#-*- coding:utf-8 -*-
from __future__ import print_function
import os, cv2
import scipy.io as sio
from scipy.misc import imsave
from sklearn.metrics import roc_auc_score, roc_curve, auc, recall_score, precision_score, f1_score
from sklearn.metrics import confusion_matrix, precision_recall_fscore_support
from PIL import Image
import numpy as np
from keras import backend as K
from keras.utils import np_utils
from keras.models import model_from_json, Model
import tensorflow as tf
(w_in, h_in, s_in, c_in) = (32, 32, 5, 5)
margin = 5
samples_n = 24
samples = range(0, samples_n)
# util function to convert a tensor into a valid image
def deprocess_image(x):
# normalize tensor: center on 0., ensure std is 0.1
x -= x.mean()
x /= (x.std() + K.epsilon())
x *= 0.1
# clip to [0, 1]
x += 0.5
x = np.clip(x, 0, 1)
# convert to RGB array
x *= 255
if K.image_data_format() == 'channels_first':
x = x.transpose((1, 2, 0))
x = np.clip(x, 0, 255).astype('uint8')
return x
def loadSplitData(path, classes):
train_data = []
train_labels = []
fp = open(os.path.join(path, "visual_train.txt"), 'r')
# fp = open(os.path.join(path, "train.txt"), 'r')
line = fp.readline() # 调用文件的 readline()
while len(line):
# print '*'*10,line
Level = int(line[0])
imgpath = line[2:-1]
mat = sio.loadmat(imgpath)
# print mat.keys()
train_data.append(mat['P'])
train_labels.append(Level)
line = fp.readline()
fp.close()
train_labels = np_utils.to_categorical(train_labels, classes)
train_data = np.asarray(train_data, dtype="float32")
X_train, y_train = train_data, train_labels
test_data = []
test_labels = []
fp = open(os.path.join(path, "visual_test.txt"), 'r')
# fp = open(os.path.join(path, "test.txt"), 'r')
line = fp.readline() # 调用文件的 readline()
while len(line):
# print '*'*10,line
Level = int(line[0])
imgpath = line[2:-1]
mat = sio.loadmat(imgpath)
# print mat.keys()
test_data.append(mat['P'])
test_labels.append(Level)
line = fp.readline()
fp.close()
test_labels = np_utils.to_categorical(test_labels, classes)
test_data = np.asarray(test_data, dtype="float32")
X_test, y_test = test_data, test_labels
return (X_train, y_train), (X_test, y_test)
def combData(path, C, dict, isTrans):
(trainData, trainLabels), (testData, testLabels) = loadSplitData(path, 3)
classes = 3
Train_list = []
Test_list = []
for modual in C:
if len(modual) == 1 or len(modual) == 3:
input_shape = (trainData.shape[1], trainData.shape[2])
Train_list.append(trainData[:,:,:,dict[modual]].reshape(trainData.shape[0], input_shape[0], input_shape[1]))
Test_list.append(testData[:,:,:,dict[modual]].reshape(testData.shape[0], input_shape[0], input_shape[1]))
else:
input_shape = (trainData.shape[1], trainData.shape[2], 5)
Train_list.append(trainData[:,:,:,dict[modual]].reshape(trainData.shape[0], input_shape[0], input_shape[1], 5))
Test_list.append(testData[:,:,:,dict[modual]].reshape(testData.shape[0], input_shape[0], input_shape[1], 5))
if len(C[0]) == 1 or len(C[0]) == 3:
X_train = tf.cast(tf.stack(Train_list, axis=3), tf.float32)
X_test = tf.cast(tf.stack(Test_list, axis=3), tf.float32)
else:
X_train = tf.cast(tf.stack(Train_list, axis=4), tf.float32)
X_test = tf.cast(tf.stack(Test_list, axis=4), tf.float32)
if isTrans:
# # 转置
X_train = tf.transpose(X_train, perm=[0,1,2,4,3])
