CODEWITHSUNDEEP
pythonkeras
JayJay81
JayJay81

Reputation: 203

Validating the structure of a convolutional encoder

I am building a convolutional encoder to deal with some 128x128 images - like this.

I have been testing the structure by testing it with a image set of 500 images. The resulting decoded images are basically entirely black (not want I want!)

I was hoping to get some advice on here as I think I am making some obvious mistakes.

A small group of images can be downloaded here -> https://www.dropbox.com/sh/0oj1p6sqal32cvx/AAAYQJSK2SPfynD8wYMSo9bPa?dl=0

Current Code

################################# SETUP #######################################

import glob
import pandas as pd
import numpy as np
import sys
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
import random

np.set_printoptions(threshold=np.nan)

######################### DATA PREPARATION #####################################

# create a list of XML files within the raw data folder
image_list = glob.glob("Images/test_images/*.jpeg")
print(image_list)

l = []
for i in image_list:
    img = np.array(cv2.imread(i, 0))
    l.append(img)
T = np.array(l)

# split into training and testing sets
labels = image_list
data_train, data_test, labels_train, labels_test = train_test_split(T, labels, test_size=0.5, random_state=42)

# convert to 0-1 floats (reconversion by * 255)
data_train = data_train.astype('float32') / 255.
data_test = data_test.astype('float32') / 255.
print(data_train.shape)

# reshape from channels first to channels last
data_train = np.rollaxis(data_train, 0, 3)
data_test = np.rollaxis(data_test, 0, 3)

######################### ENCODER MODELING #####################################

from keras.layers import Input, Dense, Conv2D, MaxPooling2D, UpSampling2D
from keras.models import Model
from keras import backend as K

input_img = Input(shape=(128, 128, 1))  # adapt this if using `channels_first` image data format

x = Conv2D(64, (3, 3), activation='relu', padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(16, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
encoded = MaxPooling2D((2, 2), padding='same')(x)

x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(16, (3, 3), activation='relu', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(64, (3, 3), activation='relu', padding='same')(x)
x = UpSampling2D((2, 2))(x)
decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)

# create the model
autoencoder = Model(input_img, decoded)
autoencoder.compile(optimizer='adadelta', loss='mse')

# reference for reordering
data_train_dimensions =  data_train.shape
data_test_dimensions =  data_test.shape

# reshape the data sets
data_train = np.reshape(data_train, (data_train_dimensions[2], 128, 128, 1))  # adapt this if using `channels_first` image data format
data_test = np.reshape(data_test, (data_test_dimensions[2], 128, 128, 1))

from keras.callbacks import TensorBoard
autoencoder.fit(data_train, data_test,
                epochs=10,
                batch_size=128,
                shuffle=True,
                validation_data=(data_train, data_test),
                callbacks=[TensorBoard(log_dir='/tmp/autoencoder')])

# create decoded images from model
decoded_imgs = autoencoder.predict(data_test)

# reorder columns
decoded_imgs = np.rollaxis(decoded_imgs, 3, 1)

# reshape from channels first to channels last
data_train = np.rollaxis(data_train, 0, 3)
data_test = np.rollaxis(data_test, 0, 3)

# convert to 0-1 floats (reconversion by * 255)
data_train = data_train.astype('float32') * 255.
data_test = data_test.astype('float32') * 255.

Upvotes: 1

Views: 76

Answers (1)

Mikhail Berlinkov
Mikhail Berlinkov

Reputation: 1624

I think the main problem is that you fit on data_train, data_test instead of data_train, labels_train, that is, you should fit your model on samples and corresponding outputs but you train it only on inputs which are by a coincidence are of compatible shapes because of the 50/50 split.

If the intention of the model is to reproduce an image from compressed representation, then you could train fit(data_train, data_train, ..., validation_data=(data_test, data_test)).

Upvotes: 1

Related Questions