tf.GradientTape 给出无渐变

Question

I'm trying to write a custom training loop. After creating the model, I have added some extra trainable parameter to some layers of my model. I have used these extra parameters to update my original parameter on every forward pass. But when I'm calculating the gradient, it's giving None for the extra parameter that i have added last. Code is given below:

model = Sequential()
model.add(tf.keras.layers.Flatten(input_shape=(1,1)))
model.add(Dense(1, activation='relu'))
model.add(Dense(2, activation='softmax'))

model.layers[1].add_weight(name="x1", shape=(1,), initializer=tf.keras.initializers.Constant(value=1.0),trainable=True)

dataset = tf.data.Dataset.from_tensor_slices((feature, labels))

for i, (x_batch_train, y_batch_train) in enumerate(dataset):
    with tf.GradientTape() as tape:
        for par in model.layers[1].trainable_weights:
            if "x1" in par.name:
                bits = tf.convert_to_tensor(par)
        for par in model.layers[1].trainable_weights:
            if "kernel" in par.name:
                par = bits + 1.0    
        x = model(x_batch_train, training = True)
        loss = tf.keras.losses.SparseCategoricalCrossentropy(y_batch_train, x)
        val = tape.gradient(loss, model.trainable_weights)
        for v in val:
            print(v)

Here, I have added one extra parameter called x1 and it's updating the kernel of Dense layer. But I'm getting None gradient for x1 parameter. The output is:

tf.Tensor([[0.]], shape=(1, 1), dtype=float32)
tf.Tensor([-0.], shape=(1,), dtype=float32)
None
tf.Tensor([[0. 0.]], shape=(1, 2), dtype=float32)
tf.Tensor([-0.5  0.5], shape=(2,), dtype=float32)

Why it's happening?

Answer 1

The problem is that the changes you are making to the layer's weights have no direct connection to the output of the model in the context of tf.GradientTape and are therefore not tracked. You could solve this with a simple custom layer:

import tensorflow as tf

class DenseLayer(tf.keras.layers.Layer):
    def __init__(self, units=1):
        super(DenseLayer, self).__init__()
        self.units = units
    def build(self, input_shape):
        self.w = self.add_weight("kernel",
                              shape=[int(input_shape[-1]),
                                      self.units], trainable=True)
        self.b = self.add_weight(shape=(self.units,), initializer="zeros", trainable=True)
        self.bits = self.add_weight(name="x1", shape=[int(input_shape[-1]),
                                      self.units], initializer=tf.keras.initializers.ones(), trainable=True)

    def call(self, inputs):
        return tf.nn.relu(tf.matmul(inputs, (self.w + self.bits + 1.0)) + self.b)

dense_layer = DenseLayer(1)
model = tf.keras.Sequential()
model.add(tf.keras.layers.Flatten(input_shape=(1,1)))
model.add(dense_layer)
model.add(tf.keras.layers.Dense(2, activation='softmax'))
print(model.summary())
dataset = tf.data.Dataset.from_tensor_slices((tf.random.normal((50, 1, 1)), tf.random.uniform((50, ), maxval=2, dtype=tf.int32))).batch(2)
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy()
optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)
for i, (x_batch_train, y_batch_train) in enumerate(dataset):
    with tf.GradientTape() as tape:
        y = model(x_batch_train, training = True)
        loss = loss_fn(y_batch_train, y)
        val = tape.gradient(loss, model.trainable_weights)
        for v in val:
            print(v)
    optimizer.apply_gradients(zip(val, model.trainable_variables))

Answer 2

Your idea is good I didn't extend from the last answer but this question is asked once about the custom layer and that you can do it for lstm by training as model.fit( ... )

It is not about the Gradient Tape.

[ Sample - Dense ]:

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Class / Function
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
class MyDenseLayer(tf.keras.layers.Layer):
    def __init__(self, num_outputs, num_add):
        super(MyDenseLayer, self).__init__()
        self.num_outputs = num_outputs
        self.num_add = num_add
        
    def build(self, input_shape):
        self.kernel = self.add_weight("kernel",
        shape=[int(input_shape[-1]),
        self.num_outputs])

    def call(self, inputs):
        temp = tf.add( inputs, self.num_add )
        temp = tf.matmul(temp, self.kernel)
        return temp

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Model Initialize
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
model = tf.keras.models.Sequential([
    tf.keras.layers.InputLayer(input_shape=( 32, 32, 4 )),
    tf.keras.layers.Normalization(mean=3., variance=2.),
    tf.keras.layers.Normalization(mean=4., variance=6.),
    tf.keras.layers.Conv2D(32, (3, 3), activation='relu'),
    tf.keras.layers.MaxPooling2D((2, 2)),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Reshape((128, 225)),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(96, return_sequences=True, return_state=False)),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(96)),
])

layer = MyDenseLayer(10, 5)

model.add(layer)
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(192, activation='relu'))
model.add(tf.keras.layers.Dense(10, activation='softmax'))
model.summary()

[ Output ]: