Q&A
失礼します。
CNN-Autoencoderを用いた画像分類器を作ろうとしていて、folder1内の28*28画像(460枚)を読み込んでCNN-Autoencoderで学習させて、更にfolder1内の画像を用いてテストした後にテスト画像の復元(decoded)を出力しようとしています。学習を進めていくと以下の様にlossやAccuracyの値が出ません。更には復元画像もほとんど真っ白な画像しか出ず、明らかに学習できていません。
step, loss, accuracy = 1: nan 0.935 step, loss, accuracy = 2: nan 0.935 step, loss, accuracy = 3: nan 0.891 step, loss, accuracy = 4: nan 0.891 step, loss, accuracy = 5: nan 0.913 step, loss, accuracy = 6: nan 0.978 step, loss, accuracy = 7: nan 0.913 step, loss, accuracy = 8: nan 0.935 step, loss, accuracy = 9: nan 0.848 step, loss, accuracy = 10: nan 0.913 step, loss, accuracy = 11: nan 0.935 step, loss, accuracy = 12: nan 0.870 step, loss, accuracy = 13: nan 0.848 step, loss, accuracy = 14: nan 0.913 step, loss, accuracy = 15: nan 0.935 step, loss, accuracy = 16: nan 0.935 step, loss, accuracy = 17: nan 0.913 step, loss, accuracy = 18: nan 0.913 step, loss, accuracy = 19: nan 0.913 step, loss, accuracy = 20: nan 0.826 loss (test) = nan accuracy(test) = 0.9347826
以下のコードを使用しています。
コードはこの方 https://qiita.com/TomokIshii/items/26b7414bdb3cd3052786 のコードを使わせて頂きました。ネットワークの定義はより高レベルに書き換えました(tf.nn.conv2dでなくtf.layers.conv2dを使用)。
cnn_ae.py
import numpy as np import matplotlib as mpl import random mpl.use('Agg') import matplotlib.pyplot as plt import tensorflow as tf import glob from PIL import Image from my_nn_lib import Convolution2D, MaxPooling2D from my_nn_lib import FullConnected, ReadOutLayer image_size = 28 one_layer = 14 two_layer = 7 # Create the model def mk_nn_model(x): # Encoding phase x_image = tf.reshape(x, [-1, image_size, image_size, 1]) conv1 = tf.layers.conv2d(inputs=x_image, filters=16, kernel_size=(3, 3), padding='same', activation=tf.nn.relu) pool1 = tf.layers.max_pooling2d(conv1, pool_size=(2, 2), strides=(2, 2), padding='same') conv2 = tf.layers.conv2d(inputs=pool1, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu) pool2 = tf.layers.max_pooling2d(conv2, pool_size=(2, 2), strides=(2, 2), padding='same') conv3 = tf.layers.conv2d(pool2, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu) encoded = tf.layers.max_pooling2d(conv3, pool_size=(2, 2), strides=(2, 2), padding='same') # at this point the representation is (8, 4, 4) i.e. 128-dimensional # Decoding phase upsample1 = tf.image.resize_images(encoded, size=(two_layer, two_layer), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR) conv4 = tf.layers.conv2d(inputs=upsample1, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu) upsample2 = tf.image.resize_images(conv4, size=(one_layer, one_layer), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR) conv5 = tf.layers.conv2d(inputs=upsample2, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu) upsample3 = tf.image.resize_images(conv5, size=(image_size, image_size), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR) conv6 = tf.layers.conv2d(inputs=upsample3, filters=16, kernel_size=(3, 3), padding='same', activation=tf.nn.relu) logits = tf.layers.conv2d(inputs=conv6, filters=1, kernel_size=(3, 3), padding='same', activation=None) decoded = tf.nn.sigmoid(logits) decoded = tf.reshape(decoded, [-1, image_size*image_size]) cross_entropy = -1. * x * tf.log(decoded) - (1. - x) * tf.log(1. - decoded) loss = tf.reduce_mean(cross_entropy) prediction_match = tf.equal(tf.argmax(decoded, axis=1), tf.argmax(x, axis=1)) accuracy = tf.reduce_mean(tf.cast(prediction_match, tf.float32), name='accuracy') return loss, decoded, accuracy if __name__ == '__main__': all_files = glob.glob('folder1/*.jpg') filelist = [] for f in all_files: image = Image.open(f) np_image = np.array(image) list_image = np_image.tolist() filelist.append(list_image) print("success") data_number = len(filelist) fileset = np.array(filelist).reshape((data_number, image_size*image_size)) # Variables x = tf.placeholder(tf.float32, (None, image_size*image_size)) learning_rate = 0.001 loss, decoded, accuracy = mk_nn_model(x) train_step = tf.train.AdagradOptimizer(learning_rate).minimize(loss) epochs = 20 batch_size = data_number // 10 init = tf.initialize_all_variables() # Train with tf.Session() as sess: sess.run(init) print('Training...') for i in range(epochs): X = np.random.permutation(fileset) for ii in range(data_number // batch_size): batch = X[ii * batch_size:(ii + 1) * batch_size] train_loss, train_accuracy, _ = sess.run([loss, accuracy, train_step], feed_dict={x: batch}) print(' step, loss, accuracy = %6d: %6.3f %6.3f' % (i+1, train_loss, train_accuracy)) # generate decoded image with test data X = np.random.permutation(fileset) test_image = X[0:batch_size] decoded_imgs, test_loss, test_accuracy = sess.run([decoded, loss, accuracy], feed_dict={x: test_image}) print('loss (test) = ', test_loss) print('accuracy(test) = ', test_accuracy) x_test = test_image n = 10 # how many digits we will display plt.figure(figsize=(20, 4)) for i in range(n): # display original ax = plt.subplot(2, n, i + 1) plt.imshow(x_test[i].reshape(image_size, image_size)) plt.gray() ax.get_xaxis().set_visible(False) ax.get_yaxis().set_visible(False) # display reconstruction ax = plt.subplot(2, n, i + 1 + n) plt.imshow(decoded_imgs[i].reshape(image_size, image_size)) plt.gray() ax.get_xaxis().set_visible(False) ax.get_yaxis().set_visible(False) # plt.show() plt.savefig('images.png')
####実行環境
OS Windows10 64bit
プロセッサ Intel(R) Core(TM)i7-8700k CPU @ 3.70GHz
RAM 32.0 GB
Anaconda Prompt
####試したこと
・decoded直前でsigmoidをsoftmaxにした。その結果、nanばかり表示されてしまい、全く改善せず。
・元々、256256の画像でやってみてうまくいかないと考えて、2828に圧縮してやってみたが全く改善しなかった。よって、画像サイズの問題ではなく、ナットワーク設計そのものに問題があると考えられる。
・もちろんサンプルコードより大きなepoch数(100程度)で試したが結果は変わらず。
おそらくネットワーク設計(重みの設定、誤差関数の設定など)を改善するべきなのでしょうが、具体的にどこをどう確かめれば良いのか方向性が掴めません。lossがnanになった途端にaccuracyが1に近づいたので、外れ値(-∞か+∞)がsigmoidを通って、0か1を吐き出したのだろうということくらいしか分かりません。
まだ機械学習をやり始めて経験が浅くどうすればよいか分からない状況ですので、何卒助言頂ければ幸いです。
あなたの回答
tips
プレビュー
