conv_auto.py

# This code is for testing the performace of autoencoders with convolutional layers in denoising gene expression profiles.

from keras.layers import Input, Dense, Conv2D, MaxPooling2D, UpSampling2D, ZeroPadding2D
from keras.models import Model
from keras import backend as K
import numpy as np
from keras.callbacks import TensorBoard

input_profile = Input(shape=(24, 24, 1))
#x = ZeroPadding2D(padding=((0,1), (0,1)))(input_profile)

x = Conv2D(16, (3, 3), activation='relu', padding='same')(input_profile)
print('Conv2d', x.shape)
x = MaxPooling2D((2, 2), padding='same')(x)
print('MaxPooling2D1', x.shape)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
print('Conv2D2', x.shape)
x = MaxPooling2D((2, 2), padding='same')(x)
print('MaxPooling2D2', x.shape)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
print('Conv2D3', x.shape)
encoded = MaxPooling2D((2, 2), padding='same')(x)
print('MaxPooling2D3', x.shape)
# at this point the representation is (4, 42754503, 8) i.e. 128-dimensional

x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
print(x.shape)
x = UpSampling2D((2, 2))(x)
print(x.shape)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
print(x.shape)
x = UpSampling2D((2, 2))(x)
print(x.shape)
x = ZeroPadding2D(padding=(1,1))(x)
x = Conv2D(16, (3, 3), activation='relu')(x)
print(x.shape)
x = UpSampling2D((2, 2))(x)
print(x.shape)
decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)
print(decoded.shape)
autoencoder = Model(input_profile, decoded)
autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')

X_GTEx = np.load('GTEx_X_float64.npy')
x_train = X_GTEx 
x_train = np.reshape(np.pad(x_train.flatten(), (0,505), mode='constant', constant_values=0), (4783, 24, 24, 1))
np.random.shuffle(x_train)
MSE = []
for n in range(11):
    
    x_noisy = x_train + noise_factor *\
        np.random.normal(loc=0.0, scale=1.0, size=x_train.shape)
    np.random.shuffle(x_noisy)

    autoencoder.fit(x_train, x_train,\
                    epochs=10, batch_size=127,\
                    shuffle=True, validation_data=(x_train, x_train),\
                    callbacks=[TensorBoard(log_dir='/tmp/autoencoder')])

    denoised_profile = autoencoder.predict(x_train)
    mse = ((x_train - denoised_profile)**2).mean(axis=None)
    print(mse)
    MSE.append(MSE)

dir_ = ''
np.save(dir_+'')
print('loss data saved in', dir_)