Python 3.x 多输入误差预测神经网络

我创建了一个带有多输入和单输出的NN。我的简单网络每个观察有3个输入节点，在隐藏层（中间）只有一个节点。之后，我将它们连接起来。最后是一个单一输出，具有与输入节点一样多的节点

num_vars_per_input = 3
num_single_inputs = 2
num_middle_units = 1
output_shape= num_single_inputs

single_input_shape = (num_vars_per_input,)

input_layers = [keras.Input(shape=single_input_shape) for _ in range(num_single_inputs)]


middle_layer = layers.Dense(num_middle_units)


middle_layers_outs = []
for i,inp in enumerate(input_layers):
    middle_layers_outs.append(middle_layer(inp))
    
middle_layer_concat_out = Concatenate()(middle_layers_outs)

output_layer = Dense(num_single_inputs ,activation="softmax")

output_layer_out = output_layer(middle_layer_concat_out)


model = keras.Model(inputs = input_layers, outputs = output_layer_out,name= "player")
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

#Prediction:

#Data
x = np.array([[90., 89., 91.],[68., 61., 73.]])

model.predict(x)

**ERROR**:
AssertionError: Could not compute output Tensor("dense_59/Softmax:0", shape=(None, 2), dtype=float32)

错误是什么？连接方法还是该零件后的形状