorm = unnorm_img(diff_0_g)
diff_0_np = np.concatenate((diff_0_r_unnorm, diff_0_g_unnorm, diff_0_b_unnorm), axis=3)
img_batch_1_PIL = convert_np2pil(img_batch_1_unn)
img_batch_0_PIL = convert_np2pil(img_batch_0_unn)
x_z_x_1_PIL = convert_np2pil(x_z_x_1_unn)
x_z_x_0_PIL = convert_np2pil(x_z_x_0_unn)
diff_1_PIL = convert_np2pil(diff_1_np)
diff_0_PIL = convert_np2pil(diff_0_np)
score_a_1_PIL = make_score_bar(score_a_1)
score_a_0_PIL = make_score_bar(score_a_0)
wide_image_np = np.ones(((img1_h + 1) * data_num - 1, (img1_w + 1) * 8 - 1, 3), dtype=np.uint8) * 255
wide_image_PIL = Image.fromarray(wide_image_np)
for num, (ori_1, ori_0, xzx1, xzx0, diff1, diff0, score_1, score_0) in enumerate(zip(img_batch_1_PIL, img_batch_0_PIL ,x_z_x_1_PIL, x_z_x_0_PIL, diff_1_PIL, diff_0_PIL, score_a_1_PIL, score_a_0_PIL)):
wide_image_PIL.paste(ori_1, (0, num * (img1_h + 1)))
wide_image_PIL.paste(xzx1, (img1_w + 1, num * (img1_h + 1)))
wide_image_PIL.paste(diff1, ((img1_w + 1) * 2, num * (img1_h + 1)))
wide_image_PIL.paste(score_1, ((img1_w + 1) * 3, num * (img1_h + 1)))
wide_image_PIL.paste(ori_0, ((img1_w + 1) * 4, num * (img1_h + 1)))
wide_image_PIL.paste(xzx0, ((img1_w + 1) * 5, num * (img1_h + 1)))
wide_image_PIL.paste(diff0, ((img1_w + 1) * 6, num * (img1_h + 1)))
wide_image_PIL.paste(score_0, ((img1_w + 1) * 7, num * (img1_h + 1)))
wide_image_PIL.save(out_img_dir + "/resultImage_"+ log_file_name + '_' + str(epoch) + ".png")
def make_output_img_test(img_batch_test, x_z_x_test, score_A_np_tmp, log_file_name, out_img_dir):
(data_num, img1_h, img1_w, _) = img_batch_test.shape
img_batch_test_unn = np.tile(unnorm_img(img_batch_test), (1, 1, 3))
x_z_x_test_unn = np.tile(unnorm_img(x_z_x_test), (1, 1, 3))
diff_test = img_batch_test - x_z_x_test
diff_test_r = (2.0 * np.maximum(diff_test, 0.0)) - 1.0 #(0.0, 1.0) -> (-1.0, 1.0)
diff_test_b = (2.0 * np.abs(np.minimum(diff_test, 0.0))) - 1.0 #(-1.0, 0.0) -> (1.0, 0.0) -> (1.0, -1.0)
diff_test_g = diff_test_b * 0.0 - 1.0
diff_test_r_unnorm = unnorm_img(diff_test_r)
diff_test_b_unnorm = unnorm_img(diff_test_b)
diff_test_g_unnorm = unnorm_img(diff_test_g)
diff_test_np = np.concatenate((diff_test_r_unnorm, diff_test_g_unnorm, diff_test_b_unnorm), axis=3)
img_batch_test_PIL = convert_np2pil(img_batch_test_unn)
x_z_x_test_PIL = convert_np2pil(x_z_x_test_unn)
diff_test_PIL = convert_np2pil(diff_test_np)
score_a = score_A_np_tmp[:, 1:]
#tars = score_A_np_tmp[:, 0]
score_a_PIL = make_score_bar_predict(score_A_np_tmp)
wide_image_np = np.ones(((img1_h + 1) * data_num - 1, (img1_w + 1) * 8 - 1, 3), dtype=np.uint8) * 255
wide_image_PIL = Image.fromarray(wide_image_np)
for num, (ori_test, xzx_test, diff_test, score_test) in enumerate(zip(img_batch_test_PIL, x_z_x_test_PIL, diff_test_PIL, score_a_PIL)):
wide_image_PIL.paste(ori_test, (0, num * (img1_h + 1)))
wide_image_PIL.paste(xzx_test, (img1_w + 1, num * (img1_h + 1)))
wide_image_PIL.paste(diff_test, ((img1_w + 1) * 2, num * (img1_h + 1)))
wide_image_PIL.paste(score_test, ((img1_w + 1) * 3, num * (img1_h + 1)))
wide_image_PIL.save(out_img_dir + "/resultImage_"+ log_file_name + "_test.png")
def save_list_to_csv(list, filename):
f = open(filename, 'w')
writer = csv.writer(f, lineterminator='\n')
