利用tensorflow实现3DCNN

#!/usr/bin/env python2 # -*- coding: utf-8 -*- """ Created on Thu Feb 23 10:51:28 2017 @author: cdn """ import numpy as np np.random.seed(1234) import timeit import os import matplotlib.pyplot as plt import scipy.io as sio from sklearn.cross_validation import StratifiedKFold import tensorflow as tf from tensorflow.contrib.layers import fully_connected, convolution2d, flatten, dropout from tensorflow.python.layers.pooling import max_pooling3d from tensorflow.python.ops.nn import relu,softmax from tensorflow.python.framework.ops import reset_default_graph import six.moves.cPickle as pickle from confusionmatrix import ConfusionMatrix def onehot(t, num_classes): out = np.zeros((t.shape[0], num_classes)) for row, col in enumerate(t): out[row, col] = 1 return out def load_cv(cv_name = 'index_10fold.pkl',fold_idx = 0): input_doc = open(cv_name,'rb') in_data = pickle.load(input_doc) input_doc.close() train_idx,test_idx = in_data train_index0 = train_idx[fold_idx] test_index = test_idx[fold_idx] val_index = test_idx[fold_idx-1] train_index = list(set(train_index0)-set(val_index)) train_index = np.array(train_index) return train_index,test_index,val_index def load_data(fold_index): data = sio.loadmat('ADNI/PET/Data_PET_d3.mat') AD_data = data['AD_data_PET_d3'] NC_data = data['NORMAL_data_PET_d3'] # AD_data = data['AffineAD_128'] # NC_data = data['AffineNC_128'] ad_num,sizeX,sizeY,sizeZ = AD_data.shape nc_num,_,_,_ = NC_data.shape size_input = [1,sizeX,sizeY,sizeZ] np.random.seed(1234) random_idx = np.random.permutation(ad_num+nc_num) adnc_data = np.concatenate((AD_data,NC_data),axis=0)[random_idx] labels = np.hstack((np.ones((ad_num,)),np.zeros((nc_num,))))[random_idx] trainid,testid,validid = load_cv(fold_idx = fold_index) x_train = adnc_data[trainid] y_train = labels[trainid] x_test = adnc_data[testid] y_test = labels[testid] x_valid = adnc_data[validid] y_valid = labels[validid] return x_train,y_train,x_test,y_test,x_valid,y_valid,size_input n_fold = 10 train_accuracy = np.zeros((n_fold,)) test_accuracy = np.zeros((n_fold,)) valid_accuracy = np.zeros((n_fold,)) t1_time = timeit.default_timer() #for fi in range(n_fold): num_classes = 2 num_filters_conv1 = 10 num_filters_conv2 = 25 num_filters_conv3 = 40 num_filters_conv4 = 40 dense_num = 100 size_conv = 3 # [height, width] pool_size = 2 batch_size = 5 nb_epoch = 50 fi = 0 X_train,y_train,X_test,y_test,X_val,y_val,size_input = load_data(fi) X_train = X_train.reshape(X_train.shape[0], 1, X_train.shape[1], X_train.shape[2],X_train.shape[3]) X_val = X_val.reshape(X_val.shape[0], 1, X_val.shape[1],X_val.shape[2],X_val.shape[3]) X_test = X_test.reshape(X_test.shape[0], 1, X_test.shape[1], X_test.shape[2],X_test.shape[3]) print('X_train shape:', X_train.shape) print(X_train.shape[0], 'train samples') print(X_val.shape[0], 'validate samples') print(X_test.shape[0], 'test samples') train_accuracy = np.zeros((n_fold,)) test_accuracy = np.zeros((n_fold,)) valid_accuracy = np.zeros((n_fold,)) t1_time = timeit.default_timer() for fi in range(n_fold): print('Now running on fold %d'%(fi+1)) num_classes = 2 x_train,y_train,x_test,y_test,x_valid,y_valid,size_input = load_data(fi) nchannels,rows,cols,deps = size_input x_train = x_train.astype('float32') x_train = x_train.reshape((-1,nchannels,rows,cols,deps)) targets_train = y_train.astype('int32') x_valid = x_valid.astype('float32') x_valid = x_valid.reshape((-1,nchannels,rows,cols,deps)) targets_valid = y_valid.astype('int32') x_test = x_test.astype('float32') x_test = x_test.reshape((-1,nchannels,rows,cols,deps)) targets_test = y_test.astype('int32') # define