python - Simple Autoencoder on Tensorflow using pokemon -

i have implemented simple autoencoder close reference to: https://hackernoon.com/how-to-autoencode-your-pok%c3%a9mon-6b0f5c7b7d97 . results different expected (as unable post images now, please check them out in github: https://github.com/notha99y/autoencoder)

does know why model seems not learning?

thanks in advance.

my implementation is:

1) flatten out images. (none,64,64,3) => (none, 12288)

2) mean , variance normalized

3) encoding: 12288 => 1024 => 64 => 4. decoding: 4 => 64 => 1024 => 1228

4) multiply variance , add mean

code:

import numpy np import matplotlib.pyplot plt import os scipy.misc import imread import time sklearn.utils import shuffle import tensorflow tf os.environ['tf_cpp_min_log_level']='2'  def getimages(batchsz = 100,random = false):     '''access directory holding pokemon , returns images in     array of shape (batchsz,width,breath,channels)     if random set true, function take randomly selected set     of images     '''     directory = 'data/pokemon/jpeg/'     names = os.listdir(directory)     if random == true:         np.random.shuffle(names)      filenames = [os.path.join(directory,name) name in names if '.jpg' in name]     # print('total number of images: {}'.format(len(names)))     if batchsz == 'all':         # print('taking entire batch size of {}'.format(len(filenames)))         labels = [x.replace('.jpg','') x in names]         images = np.array([plt.imread(x) x in filenames])     else:         # print('taking batch size of {}'.format(batchsz))         labels = [x.replace('.jpg','') x in names[:batchsz]]         images = np.array([plt.imread(x) x in filenames[:batchsz]])         # print(images.shape)     return images,labels  def montage(images,labels,title = none):     '''plots montage of first 100 images corresponding labels     '''     f, axs = plt.subplots(10, 10, sharex='all', sharey='all')     f.suptitle(title)     axs = axs.ravel()     in range(100):         axs[i].imshow(images[i].astype(np.uint8))         axs[i].set_title(labels[i], fontsize = '6')         axs[i].axis('off')      # plt.show()  def meanimage(images):     '''takes in set of images , plot mean image     '''     mean = np.mean(images,axis =0).astype(np.uint8)     plt.figure()     plt.imshow(mean)     # plt.show()  def stdimage(images):     '''takes in set of images , plot stdev image     '''     stdev = np.std(images,axis=0).astype(np.uint8)     plt.figure()     plt.imshow(stdev)  def preprocess(images, labels, allimages):     '''preprocesses images minus-ing mean , diving stdev     '''     mean = np.mean(allimages, axis = 0)     stdev = np.std(allimages, axis = 0)     newimages = ((images - mean)/stdev)     return newimages  def postprocess(allimages,labels,normimag,plot = false, epoch = none):     '''postprocesses images multiplying stdev , adding mean     '''     mean = np.mean(allimages,axis = 0)     stdev = np.std(allimages,axis = 0)     newimages = normimag*stdev + mean     # clipped = np.clip(newimages, 0, 255)     if plot == true:         f, axs = plt.subplots(10, 10, sharex='all', sharey='all')         f.suptitle('epoch: {}'.format(epoch))         axs = axs.ravel()         in range(100):             axs[i].imshow(newimages[i].astype(np.uint8))             axs[i].set_title(labels[i], fontsize = '6')             axs[i].axis('off')     return newimages  def autoencoder(allimages,alllabels):     '''     12288 => 1024 => 64 => 4 => 64 => 1024 => 12288     '''     #variables     x = tf.placeholder(tf.float32, shape=[none,12288], name='inputs')     #encoding     w1 = tf.get_variable(name = 'encoder1', shape=[12288,1024], dtype=tf.float32,initializer=tf.contrib.layers.xavier_initializer())     w2 = tf.get_variable(name = 'encoder2', shape=[1024,64], dtype=tf.float32,initializer=tf.contrib.layers.xavier_initializer())     w3 = tf.get_variable(name = 'encoder3', shape=[64,4], dtype=tf.float32,initializer=tf.contrib.layers.xavier_initializer())     # w1 = tf.variable(tf.truncated_normal(shape=[12288, 