writer.writerows(list)
f.close()
predict_BiGAN.py
predict_BiGAN.py
import numpy as np
import os
import tensorflow as tf
import utility as Utility
import argparse
import matplotlib.pyplot as plt
from model_BiGAN import BiGAN as Model
from make_datasets_predict import Make_datasets_predict as Make_datasets
def parser():
parser = argparse.ArgumentParser(description='train LSGAN')
parser.add_argument('--batch_size', '-b', type=int, default=300, help='Number of images in each mini-batch')
parser.add_argument('--log_file_name', '-lf', type=str, default='anpanman', help='log file name')
parser.add_argument('--epoch', '-e', type=int, default=1, help='epoch')
#parser.add_argument('--file_train_data', '-ftd', type=str, default='./mnist.npz', help='train data')
#parser.add_argument('--test_true_data', '-ttd', type=str, default='./mnist.npz', help='test of true_data')
#parser.add_argument('--test_false_data', '-tfd', type=str, default='./mnist.npz', help='test of false_data')
parser.add_argument('--test_data', '-td', type=str, default='../Test_Data/200112/', help='test of false_data')
parser.add_argument('--valid_span', '-vs', type=int, default=1, help='validation span')
parser.add_argument('--score_th', '-st', type=float, default=np.load('./score_threshold.npy'), help='validation span')
return parser.parse_args()
args = parser()
# global variants
BATCH_SIZE = args.batch_size
LOGFILE_NAME = args.log_file_name
EPOCH = args.epoch
# FILE_NAME = args.file_train_data
# TRUE_DATA = args.test_true_data
# FALSE_DATA = args.test_false_data
TEST_DATA = args.test_data
IMG_WIDTH = 100
IMG_HEIGHT = 100
IMG_CHANNEL = 1
BASE_CHANNEL = 32
NOISE_UNIT_NUM = 200
NOISE_MEAN = 0.0
NOISE_STDDEV = 1.0
TEST_DATA_SAMPLE = 5 * 5
L2_NORM = 0.001
KEEP_PROB_RATE = 0.5
SEED = 1234
SCORE_ALPHA = 0.9 # using for cost function
VALID_SPAN = args.valid_span
np.random.seed(seed=SEED)
BOARD_DIR_NAME = './tensorboard/' + LOGFILE_NAME
OUT_IMG_DIR = './out_images_BiGAN' #output image file
out_model_dir = './out_models_BiGAN/' #output model_ckpt file
# Load_model_dir = '../model_ckpt/' #Load model_ckpt file
OUT_HIST_DIR = './out_score_hist_BiGAN' #output histogram file
CYCLE_LAMBDA = 1.0
SCORE_TH = args.score_th
make_datasets = Make_datasets(TEST_DATA, IMG_WIDTH, IMG_HEIGHT, SEED)
model = Model(NOISE_UNIT_NUM, IMG_CHANNEL, SEED, BASE_CHANNEL, KEEP_PROB_RATE)
z_ = tf.placeholder(tf.float32, [None, NOISE_UNIT_NUM], name='z_') #noise to generator
x_ = tf.placeholder(tf.float32, [None, IMG_HEIGHT, IMG_WIDTH, IMG_CHANNEL], name='x_') #image to classifier
d_dis_f_ = tf.placeholder(tf.float32, [None, 1], name='d_dis_g_') #target of discriminator related to generator
d_dis_r_ = tf.placeholder(tf.float32, [None, 1], name='d_dis_r_') #target of discriminator related to real image
is_training_ = tf.placeholder(tf.bool, name = 'is_training')
with tf.variable_scope('encoder_model'):
z_enc = model.encoder(x_, reuse=False, is_training=is_training_)
with tf.variable_scope('decoder_model'):
x_dec = model.decoder(z_, reuse=False, is_training=is_training_)
x_z_x = model.decoder(z_enc, reuse=True, is_training=is_training_) # for cycle consistency
with tf.variable_scope('discriminator_model'):
#stream around discriminator
