防止 Keras3 中某些样本的权重更新

Question

我有一个名为

ControlledBackward

的自定义图层，它接受

prev_layer

和

input_layer_of_mask

:

class ControlledBackward(Layer):
  def __init__(self, **kwargs):
    super(ControlledBackward, self).__init__(**kwargs)

  def call(self, inputs):
    mask, prev_layer = inputs

    # Cast the mask to the same type as prev_layer
    mask = Lambda(lambda x: ops.cast(x, dtype=prev_layer.dtype))(mask)
    mask_h = Lambda(lambda x: ops.logical_not(x))(mask)

    # Apply the stop_gradient function on the masked parts
    stopped_gradient_part = Lambda(lambda x: ops.stop_gradient(x))(prev_layer)

    # Multiply stopped gradient part with mask_h
    stopped_gradient_masked = Multiply()([stopped_gradient_part, mask_h])

    # Multiply normal (non-stopped) part with mask
    non_stopped_gradient_part = Multiply()([prev_layer, mask])

    # Add the stopped and non-stopped parts
    return Add()([stopped_gradient_masked, non_stopped_gradient_part])

这对于隐藏层工作正常，但不适用于输出层。考虑这个测试代码，其中

X_mask

是某些样本的梯度阻挡器的标记：

import numpy as np

input_layer = Input(shape=(1,), name='input_layer')
gradient_blocker_mask = Input(shape=(1,), dtype='bool', name='cg')

hidden_a = Dense(1, name='hidden_a')(input_layer)
controlled_hidden_a = ControlledBackward(name='gradient_blocker')([gradient_blocker_mask, hidden_a])

output_a = Dense(1, name='output_a')(controlled_hidden_a)

model = Model(inputs=[input_layer, gradient_blocker_mask], outputs=[output_a])

# print weights
print('hidden_weights', model.get_layer('hidden_a').get_weights())
print('output_weights',model.get_layer('output_a').get_weights())

# dummy data
X = np.array([[42], [3]])
X_mask = np.array([[False], [False]])
y = np.array([[7], [5]])

# pred
y_pred = model.predict([X, X_mask], verbose=0)
print(y_pred)

print('')

# fit
model.compile(optimizer='adam', loss='mse')
model.fit([X, X_mask], y, epochs=100, verbose=0)

# print weights
print('hidden_weights', model.get_layer('hidden_a').get_weights())
print('output_weights',model.get_layer('output_a').get_weights())

# predict
y_pred = model.predict([X, X_mask], verbose=0)
print(y_pred)

# display(plot_model(model, show_shapes=True, show_layer_names=True))

输出：

# Before Train
hidden_weights [array([[1.2073823]], dtype=float32), array([0.], dtype=float32)]
output_weights [array([[0.7683755]], dtype=float32), array([0.], dtype=float32)]
[[38.964367]
 [ 2.783169]]

# After Train
hidden_weights [array([[1.2073823]], dtype=float32), array([0.], dtype=float32)]
output_weights [array([[0.6712855]], dtype=float32), array([-0.09659182], dtype=float32)]
[[33.944336 ]
 [ 2.3349032]]

如您所见，

hidden layer

的权重符合预期；它没有改变权重，但

output layer

仍然更新了它的权重。

那么，如何防止

output layer

的权重根据

X_mask

更新其权重呢？

Answer 1

原来我只需要将

output_layer

参数设置为

prev_layer

:

import numpy as np

input_layer = Input(shape=(1,), name='input_layer')
gradient_blocker_mask = Input(shape=(1,), dtype='bool', name='cg')

hidden_a = Dense(1, name='hidden_a')(input_layer)
output_a = Dense(1, name='output_a')(hidden_a)

controlled_output_a = ControlledBackward(name='gradient_blocker')([gradient_blocker_mask, output_a])

model = Model(inputs=[input_layer, gradient_blocker_mask], outputs=[controlled_output_a])

# print weights
print('hidden_weights', model.get_layer('hidden_a').get_weights())
print('output_weights',model.get_layer('output_a').get_weights())

# dummy data
X = np.array([[42], [3]])
X_mask = np.array([[False], [False]])
y = np.array([[7], [5]])

# pred
y_pred = model.predict([X, X_mask], verbose=0)
print(y_pred)

print('')

# fit
model.compile(optimizer='adam', loss='mse')
model.fit([X, X_mask], y, epochs=100, verbose=0)

# print weights
print('hidden_weights', model.get_layer('hidden_a').get_weights())
print('output_weights',model.get_layer('output_a').get_weights())

# predict
y_pred = model.predict([X, X_mask], verbose=0)
print(y_pred)

display(plot_model(model, show_shapes=True, show_layer_names=True))

输出：

hidden_weights [array([[-0.74082583]], dtype=float32), array([0.], dtype=float32)]
output_weights [array([[1.0905765]], dtype=float32), array([0.], dtype=float32)]
[[-33.932945 ]
 [ -2.4237816]]

hidden_weights [array([[-0.74082583]], dtype=float32), array([0.], dtype=float32)]
output_weights [array([[1.0905765]], dtype=float32), array([0.], dtype=float32)]
[[-33.932945 ]
 [ -2.4237816]]