在 Functional API 模型中添加截断层

Add a Truncated layer in Functional API model

我正在尝试在以下代码中的 Conv2D 层之后添加截断层:

input_layer = Input(shape=(256, 256, 1))
conv = Conv2D(8, (5, 5), padding='same', strides=1, use_bias=False)(input_layer)
output_layer = Activation(activation='tanh')(lambda_layer)
output_layer = AveragePooling2D(pool_size= (5, 5), strides=2)(output_layer)
output_layer = BatchNormalization()(output_layer)

截断层必须满足:

−T if x < −T
x if −T ≤ x ≤ T
T  x > T 

where `T` is a threshold value, x= the output of convolution layer`

有人可以帮我构建这一层吗?

谢谢

您可以使用 tf.clip_by_valuetf.stop_gradient 来保留梯度,因为 tf.clip_by_value 不可微分。最后,将其包裹在 Lambda 层中:

import functools
def clip_preserve_grad(inp, clip_min, clip_max):
  return inp + tf.stop_gradient(tf.clip_by_value(inp, clip_min, clip_max) - inp)

T = 0.5
trunc_func = functools.partial(clip_preserve_grad, clip_min=-T, clip_max=T)
trunc = tf.keras.layers.Lambda(trunc_func)

使用 Lambda 层:

>>> a = tf.random.normal((1,10))
>>> a
<tf.Tensor: shape=(1, 10), dtype=float32, numpy=
array([[-1.8041286 , -0.11153453, -0.84555113,  0.8489615 ,  0.12237629,
         1.3350475 ,  0.619644  , -0.5498301 , -0.6082269 ,  0.8465021 ]],
      dtype=float32)>
>>> trunc(a)
<tf.Tensor: shape=(1, 10), dtype=float32, numpy=
array([[-0.5       , -0.11153453, -0.5       ,  0.5       ,  0.12237629,
         0.5       ,  0.5       , -0.5       , -0.5       ,  0.5       ]],
      dtype=float32)>

您可以使用 tensorflow.keras.backed.switch 构建所需的函数并将其包装在 Lambda 层中

构建并测试函数:

T = 5
X = tf.constant(np.random.uniform(-10, 10, (3,5)))

def switch_func(X, T):
    
    zeros = tf.zeros_like(X)
    T_matrix = tf.ones_like(X) * T

    cond1 = K.switch(X < -T_matrix, -T_matrix, zeros)
    cond2 = K.switch(X > T_matrix, T_matrix, zeros)
    cond3 = K.switch(tf.abs(cond1 + cond2) == T, zeros, X)
    res = cond1 + cond2 + cond3
    return res

switch_func(X, T)

<tf.Tensor: shape=(3, 5), dtype=float64, numpy=
array([[-5.        ,  0.65807168, -4.93481499, -5.        , -2.94954848],
       [-1.25114075, -5.        ,  2.97657545,  5.        , -0.8958152 ],
       [-1.26611956,  5.        , -3.38477137,  5.        , -3.53358454]])>

模型内部用法:

X = np.random.uniform(0,1, (100,10))
y = np.random.uniform(0,1, (100,))

inp = Input((10,))
x = Dense(8)(inp)
x = Lambda(lambda x: switch_func(x, T=0.5))(x)
out = Dense(1)(x)

model = Model(inp, out)
model.compile('adam', 'mse')
model.fit(X,y, epochs=3)