1024], stddev=0.1, name='encoder1'))     # w2 = tf.variable(tf.truncated_normal(shape=[1024, 64], stddev=0.1, name='encoder2'))     # w3 = tf.variable(tf.truncated_normal(shape=[64, 4], stddev=0.1, name='encoder3'))     b1 = tf.variable(tf.constant(0.0,shape = [1024],name = 'encoderbias1'))     b2 = tf.variable(tf.constant(0.0,shape = [64],name = 'encoderbias2'))     b3 = tf.variable(tf.constant(0.0,shape = [4],name = 'encoderbias3'))     #decoding     w4 = tf.transpose(w3,name='decoder1')     w5 = tf.transpose(w2, name='decoder2')     w6 = tf.transpose(w1, name='decoder3')     b4 = tf.variable(tf.constant(0.0,shape = [64],name = 'decoderbias1'))     b5 = tf.variable(tf.constant(0.0,shape = [1024],name = 'decoderbias2'))     b6 = tf.variable(tf.constant(0.0,shape = [12288],name = 'decoderbias3'))     #operations     initop = tf.global_variables_initializer()     encode1 = tf.nn.relu(tf.matmul(x,w1) +b1, name = 'reluencode1') #12288 => 1024     encode2 = tf.nn.relu(tf.matmul(encode1,w2) +b2, name = 'reluencode2') #1024 => 64     encode3 = tf.nn.relu(tf.matmul(encode2,w3) +b3, name = 'reluencode3') #64 => 4     # encode1 = tf.matmul(x,w1) +b1 #12288 => 1024     # encode2 = tf.matmul(encode1,w2) +b2 #1024 => 64     # encode3 = tf.matmul(encode2,w3) +b3 #64 => 4     decode1 = tf.nn.relu(tf.add(tf.matmul(encode3,w4) , b4), name = 'reludecode1') #4 => 64     decode2 = tf.nn.relu(tf.add(tf.matmul(decode1,w5) , b5), name = 'reludecode2') #64 => 1024     y = tf.nn.relu(tf.add(tf.matmul(decode2,w6) , b6), name = 'outputs') #1024 => 12288     #checking shape     print('shape of w1: ', w1.shape)     #training     cost = tf.reduce_mean(tf.squared_difference(x,y), name = 'costfunction')     train = tf.train.gradientdescentoptimizer(learning_rate = 10e-3).minimize(cost)     #session     tf.session() sess:         tic = time.time()         sess.run(initop)         # print(type(w1.eval()))         assert np.allclose(w1.eval(), tf.transpose(w6).eval()),'w1: {} w6t: {}'.format(w1.eval(), tf.transpose(w6).eval())         err = []         epoch = 3001          firstbatch = getimages(100,true)         in range(epoch):             images, labels = getimages(100,true)             inputimages = preprocess(images,labels,allimages).reshape([-1,12288])             error = cost.eval(feed_dict={x: inputimages})             err.append(error)             if i%100 == 0:                 toc = time.time()                 print('epoch: ', i, 'error: ',error, 'time elapsed: ', toc-tic )                 recon = y.eval(feed_dict = {x:firstbatch[0].reshape([-1,12288])}).reshape([-1,64,64,3])                 output = postprocess(allimages,firstbatch[1],recon,true,i)                 # reconplot = montage(recon,labels,'reconstructed')                 # postplot = montage(output,labels, 'postprocess')                 # plt.show()                 # print('yhat:', yhat.eval(feed_dict = {x:batch[0], y:batch[1], keepprob : 1}))                 # print('predicted y: ', tf.argmax(yhat,1).eval(feed_dict={x:batch[0], y:batch[1], keepprob : 1}), ' true y: ', tf.argmax(y,1).eval(feed_dict={x:batch[0], y:batch[1], keepprob : 1}))             sess.run(train,feed_dict={x: inputimages})         originalplot = montage(firstbatch[0],firstbatch[1],'original')         # normplot = montage(preprocess(images,labels,allimages),labels, 'normalized images')         plt.show()         # savepath = saver.save(sess,'savenettest/mnistcnn.ckpt')         toc = time.time()         # print('neural net saved in {}. total time elapsed: {}'.format(savepath,toc-tic))         plt.plot(err)         plt.xlabel('epoch')         plt.ylabel('mean squared error')         plt.title('autoencoder')         plt.show()   # images, labels = getimages(100,true) allimages, alllabels = getimages('all') # testplot1 = montage(images,labels) # test1 = preprocess(images,labels) # testplot2 = montage(test1,labels) # test2 = postprocess(images,labels,test1,true) # plt.show() testautoencoder = autoencoder(allimages,alllabels)