train_group.py

# -*- coding: utf-8 -*-
"""
Created on Tue Jan 14 11:32:37 2020

@author: pku
"""
#%%
import os
import numpy as np
from PIL import Image
from tensorflow.keras import applications
from tensorflow.keras import optimizers
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.callbacks import ModelCheckpoint,EarlyStopping
from tensorflow.keras.layers import Flatten, Dense, Dropout

import matplotlib.pyplot as plt 
import h5py
#%%


nb_train_samples = 16000
nb_validation_samples = 500


#%%

data = np.load('../Rsp.npy')
test = np.load('../valRsp.npy')


n = nb_train_samples
train_x1 = np.load('../train_r132.npy')
val_x1 = np.load('../val_r132.npy')

#%%

n1 = 11*11*1024
STD1 = np.std(train_x1)


train_x = np.reshape(train_x1,(n,n1))/STD1
val_x = np.reshape(val_x1,(nb_validation_samples,n1))/STD1


train_y = np.reshape(data,(n,128*128))
val_y   = np.reshape(test,(nb_validation_samples,128*128))
val_y1 = np.reshape(val_y,(nb_validation_samples,128,128,1))


fnum = 128*128
ROI = np.mean(test**2,0)>0
roi = np.reshape(ROI,(fnum))
Cnum = 200

L1 = 1e-4
#%%
from tensorflow.keras.layers import Conv2D,BatchNormalization, MaxPooling2D, DepthwiseConv2D, Activation, GaussianNoise, LocallyConnected1D, Reshape, Conv2DTranspose
from tensorflow.keras import regularizers
from tensorflow.keras import activations
from tensorflow.keras.models import load_model
from tensorflow.keras import backend as K
import tensorflow as tf



def mean_squared_error_noise(y_true, y_pred):
    return K.mean(K.square(K.relu(K.abs(y_pred - y_true)-0.0)), axis=-1)

# to deal with Failed to get convolution algorithm. 
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.compat.v1.InteractiveSession(config=config)

#alexnet = load_model('alexnet.h5')
model = Sequential()
model.add(Dropout(0.1,input_shape=(n1,)))
model.add(Dense(Cnum, kernel_regularizer=regularizers.l1(L1/Cnum), activation='elu'))
model.add(Dense(fnum, use_bias=True))


w1 = model.layers[-1].get_weights()
w2 = np.zeros((Cnum, fnum))
w3 = np.zeros((fnum))
w3[roi] = 0.001
for i in range(Cnum):
    w2[i,:] = w3

w1[0] = w2
model.layers[-1].set_weights(w1)


#adadelta=optimizers.Adadelta(lr=0.001, rho=0.95, epsilon=1e-06)
Adam = optimizers.Adam(lr=0.00002, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
#model.layers[1].kernel_regularizer=regularizers.l1(1e-2/Cnum)

model.compile(loss=mean_squared_error_noise, optimizer=Adam, metrics=['mse'])
model.summary()
model.save_weights('initial.hdf5')

    

#%%

filepath="Cell.hdf5"
model.load_weights('initial.hdf5')

#earlyStopping=EarlyStopping(monitor='val_mse', patience=30, verbose=1, mode='auto')
saveBestModel = ModelCheckpoint(filepath, monitor='val_mse', verbose=1, save_best_only=True, mode='auto')
#callbacks_list = [earlyStopping,saveBestModel] 
callbacks_list = [saveBestModel] 


model.fit(train_x, train_y, epochs=20, batch_size=20, validation_data=(val_x, val_y), callbacks=callbacks_list)

#%% test trained model
model.load_weights(filepath)
pred = model.predict(val_x)
pred1 = np.reshape(pred,(nb_validation_samples,128,128,1))

R = np.zeros((128,128))
VE = np.zeros((128,128))
for i in range(128):
    for j in range(128):
        if np.sum(np.abs(val_y1[:,i,j,0]))>0:
            
            u2=np.zeros((2,nb_validation_samples))
            u2[0,:]=np.reshape(pred1[:,i,j,0],(nb_validation_samples))
            u2[1,:]=np.reshape(val_y1[:,i,j,0],(nb_validation_samples))
        
        
                
            c2=np.corrcoef(u2)
            R[i,j] = c2[0,1]
            VE[i,j] = 1-np.var(pred1[:,i,j,0]-val_y1[:,i,j,0])/np.var(val_y1[:,i,j,0])


np.save('R.npy',R)
np.save('VE.npy',VE)

print('L1:',L1)
print('R:',np.mean(R[ROI]))
    
K.clear_session()