def on_epoch_end(self, epoch, logs={}):
BatchSize = 64
INPUT_SHAPE = (224, 224, 3)
annotation = r'H:\Qiulin\ODIR\ODIR_dataset\ODIR-5K_training-Chinese_modified_v1.xlsx'
ODIR_valDir = r'H:\Qiulin\ODIR\ODIR_dataset\ODIR-5K_training_cvt\v3_val_aug_equ_c'
val_gen = data_generator(annotation,
ODIR_valDir,
BatchSize,
INPUT_SHAPE,
mode='valid')
x_val, y_val = next(val_gen.get_mini_batch())
predict = self.model.predict_on_batch(x_val)
steps = val_gen.num_of_images // val_gen.batch_size #有多少个steps
for i in range(steps):
if i == 0:
continue
datas, labels = next(val_gen.get_mini_batch())
pre = self.model.predict_on_batch(datas)
predict = np.vstack((predict, pre))
y_val = np.vstack((y_val, labels))
th = 0.5
gt = y_val.flatten()
pr = predict.flatten()
kappa = metrics.cohen_kappa_score(gt, pr > th)
f1 = metrics.f1_score(gt, pr > th, average='micro')
auc = metrics.roc_auc_score(gt, pr)
final_score = (kappa + f1 + auc) / 3.0
print('val_kappa: %.3f' % kappa)
print('val_f1 : %.3f' % f1)
print('val_auc : %.3f' % auc)
print('val_final_score:%.3f' % final_score)
logs['final_score'] = final_score
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["TF_CPP_MIN_LOG_LEVEL"] = '2' #只显示warning 和 error
if __name__ == '__main__':
BatchSize = 32
epochs = 20
INPUT_SHAPE = (224, 224, 3)
model_savePath = r'H:\Qiulin\ODIR\model\v6\vgg16_bv_v1.h5'
annotation = r'H:\Qiulin\ODIR\ODIR_dataset\ODIR-5K_training-Chinese_modified_v1.xlsx'
ODIR_trainDir = r'H:\Qiulin\ODIR\ODIR_dataset\ODIR-5K_training_cvt\v3_train_aug_bv'
'''generator'''
train_gen = data_generator(annotation,
ODIR_trainDir,
BatchSize,
INPUT_SHAPE,
mode='train')
'''training'''
model = get_VGG16(INPUT_SHAPE)
#定义callback需要做的事情
cb = [
Roc_Callback(), # include it before EarlyStopping!
# EarlyStopping(monitor='final_score', patience=10, verbose=2, mode='max')
ModelCheckpoint(model_savePath,
monitor='final_score',
verbose=1,
save_best_only=True,
mode='max'),
ReduceLROnPlateau(monitor='final_score',