# Compute class weights to deal with unbalanced data
class_weight = class_weight.compute_class_weight('balanced', np.unique(trainY),
trainY)
BATCH_SIZE = 64
learning_rate = 0.005
nb_epochs = 400
dropout_rate = 0.7
optimizer = tf.keras.optimizers.Adamax(learning_rate)
loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
model = CNN(dropout_rate=dropout_rate)
model.compile(optimizer=optimizer,
loss=loss_object,
metrics=['sparse_categorical_accuracy'])
# Use early stopping to avoid overfitting
earlystopping = EarlyStopping(monitor='val_loss', min_delta=1e-4, patience=10)
# Train the model on the original dataset
def train():
model.fit(trainX,
trainY,
batch_size=BATCH_SIZE,
epochs=400,
validation_data=(valX, valY),
class_weight=class_weight,
callbacks=[earlystopping])
def create_model_optimizer(optimizer=tf.keras.optimizers.SGD(0.001)):
model = CNN()
loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
model.compile(optimizer=optimizer, loss=loss_object, metrics=['accuracy'])
return model
)
num_batches = (train_x.shape[0] // args.batch) + 1
csvlogger = CSVLogger(loggerfile)
earlystopping = EarlyStopping(monitor='val_acc',
patience=10,
verbose=1,
mode='max')
checkpoint = ModelCheckpoint(paramfile,
monitor='val_acc',
save_best_only=True,
save_weights_only=True,
verbose=0,
mode='max')
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
print('Start training...')
if args.model != 'DNN':
model.fit_generator(datagen.flow(train_x,
train_y,
batch_size=args.batch),
steps_per_epoch=5 * num_batches,
epochs=args.epoch,
verbose=1,
validation_data=(valid_x, valid_y),
workers=8,
callbacks=[checkpoint, earlystopping, csvlogger])
else:
model.fit(train_x,
train_y,
_, x_valid, y_valid = Fer2013().gen_valid()
_, x_test, y_test = Fer2013().gen_test()
# target编码
y_train = to_categorical(y_train).reshape(y_train.shape[0], -1)
y_valid = to_categorical(y_valid).reshape(y_valid.shape[0], -1)
# 为了统一几个数据集，必须增加一列为0的
y_train = np.hstack((y_train, np.zeros((y_train.shape[0], 1))))
y_valid = np.hstack((y_valid, np.zeros((y_valid.shape[0], 1))))
print(
"load fer2013 dataset successfully, it has {} train images and {} valid iamges"
.format(y_train.shape[0], y_valid.shape[0]))
model = CNN(input_shape=(48, 48, 1), n_classes=8)
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
callback = [
# EarlyStopping(monitor='val_loss', patience=50, verbose=True),
# ReduceLROnPlateau(monitor='lr', factor=0.1, patience=20, verbose=True),
ModelCheckpoint('./models/cnn_best_weights.h5',
monitor='val_acc',
verbose=True,
save_best_only=True,
save_weights_only=True)
]
train_generator = ImageDataGenerator(rotation_range=10,
width_shift_range=0.05,
height_shift_range=0.05,
horizontal_flip=True,
