Classify mnist numbers by unsupervised learning with keras [Autoencoder]
Introduction
Unsupervised learning is generally less accurate than supervised learning, but at the cost of many benefits. Specifically, as a scene where unsupervised learning is useful
**-Data whose pattern is not well understood --Time-varying data --Unlabeled data **
And so on.
Unsupervised learning learns the structure behind the data from the data itself. This allows you to take advantage of more unlabeled data, which may pave the way for new applications.
Last time, we classified by unsupervised learning using PCA and t-SNE. https://qiita.com/nakanakana12/items/af08b9f605a48cad3e4e
But after all I want to use the trendy deep learning, so in this article ** Unsupervised learning with autoencoder ** to hold. A detailed explanation of the autoencoder itself is omitted. Please refer to the references.
Library import
python
import keras
import random
import matplotlib.pyplot as plt
from matplotlib import cm
import seaborn as sns
import pandas as pd
import numpy as np
import plotly.express as px
import os
from sklearn.decomposition import PCA
from sklearn.cluster import KMeans
from sklearn.metrics import confusion_matrix
from sklearn.manifold import TSNE
from keras import backend as K
from keras.models import Sequential, Model, clone_model
from keras.layers import Activation, Dense, Dropout, Conv2D,MaxPooling2D,UpSampling2D
from keras import callbacks
from keras.layers import BatchNormalization, Input, Lambda
from keras import regularizers
from keras.losses import mse, binary_crossentropy
sns.set("talk")
Data preparation
Download mnist data and perform preprocessing. In the pre-processing, normalization and channel position adjustment are performed.
python
fashion_mnist = keras.datasets.mnist
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()
#Normalization
train_images = (train_images - train_images.min()) / (train_images.max() - train_images.min())
test_images = (test_images - test_images.min()) / (test_images.max() - test_images.min())
print(train_images.shape,test_images.shape)
#Adjusting the channel position
image_height, image_width = 28,28
train_images = train_images.reshape(train_images.shape[0],28*28)
test_images = test_images.reshape(test_images.shape[0],28*28)
print(train_images.shape, test_images.shape)
Creating an autoencoder
Create a model of the autoencoder. I was impressed by the extremely small number of codes.
Here, we will create an autoencoder ** that compresses to ** 36 dimensions. By simply connecting two fully connected layers, the first layer compresses to 36 dimensions and the second layer returns to the original size. In other words, the first layer is the encoder and the second layer is the decoder.
The writing style here is the same as ordinary supervised learning.
python
model = Sequential()
#Encoder
model.add(Dense(36, activation="relu", input_shape=(28*28,)))
#decoder
model.add(Dense(28*28,activation="sigmoid"))
model.compile(optimizer="adam",loss="binary_crossentropy")
model.summary()
Learning autoencoder
Next, let's train the autoencoder. ** The point here is that the correct answer data uses image data instead of labels. ** ** Learning was completed at 156 epoch in my environment.
python
fit_callbacks = [
callbacks.EarlyStopping(monitor='val_loss',
patience=5,
mode='min')
]
#Train the model
#Train to correct data_Use images
model.fit(train_images, train_images,
epochs=200,
batch_size=2024,
shuffle=True,
validation_data=(test_images, test_images),
callbacks=fit_callbacks,
)
Let's check the learning result. You can see that the loss has converged to a certain value.
python
#Check for test data loss
score = model.evaluate(test_images, test_images, verbose=0)
print('test xentropy:', score)
#Visualize test data loss
score = model.evaluate(test_images, test_images, verbose=0)
print('test xentropy:', score)
Autoencoder model creation
Next, take out only the encoder part from the previous model and make a model.
#Dimensional compression model
encoder = clone_model(model)
encoder.compile(optimizer="adam", loss="binary_crossentropy")
encoder.set_weights(model.get_weights())
#Delete the last layer
encoder.pop()
Visualize 36-dimensional data using the extracted autoencoder. ** The image is 36-dimensional in the middle layer, but you can see that the original data has been restored in the output layer. ** ** It's kind of strange.
python
#Visualize by selecting 10 points from test data
p = np.random.randint(0, len(test_images), 10)
x_test_sampled = test_images[p]
#Apply the selected sample to AutoEncoder
x_test_sampled_pred = model.predict(x_test_sampled,verbose=0)
#Call only encoder
x_test_sampled_enc = encoder.predict(x_test_sampled,verbose=0)
#Visualize the processing result
fig, ax = plt.subplots(3, 10,figsize=[20,10])
for i, label in enumerate(test_labels[p]):
#The original image
img = x_test_sampled[i].reshape(image_height, image_width)
ax[0][i].imshow(img, cmap=cm.gray_r)
ax[0][i].axis('off')
#Image compressed by AutoEncoder
enc_img = x_test_sampled_enc[i].reshape(6, 6)
ax[1][i].imshow(enc_img, cmap=cm.gray_r)
ax[1][i].axis('off')
#Image restored by AutoEncoder
pred_img = x_test_sampled_pred[i].reshape(image_height, image_width)
ax[2][i].imshow(pred_img, cmap=cm.gray_r)
ax[2][i].axis('off')
Classification of images by k-means
Finally, classification by k-means is performed using 36-dimensional data. It is classified into 10 clusters, and the highest number for each cluster is used as the prediction label.
python
#Creating dimensionally reduced data
x_test_enc = encoder.predict(train_images)
print(x_test_enc.shape)
#k-Classification by menas
KM = KMeans(n_clusters = 10)
result = KM.fit(x_test_enc)
#Evaluation by confusion matrix
df_eval = pd.DataFrame(confusion_matrix(train_labels,result.labels_))
df_eval.columns = df_eval.idxmax()
df_eval = df_eval.sort_index(axis=1)
df_eval
When I checked the confusion matrix, it seems that I couldn't classify it into 10 well.
Now let's visualize the image of each cluster.
python
#Display 5 images from each cluster
fig, ax = plt.subplots(5,10,figsize=[15,8])
for col_i in range(10):
idx_list = random.sample(set(np.where(result.labels_ == col_i)[0]), 5)
ax[0][col_i].set_title("cluster:" + str(col_i), fontsize=12)
for row_i, idx_i in enumerate(idx_list):
ax[row_i][col_i].imshow((train_images[idx_i].reshape(image_height, image_width)), cmap=cm.gray_r)
ax[row_i][col_i].axis('off')
** If you look at each cluster, you can see that they have similar features in the image even if the numbers are different. ** ** For example, cluster 0 is thick and cluster 4 is thin. From this, it can be seen that ** the data is well characterized by the image itself, not just the data represented by the label **.
It's interesting to be able to classify by information outside the label. It means that you can add new labels.
At the end
This time, I used an autoencoder to classify mnist numbers without supervised learning. ** I couldn't classify according to the data on the label, but when I visualized the cluster, the information outside the label became apparent **.
The great thing about unsupervised learning is that you can get information outside the label.
If you find it helpful, it would be encouraging if you could use LGTM.
References
What is an autoencoder? Explain the mechanism of pre-learning and how to use it now! !! https://it-trend.jp/development_tools/article/32-0024
AutoEncoder with Keras https://qiita.com/fukuit/items/2f8bdbd36979fff96b07
Python: Write AutoEncoder in Keras https://blog.amedama.jp/entry/keras-auto-encoder
Recommended Posts