X_test = tf.transpose(X_test, perm=[0,1,2,4,3])
with tf.Session() as sess:
X_train, X_test = X_train.eval(), X_test.eval()
return (X_train, trainLabels), (X_test, testLabels), classes
def accuracy(y_true, y_pred):
# 计算混淆矩阵
y = np.zeros(len(y_true))
y_ = np.zeros(len(y_true))
for i in range(len(y_true)):
y[i] = np.argmax(y_true[i,:])
y_[i] = np.argmax(y_pred[i,:])
cnf_mat = confusion_matrix(y, y_)
# Acc = 1.0*(cnf_mat[1][1]+cnf_mat[0][0])/len(y_true)
# Sens = 1.0*cnf_mat[1][1]/(cnf_mat[1][1]+cnf_mat[1][0])
# Spec = 1.0*cnf_mat[0][0]/(cnf_mat[0][0]+cnf_mat[0][1])
# # 绘制ROC曲线
# fpr, tpr, thresholds = roc_curve(y_true[:,0], y_pred[:,0])
# Auc = auc(fpr, tpr)
# 计算多分类评价值
Sens = recall_score(y, y_, average='macro')
Prec = precision_score(y, y_, average='macro')
F1 = f1_score(y, y_, average='weighted')
Support = precision_recall_fscore_support(y, y_, beta=0.5, average=None)
return Sens, Prec, F1, cnf_mat
def save_map_img(layer_name, ex_model, (map_w, map_h, map_s), map_n):
width = map_w * map_n + (map_n-1) * margin
height = map_h * len(samples) + margin * (len(samples) - 1)
if map_s == 1 or map_s == 3:
stitched_filters = np.zeros((width, height, map_s))
else:
raise RuntimeError('Channel Error')
for j, n in enumerate(samples):
tr_features = ex_model.predict(trainData[n,:,:,:,:].reshape(1,w_in, h_in, s_in, c_in))
feature_map = deprocess_image(tr_features)
# fill the picture with our saved filters
for i in range(map_n):
stitched_filters[(map_w+margin)*i : (map_w+margin)*i+map_w,
(map_h+margin)*j : (map_h+margin)*j+map_h, :] = feature_map[0,:,:,:,i]
# save the result to disk
# imsave(layer_name+'feature.png', stitched_filters)
cv2.imwrite('../models/MFC-3DCNN_models/'+layer_name+'feature.png', stitched_filters)
# build the VGG16 network with ImageNet weights
model = model_from_json(open('../models/MFC-3DCNN_models/architecture.json').read())
model.load_weights('../models/MFC-3DCNN_models/weights.h5')
print('Model loaded.')
model.summary()
# # get the symbolic outputs of each "key" layer (we gave them unique names).
# layer_dict = dict([(layer.name, layer) for layer in model.layers[0:]])
# for layer in model.layers[0:]:
# print(layer)
#读取数据
path = '../data_loader/ABKEFGHIJ_EFGHIJ_FGHIJ-E_FGHIJ5-E'
dict = {'A': 0, 'B': 1, 'K': 2, 'E': 3, 'F': 4, 'G': 5, 'H': 6, 'I': 7, 'J': 8, \
'E5':range(9,14), 'F5':range(14,19), 'G5':range(19,24), 'H5':range(24,29), 'I5':range(29,34), 'J5':range(34,39), \
'F-E':39, 'G-E':40, 'H-E':41, 'I-E':42, 'J-E':43, \
'F-E5':range(44,49), 'G-E5':range(49,54), 'H-E5':range(54,59), 'I-E5':range(59,64), 'J-E5':range(64,69)}
C = ['E5', 'F5', 'G5', 'H5', 'I5']
(trainData, trainLabels), (testData, testLabels), classes = combData(path, C, dict, 1)
print(trainLabels[:samples_n, :])
#验证模型
pred_test = model.predict(trainData)
sens, prec, f1, cnf_mat = accuracy(testLabels, pred_test)
print(sens, prec, f1, '\n', cnf_mat)
# the name of the layer we want to visualize
#绘制第二个卷积层特征图
layer_name = 'conv1'
ex_model = Model(inputs=model.input, outputs=model.get_layer(layer_name).output)
#绘制图像
save_map_img(layer_name, ex_model, (30, 30, 3), 6)
#绘制第二个卷积层特征图
layer_name = 'conv2'
ex_model = Model(inputs=model.input, outputs=model.get_layer(layer_name).output)
save_map_img(layer_name, ex_model, (12, 12, 1), 8)