a simple feed forward neural network # hyperameters of the model num_classes = 2 channels = x_train.shape[1] height = x_train.shape[2] width = x_train.shape[3] depth = x_train.shape[4] num_filters_conv1 = 10 num_filters_conv2 = 25 num_filters_conv3 = 40 num_filters_conv4 = 40 kernel_size_conv1 = [3, 3, 3] # [height, width] pool_size = [2,2,2] stride_conv1 = [1,1,1] # [stride_height, stride_width] num_l1 = 100 # resetting the graph ... reset_default_graph() # Setting up placeholder, this is where your data enters the graph! x_pl = tf.placeholder(tf.float32, [None, channels, height, width, depth]) l_reshape = tf.transpose(x_pl, [0, 2, 3, 4, 1]) # TensorFlow uses NHWC instead of NCHW is_training = tf.placeholder(tf.bool)#used for dropout # Building the layers of the neural network # we define the variable scope, so we more easily can recognise our variables later l_conv1 = convolution2d(l_reshape, num_filters_conv1, kernel_size_conv1, stride_conv1,activation_fn=relu, scope="l_conv1") l_maxpool1 = max_pooling3d(l_conv1,pool_size,pool_size) l_conv2 = convolution2d(l_maxpool1, num_filters_conv2, kernel_size_conv1, stride_conv1,activation_fn=relu,scope="l_conv2") l_maxpool2 = max_pooling3d(l_conv2,pool_size,pool_size) l_conv3 = convolution2d(l_maxpool2, num_filters_conv3, kernel_size_conv1, stride_conv1,activation_fn=relu,scope="l_conv3") l_maxpool3 = max_pooling3d(l_conv3,pool_size,pool_size) l_conv4 = convolution2d(l_maxpool3, num_filters_conv4, kernel_size_conv1, stride_conv1,activation_fn=relu,scope="l_conv4") l_flatten = flatten(l_conv4, scope="flatten") # use l_conv1 instead of l_reshape l1 = fully_connected(l_flatten, num_l1, activation_fn=relu, scope="l1") l1 = dropout(l1, is_training=is_training, scope="dropout") y = fully_connected(l1, num_classes, activation_fn=softmax, scope="y") # y_ is a placeholder variable taking on the value of the target batch. y_ = tf.placeholder(tf.float32, [None, num_classes]) # computing cross entropy per sample cross_entropy = -tf.reduce_sum(y_ * tf.log(y+1e-8), reduction_indices=[1]) # averaging over samples cross_entropy = tf.reduce_mean(cross_entropy) # defining our optimizer optimizer = tf.train.AdamOptimizer(learning_rate=0.001) # applying the gradients train_op = optimizer.minimize(cross_entropy) #Test the forward pass # x = np.random.normal(0,1, (45, 1,65, 52, 51)).astype('float32') #dummy data # restricting memory usage, TensorFlow is greedy and will use all memory otherwise gpu_opts = tf.GPUOptions(per_process_gpu_memory_fraction=0.2) # initialize the Session sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_opts)) sess.run(tf.global_variables_initializer()) # res = sess.run(fetches=[y], feed_dict={x_pl: x}) # res = sess.run(fetches=[y], feed_dict={x_pl: x, is_training: False}) # for when using dropout # print "y", res[0].shape #Training Loop from confusionmatrix import ConfusionMatrix batch_size = 10 num_epochs = 50 num_samples_train = x_train.shape[0] num_batches_train = num_samples_train // batch_size num_samples_valid = x_valid.shape[0] num_batches_valid = num_samples_valid // batch_size train_acc, train_loss = [], [] valid_acc, valid_loss = [], [] test_acc, test_loss = [], [] cur_loss = 0 loss = [] try: for epoch in range(num_epochs): #Forward->Backprob->Update params cur_loss = 0 for i in range(num_batches_train): idx = range(i*batch_size, (i+1)*batch_size) x_batch = x_train[idx] target_batch = targets_train[idx] # feed_dict_train = {x_pl: x_batch, y_: onehot(target_batch, num_classes)} feed_dict_train = {x_pl: x_batch, y_: onehot(target_batch, num_classes), is_training: True}