Update a subset of weights in TensorFlow

Question

Does anyone know how to update a subset (i.e. only some indices) of the weights that are used in the forward propagation?

My guess is that I might be able to do that after applying compute_gradients as follows:

optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
grads_vars = optimizer.compute_gradients(loss, var_list=[weights, bias_h, bias_v])

...and then do something with the list of tuples in grads_vars.

score 9 · Answer 1 · answered Jan 22 '16 at 01:41

9

You could use a combination of gather and scatter_update. Here's an example that doubles the values at position 0 and 2

indices = tf.constant([0,2])
data = tf.Variable([1,2,3])
data_subset = tf.gather(data, indices)
updated_data_subset = 2*data_subset
sparse_update = tf.scatter_update(data, indices, updated_data_subset)
init_op = tf.initialize_all_variables()

sess = tf.Session()
sess.run([init_op])
print "Values before:", sess.run([data])
sess.run([sparse_update])
print "Values after:", sess.run([data])

You should see

Values before: [array([1, 2, 3], dtype=int32)]
Values after: [array([2, 2, 6], dtype=int32)]

answered Jan 22 '16 at 01:41

Yaroslav Bulatov

57,332
22
139
197

Thanks, I can use that with a 1D tensor. What about a 2D matrix (weights in my case)? – Luca Jan 22 '16 at 12:01
Reshape to 1d vector, compute linear indices, reshape back into matrix – Yaroslav Bulatov Jan 22 '16 at 15:50
if I do `data = tf.Variable([[1,2,3],[4,5,6]]) data = tf.reshape(data, [-1])` I get `TypeError: Input 'ref' of 'ScatterUpdate' Op requires l-value input` – Luca Jan 23 '16 at 17:10
1

Hm...good point. One work-around is to keep your variables as 1D vectors, and apply reshape when you need it as a matrix – Yaroslav Bulatov Jan 23 '16 at 18:02

score 1 · Answer 2 · answered Apr 07 '16 at 10:00

1

Easiest way is to pull the tf.Variable into python (as a numpy array) using npvar = sess.run(tfvar), then perform some operation on it such as npvar[1, 2] = -10. Then you can upload the modified data back into tensorflow using sess.run(tfvar.assign(npvar)).

Obviously this is very slow and not really useful for training but it does work.

answered Apr 07 '16 at 10:00

user38157

351
2
2

Really great! Maybe this way slows down a bit, but if used not inside the training loop, but simply at the very beginning, shaping custom values on subparts of weights matrix it is a very effective and simple way to make initialisation. – Stefano Scarpanti Nov 08 '17 at 14:21

score 0 · Answer 3 · answered Jul 24 '21 at 09:11

# in TF2.0 you can solve with "tensor_scatter_nd_update"
# for example:
tensor = [0, 0, 0, 0, 0, 0, 0, 0]  # tf.rank(tensor) == 1
indices = [[1], [3], [4], [7]]  # num_updates == 4, index_depth == 1
updates = [9, 10, 11, 12]  # num_updates == 4
print(tf.tensor_scatter_nd_update(tensor, indices, updates))
# tf.Tensor([ 0 9  0 10  11  0  0 12], shape=(8,), dtype=int32)

Update a subset of weights in TensorFlow

3 Answers3

Linked