带有 Tensorflow 2.1 输出形状问题的自定义层

Question

我正在尝试让自定义层返回 (25,1) 的张量，但是应该传递一个 batch_size（我从下一层得到一个错误）。我找了一些例子，但不知道如何指定输出形状。

此外，我需要一个独立于输入大小的任意输出形状，因为计算（不是以下示例的一部分）将始终返回固定数量的值。

我尝试了以下方法：

class SimpleLayer(layers.Layer):
    def __init__(self, **kwargs):
        super(SimpleLayer,  self).__init__(**kwargs)
        self.baseline = tf.Variable(initial_value=0.1, trainable=True)

    def call(self, inputs):
        print ("in call inputs:", inputs.shape)
        ret = tf.zeros((25, 1)) + self.baseline
        print("Ret:", ret, "Shape", tf.shape(ret))
        return (ret)

返回：

Ret: Tensor("om/add:0", shape=(25, 1), dtype=float32) Shape Tensor("om/Shape:0", shape=(2,), dtype=int32)
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
inputs (InputLayer)          [(None, 150, 1)]          0         
_________________________________________________________________
dense (Dense)                (None, 150, 256)          512       
_________________________________________________________________
om (SimpleLayer)             (25, 1)                   1         
=================================================================

但这确实会产生输出形状 (25, 1) 而不是 (None, 25, 1)。

然后我尝试了：

class SimpleLayer(layers.Layer):
    def __init__(self, **kwargs):
        super(SimpleLayer,  self).__init__(**kwargs)
        self.baseline = tf.Variable(initial_value=0.1, trainable=True)

    def call(self, inputs):
        print ("in call inputs:", inputs.shape)
        ret = tf.zeros((25, 1)) + self.baseline
        return (ret)

并得到错误：

TypeError: Expected int32, got None of type 'NoneType' instead.

有什么建议吗？

Answer 1

我建议您使用调用方法中定义的输入数据，否则该层没有意义

我提供了一个虚拟示例并且完美运行

class SimpleLayer(tf.keras.layers.Layer):
    def __init__(self, **kwargs):
        super(SimpleLayer,  self).__init__(**kwargs)
        self.baseline = tf.Variable(initial_value=0.1, trainable=True)

    def call(self, inputs):
        ret = inputs + self.baseline
        return (ret)

    def compute_output_shape(self, input_shape):
        return (input_shape[0], input_shape[1], input_shape[2])

使用 SimpleLayer 创建模型

inp = Input(shape=(25,1))
x = SimpleLayer()(inp)
out = Dense(3)(x)
model = Model(inp, out)
model.summary()

总结：

Layer (type)                 Output Shape              Param #   
=================================================================
input_10 (InputLayer)        [(None, 25, 1)]           0         
_________________________________________________________________
simple_layer_16 (SimpleLayer (None, 25, 1)             1         
_________________________________________________________________
dense_22 (Dense)             (None, 25, 3)             6         
=================================================================
Total params: 7
Trainable params: 7
Non-trainable params: 0

---

编辑

我尝试用这种方式覆盖None维度的问题

class SimpleLayer(tf.keras.layers.Layer):
    def __init__(self, **kwargs):
        super(SimpleLayer,  self).__init__(**kwargs)
        self.baseline = tf.Variable(initial_value=0.1, trainable=True, dtype=tf.float64)

    def call(self, inputs):
        ret = tf.zeros((1, 25, 1), dtype=tf.float64) + self.baseline
        ret = tf.compat.v1.placeholder_with_default(ret, (None, 25, 1))
        return (ret)

inp = Input((150,1))
x = Dense(256)(inp)
x = SimpleLayer()(x)
x = Dense(10)(x)

model = Model(inp, x)
model.summary()

总结：

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_34 (InputLayer)        [(None, 150, 1)]          0         
_________________________________________________________________
dense_68 (Dense)             (None, 150, 256)          512       
_________________________________________________________________
simple_layer_9 (SimpleLayer) (None, 25, 1)             1         
_________________________________________________________________
dense_69 (Dense)             (None, 25, 10)            20        
=================================================================
Total params: 533
Trainable params: 533
Non-trainable params: 0