我有一个带有4个头的模型。但是，为了尽早停止，我只想监视一个汇总了两个Head的指标。
不幸的是，我收到警告说：
以指标val_control_only_metric为条件的提前停止，该指标不可用。可用的指标有：
失利，
转向损耗
节流阀损耗
smoosh_loss，
is_sim_loss，
eering_control_only_metric，
油门_控制_仅_度量，
smoosh_control_only_metric，
is_sim_control_only_metric，
val_loss，
val_steering_loss，
val_throttle_loss，
val_smoosh_loss，
val_is_sim_loss，
val_steering_control_only_metric，
val_throttle_control_only_metric，
val_smoosh_control_only_metric，
val_is_sim_control_only_metric
在验证集上运行时，我想监视control_only_metric的输出。这有可能吗？
def default_n_linear():
input_shape = (120, 160, 3)
roi_crop=(0, 0)
drop = 0.1
input_shape = adjust_input_shape(input_shape, roi_crop)
img_in = Input(shape=input_shape, name='img_in') # 0
x = img_in
x = Convolution2D(24, (5,5), strides=(2,2), activation='relu', name="conv2d_1")(x) # 1
x = Dropout(drop)(x) # 2
x = Convolution2D(32, (5,5), strides=(2,2), activation='relu', name="conv2d_2")(x) # 3
x = Dropout(drop)(x) # 4
x = Convolution2D(64, (5,5), strides=(2,2), activation='relu', name="conv2d_3")(x) # 5
x = Dropout(drop)(x) # 6
x = Convolution2D(64, (3,3), strides=(1,1), activation='relu', name="conv2d_4")(x) # 7
x = Dropout(drop)(x) # 8
x = Convolution2D(64, (3,3), strides=(1,1), activation='relu', name="conv2d_5")(x) # 9
x = Dropout(drop)(x)
x = Flatten(name='flattened')(x) # 11
x = Dense(100, activation='relu')(x) # 12
x = Dropout(drop)(x) # 13
x = Dense(50, activation='relu')(x) # 14
x = Dropout(drop)(x) # 15
steering = Dense(1, activation="linear", name="steering")(x)
throttle = Dense(1, activation="linear", name="throttle")(x)
smoosh = Dense(1, activation="sigmoid", name="smoosh")(x)
x_stop_grad = Lambda(lambda x: K.stop_gradient(x))(x)
is_sim = Dense(1, activation="sigmoid", name="is_sim")(x_stop_grad) # real=0, sim=1
outputs = [steering, throttle, smoosh, is_sim]
model = Model(inputs=[img_in], outputs=outputs)
return model
def loss(y_true, y_pred):
gt_angle = y_true[0]
gt_throttle = y_true[1]
gt_is_sim = y_true[2]
point_5 = y_true[3]
pred_angle = y_pred[0]
pred_throttle = y_pred[1]
pred_is_sim = y_pred[2]
pred_smoosh = y_pred[3]
loss = K.mean(K.square(pred_angle-gt_angle))
loss += K.mean(K.square(pred_throttle-gt_throttle))
loss += K.mean(K.square(pred_is_sim-gt_is_sim))
loss += K.mean(K.square(pred_smoosh - point_5))
return loss
def control_only_metric(y_true, y_pred):
gt_angle = y_true[0]
gt_throttle = y_true[1]
pred_angle = y_pred[0]
pred_throttle = y_pred[1]
return K.mean(K.square(pred_angle-gt_angle)) +\
K.mean(K.square(pred_throttle-gt_throttle))
model = default_n_linear()
model.compile(optimizer=self.optimizer, loss=loss, metrics=[control_only_metric])
early_stop = keras.callbacks.EarlyStopping(monitor="val_control_only_metric",
min_delta=0.05,
patience=5
mode='min')
kl.model.fit_generator(train_gen,
steps_per_epoch=steps_per_epoch,
epochs=epochs,
validation_data=val_gen,
callbacks=[early_stop],
validation_steps=val_steps,
workers=workers_count,
use_multiprocessing=use_multiprocessing)