drop3_r, logits_r = model.discriminator(x_, z_enc, reuse=False, is_training=is_training_) #real pair
drop3_f, logits_f = model.discriminator(x_dec, z_, reuse=True, is_training=is_training_) #real pair
drop3_re, logits_re = model.discriminator(x_z_x, z_enc, reuse=True, is_training=is_training_) #fake pair
with tf.name_scope("loss"):
loss_dis_f = tf.reduce_mean(tf.square(logits_f - d_dis_f_), name='Loss_dis_gen') #loss related to generator
loss_dis_r = tf.reduce_mean(tf.square(logits_r - d_dis_r_), name='Loss_dis_rea') #loss related to real image
#total loss
loss_dis_total = loss_dis_f + loss_dis_r
loss_dec_total = loss_dis_f
loss_enc_total = loss_dis_r
with tf.name_scope("score"):
l_g = tf.reduce_mean(tf.abs(x_ - x_z_x), axis=(1,2,3))
l_FM = tf.reduce_mean(tf.abs(drop3_r - drop3_re), axis=1)
score_A = SCORE_ALPHA * l_g + (1.0 - SCORE_ALPHA) * l_FM
with tf.name_scope("optional_loss"):
loss_dec_opt = loss_dec_total + CYCLE_LAMBDA * l_g
loss_enc_opt = loss_enc_total + CYCLE_LAMBDA * l_g
tf.summary.scalar('loss_dis_total', loss_dis_total)
tf.summary.scalar('loss_dec_total', loss_dec_total)
tf.summary.scalar('loss_enc_total', loss_enc_total)
merged = tf.summary.merge_all()
# t_vars = tf.trainable_variables()
dec_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="decoder")
enc_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="encoder")
dis_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="discriminator")
with tf.name_scope("train"):
train_dis = tf.train.AdamOptimizer(learning_rate=0.00005, beta1=0.5).minimize(loss_dis_total, var_list=dis_vars
, name='Adam_dis')
train_dec = tf.train.AdamOptimizer(learning_rate=0.005, beta1=0.5).minimize(loss_dec_total, var_list=dec_vars
, name='Adam_dec')
train_enc = tf.train.AdamOptimizer(learning_rate=0.005, beta1=0.5).minimize(loss_enc_total, var_list=enc_vars
, name='Adam_enc')
train_dec_opt = tf.train.AdamOptimizer(learning_rate=0.005, beta1=0.5).minimize(loss_dec_opt, var_list=dec_vars
, name='Adam_dec')
train_enc_opt = tf.train.AdamOptimizer(learning_rate=0.005, beta1=0.5).minimize(loss_enc_opt, var_list=enc_vars
, name='Adam_enc')
sess = tf.Session()
ckpt = tf.train.get_checkpoint_state(out_model_dir)
saver = tf.train.Saver()
if ckpt: # checkpointがある場合
last_model = ckpt.model_checkpoint_path # 最後に保存したmodelへのパス
saver.restore(sess, last_model) # 変数データの読み込み
print("load " + last_model)
else: # 保存データがない場合
#init = tf.initialize_all_variables()
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(BOARD_DIR_NAME, sess.graph)
log_list = []
log_list.append(['epoch', 'AUC'])
# training loop
for epoch in range(1):
if epoch % VALID_SPAN == 0:
score_A_np = np.zeros((0, 2), dtype=np.float32)
val_data_num = len(make_datasets.valid_data)
img_batch_test = make_datasets.get_valid_data_for_1_batch(0, val_data_num)
score_A_ = sess.run(score_A, feed_dict={x_:img_batch_test, is_training_:False})
score_A_re = np.reshape(score_A_, (-1, 1))
tars_batch_re = np.where(score_A_re < SCORE_TH, 1, 0) #np.reshape(tars_batch, (-1, 1))
score_A_np_tmp = np.concatenate((score_A_re, tars_batch_re), axis=1)
x_z_x_test = sess.run(x_z_x, feed_dict={x_:img_batch_test, is_training_:False})
#print(score_A_np_tmp)
array_1_np, array_0_np = Utility.score_divide(score_A_np_tmp)
