Keras模型训练报错：dense层输入维度不兼容求助

尝试用Keras和TensorFlow拟合概率分布时出现维度不兼容错误，相关代码及报错信息如下：

模型构建代码 #

def build_env_model(learning_rate):
  InputLayer = Input(shape=(5,))
  Layer_1 = Dense(16,activation='tanh')(InputLayer)
  Layer_2 = Dense(16,activation='tanh')(Layer_1)

  position_mean = Dense(1, activation="linear")(Layer_2)
  position_sigma = Dense(1, activation=lambda x: tf.nn.elu(x) + 1)(Layer_2)

  velocity_mean = Dense(1, activation="linear")(Layer_2)
  velocity_sigma = Dense(1, activation=lambda x: tf.nn.elu(x) + 1)(Layer_2)

  angle_mean = Dense(1, activation="linear")(Layer_2) 
  angle_sigma = Dense(1, activation=lambda x: tf.nn.elu(x) + 1)(Layer_2)

  angular_velocity_mean = Dense(1, activation="linear")(Layer_2)
  angular_velocity_sigma = Dense(1, activation=lambda x: tf.nn.elu(x) + 1)(Layer_2)

  reward_mean = Dense(1, activation=lambda x: tf.nn.elu(x) + 1)(Layer_2)
  reward_sigma = Dense(1, activation=lambda x: tf.nn.elu(x) + 1)(Layer_2)

  y_real = Input(shape=(5,))
  lossF = cost(position_mean,position_sigma,velocity_mean,velocity_sigma,angle_mean,angle_sigma,angular_velocity_mean,angular_velocity_sigma,reward_mean,reward_sigma,y_real)
  model = Model(inputs=[InputLayer,y_real],outputs=[position_mean,position_sigma,velocity_mean,velocity_sigma,angle_mean,
                                                    angle_sigma,angular_velocity_mean,angular_velocity_sigma,reward_mean,reward_sigma])
  model.add_loss(lossF)
  adamOptimizer = adam_v2.Adam(learning_rate=learning_rate)
  model.compile(optimizer=adamOptimizer,metrics=['mse'])
  return model

模型训练代码 #

def update_env_model(self,state,action,reward,next_state,done):
    state = state[0]
    next_state = next_state[0]
    position,velocity,angle,angular_velocity = state
    ns_position,ns_velocity,ns_angle,ns_angular_velocity = next_state
    inp1 = np.array([position,velocity,angle,angular_velocity,action])
    inp2 = np.array([ns_position,ns_velocity,ns_angle,ns_angular_velocity,reward])
    self.env_model.fit([inp1,inp2],verbose=0)

报错信息 #

ValueError: in user code:

File "/usr/local/lib/python3.7/dist-packages/keras/engine/training.py", line 1051, in train_function  *
    return step_function(self, iterator)
File "/usr/local/lib/python3.7/dist-packages/keras/engine/training.py", line 1040, in step_function  **
    outputs = model.distribute_strategy.run(run_step, args=(data,))
File "/usr/local/lib/python3.7/dist-packages/keras/engine/training.py", line 1030, in run_step  **
    outputs = model.train_step(data)
File "/usr/local/lib/python3.7/dist-packages/keras/engine/training.py", line 889, in train_step
    y_pred = self(x, training=True)
File "/usr/local/lib/python3.7/dist-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
File "/usr/local/lib/python3.7/dist-packages/keras/engine/input_spec.py", line 228, in assert_input_compatibility
    raise ValueError(f'Input {input_index} of layer "{layer_name}" '

ValueError: Exception encountered when calling layer "model_1" (type Functional).

Input 0 of layer "dense_16" is incompatible with the layer: expected min_ndim=2, found ndim=1. Full shape received: (None,)

Call arguments received by layer "model_1" (type Functional):
  • inputs=('tf.Tensor(shape=(None,), dtype=float32)', 'tf.Tensor(shape=(None,), dtype=float32)')
  • training=True
  • mask=None

报错核心是模型要求输入为2维（批量维度+特征维度），但训练时传入的是1维数组。Keras的Dense层需要输入至少是2维（形如(batch_size, feature_size)），而当前inp1和inp2是1维的(5,)，传入后会被识别为(None,)，不满足层的输入要求。

修改训练代码，给输入增加批量维度，有两种实现方式：

方式1：用np.expand_dims追加维度 #

def update_env_model(self,state,action,reward,next_state,done):
    state = state[0]
    next_state = next_state[0]
    position,velocity,angle,angular_velocity = state
    ns_position,ns_velocity,ns_angle,ns_angular_velocity = next_state
    inp1 = np.array([position,velocity,angle,angular_velocity,action])
    inp2 = np.array([ns_position,ns_velocity,ns_angle,ns_angular_velocity,reward])
    # 增加批量维度，将形状从(5,)变为(1,5)
    inp1 = np.expand_dims(inp1, axis=0)
    inp2 = np.expand_dims(inp2, axis=0)
    self.env_model.fit([inp1,inp2],verbose=0)

方式2：直接创建2维数组 #

def update_env_model(self,state,action,reward,next_state,done):
    state = state[0]
    next_state = next_state[0]
    position,velocity,angle,angular_velocity = state
    ns_position,ns_velocity,ns_angle,ns_angular_velocity = next_state
    # 用双层列表直接生成(1,5)的2维数组
    inp1 = np.array([[position,velocity,angle,angular_velocity,action]])
    inp2 = np.array([[ns_position,ns_velocity,ns_angle,ns_angular_velocity,reward]])
    self.env_model.fit([inp1,inp2],verbose=0)

后续如果要处理批量数据，只需将多个样本堆叠成(batch_size,5)的形状传入即可，无需修改模型结构。

内容的提问来源于stack exchange，提问作者Tanmay Aeron