recon.py

import time
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, Reshape, Flatten
from keras.layers import Conv2D, Conv2DTranspose
from keras.optimizers import RMSprop, from keras import regularizers

def init_model():
    start_time = time.time()
    print 'Compiling Model ... '
    model = Sequential()
    model.add(Flatten(input_shape=(1,128,128) ) )
    model.add(Dense(2*128*128))    
    model.add(Activation('tanh'))    
    model.add(Dense(128*128))    
    model.add(Activation('tanh'))
    model.add(Dense(128*128))
    model.add(Activation('tanh'))    
    model.add(Reshape((1,128,128) ) )    
    model.add(Conv2D(filters=64, kernel_size=(5,5), strides=(1,1), activation = 'relu', data_format="channels_first", kernel_regularizer=regularizers.l1(0.01)) )
    model.add(Conv2D(filters=64, kernel_size=(5,5), strides=(1,1), activation = 'relu', data_format="channels_first", kernel_regularizer=regularizers.l1(0.01)) )    
    model.add(Conv2DTranspose(filters=1, kernel_size=(9,9), strides=(1,1), activation='relu', data_format="channels_first" ) )
    rms = RMSprop()
    model.compile(loss='mean_squared_error', optimizer=rms, metrics=['mean_squared_error'])
    print 'Model complied in {0} seconds'.format(time.time() - start_time)
    return model 

Thanks for sharing this code! Do you have any thoughts about the best way to normalize the input k-space to train this network?
Thanks again!