Utility.make_score_hist_test(array_1_np, array_0_np, SCORE_TH, LOGFILE_NAME, OUT_HIST_DIR)
Utility.make_output_img_test(img_batch_test, x_z_x_test, score_A_np_tmp, LOGFILE_NAME, OUT_IMG_DIR)
make_datasets_predict.py
make_datasets_predict.py
import numpy as np
import os
import glob
import re
import random
# import cv2
from PIL import Image
from keras.preprocessing import image
class Make_datasets_predict():
def __init__(self, test_data, img_width, img_height, seed):
self.filename = test_data
self.img_width = img_width
self.img_height = img_height
self.seed = seed
x_test = self.read_DATASET(self.filename)
self.valid_data = x_test
random.seed(self.seed)
np.random.seed(self.seed)
def read_DATASET(self, test_path):
test_list = os.listdir(test_path)
x_test = np.empty((0, self.img_width*self.img_height))
for img in test_list:
path_name = test_path+img
x_img = Image.open(path_name)
# サイズを揃える
x_img = x_img.resize((self.img_width, self.img_height))
# 3chを1chに変換
x_img= x_img.convert('L')
# PIL.Image.Imageからnumpy配列へ
x_img = np.array(x_img)
# 正規化
x_img = x_img / 255.0
# axisの追加
x_img = x_img.reshape((1,self.img_width, self.img_height))
# flatten
x_img = x_img.reshape(1, self.img_width*self.img_height)
x_test = np.concatenate([x_test, x_img], axis = 0)
print("x_test.shape, ", x_test.shape)
return x_test
def get_file_names(self, dir_name):
target_files = []
for root, dirs, files in os.walk(dir_name):
targets = [os.path.join(root, f) for f in files]
target_files.extend(targets)
return target_files
def divide_MNIST_by_digit(self, train_np, data1_num, data2_num):
data_1 = train_np[train_np[:,0] == data1_num]
data_2 = train_np[train_np[:,0] == data2_num]
return data_1, data_2
def read_data(self, d_y_np, width, height):
#tars = []
images = []
for num, d_y_1 in enumerate(d_y_np):
image = d_y_1.reshape(width, height, 1)
#tar = d_y_1[0]
images.append(image)
#tars.append(tar)
return np.asarray(images)#, np.asarray(tars)
def normalize_data(self, data):
# data0_2 = data / 127.5
# data_norm = data0_2 - 1.0
data_norm = (data * 2.0) - 1.0 #applied for tanh
return data_norm
def make_data_for_1_epoch(self):
self.filename_1_epoch = np.random.permutation(self.train_np)
return len(self.filename_1_epoch)
def get_data_for_1_batch(self, i, batchsize):
filename_batch = self.filename_1_epoch[i:i + batchsize]
images, _ = self.read_data(filename_batch, self.img_width, self.img_height)
images_n = self.normalize_data(images)
return images_n
def get_valid_data_for_1_batch(self, i, batchsize):
filename_batch = self.valid_data[i:i + batchsize]
images = self.read_data(filename_batch, self.img_width, self.img_height)
images_n = self.normalize_data(images)
return images_n#, tars
def make_random_z_with_norm(self, mean, stddev, data_num, unit_num):
norms = np.random.normal(mean, stddev, (data_num, unit_num))
# tars = np.zeros((data_num, 1), dtype=np.float32)
return norms
def make_target_1_0(self, value, data_num):
if value == 0.0:
target = np.zeros((data_num, 1), dtype=np.float32)
elif value == 1.0:
target = np.ones((data_num, 1), dtype=np.float32)
else:
print("target value error")
return target
GitHubのリンク
参考にした記事
https://qiita.com/masataka46/items/49dba2790fa59c29126b
https://qiita.com/underfitting/items/a0cbb035568dea33b2d7