TensorFlow，访问变量的内部元素并冻结其中一些元素

Question

在权重张量(假设的矩阵(中，如何从权重矩阵中选择一些元素并将其添加到要冻结的变量列表中，以及要在tensorflow中训练的矩阵的其余元素？ 示例：创建了一个大小为 20*20 的变量 W： 如何挑选出一些元素，如 W[0][1]，W[13][15] 并将它们冻结在优化器中

........
def rnn_cell(rnn_input, state, weight):
with tf.variable_scope('rnn_cell', reuse=True):
W = tf.get_variable('W', [n_inputs + n_neurons, n_neurons])
b = tf.get_variable('b', [1, n_neurons], 
initializer=tf.constant_initializer(0.0))
return (tf.tanh(tf.matmul(tf.concat([rnn_input, state], 1), weight) + b))

part_W = tf.scatter_nd([[0,0]], [W[0][0]], [178,150])
W_2 = part_W + tf.stop_gradient(-part_W + W)
state = init_state
rnn_outputs = []
for rnn_input in rnn_inputs:
state = rnn_cell(rnn_input, state, W_2)
rnn_outputs.append(state)
final_state = rnn_outputs[-1]
logits = fully_connected(final_state, n_outputs, activation_fn=None)
xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y,
logits=logits)
loss = tf.reduce_mean(xentropy)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
training_op = optimizer.minimize(loss)
correct = tf.nn.in_top_k(logits, y, 1)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
init = tf.global_variables_initializer()
with tf.Session() as sess:
init.run()
for epoch in range(n_epochs):
for iteration in range(mnist.train.num_examples // 
batch_size):
X_batch, y_batch = 
mnist.train.next_batch(batch_size)
X_batch = X_batch.reshape((-1, n_steps, 
n_inputs))
h=np.zeros([batch_size,n_neurons])
sess.run(training_op, feed_dict={X: X_batch, y: 
y_batch, p:h})
acc_train = accuracy.eval(feed_dict={X: X_batch, y: y_batch, 
p:h})
q=np.zeros([10000,n_neurons])
acc_test = accuracy.eval(feed_dict={X: X_test, y: y_test,p:q})
print(epoch, "Train accuracy:", acc_train, "Test 
accuracy:",acc_test)

Answer 1

我会将权重与它自己的副本重新组合tf.stop_gradient.例如

import tensorflow as tf
w = tf.Variable(tf.zeros((10, 10)))
mask = tf.cast(tf.random_uniform((10, 10), 0, 2, dtype=tf.int32), tf.bool)
w = tf.where(mask, w, tf.stop_gradient(w))