[NengoDL] Signals.scatter() shape issue

Backends
1

I’m working on porting my custom learning rule to NengoDL and feel I’m 90% there. There are just a few issue to iron out, the main one being that signals.scatter() is having trouble dealing with my Tensor shapes.

I have a matrix of resistance values that I would like to read and update at each timestep in my NengoDL learning rule.
The matrix is defined in my build_mpes() build function and has shape (post.n_neurons, pre.n_neurons), basically the same as weights and delta Signals:

memristors = Signal( shape=(out_size, in_size), name="mPES:memristors", (out_size, in_size) ) )

I then get the corresponding TensorSignal in the init() of my NengoDL SimmPESBuilder class and reshape it to (1, post.n_neurons, pre.n_neurons):

    self.memristors = signals.combine( [ op.memristors for op in ops ] )
    self.memristors = self.pos_memristors.reshape(
            (len( ops ), ops[ 0 ].memristors.shape[ 0 ], ops[ 0 ].memristors.shape[ 1 ])
            )

I do the reshape so that when I get a Tensor of shape (1, 1, post.n_neurons, pre.n_neurons) in the build_step() function by using:

memristors = signals.gather( self.memristors )

I then proceed to update memristors using tf.tensor_scatter_nd_update(), but when I try to write back to the TensorSignal by using:

signals.scatter( self.memristors, memristors )

I get the following error (with post.n_neurons = pre.n_neurons = 4):

ValueError: The outer 2 dimensions of indices.shape=[1,4,2] must match the outer 2 dimensions of updates.shape=[1,1,4,4]: Dimension 1 in both shapes must be equal, but are 4 and 1. Shapes are [1,4] and [1,1]. for ‘TensorGraph/while/iteration_0/SimmPESBuilder/TensorScatterUpdate’ (op: ‘TensorScatterUpdate’) with input shapes: [1,4,4], [1,4,2], [1,1,4,4].

I have looked into nengo_dl/signals.py and the issue seems to be caused by line 349: dst_shape[dst.minibatched] = dst.shape[0] (https://github.com/nengo/nengo-dl/blob/72242a2e826c172ddfe57cd5731ef7b6070315a2/nengo_dl/signals.py#L349):

    # align val shape with dst base shape
    val.shape.assert_is_fully_defined()
    dst_shape = list(base_shape)
    dst_shape[dst.minibatched] = dst.shape[0]
    if val.shape != dst_shape:
        val = tf.reshape(val, dst.tf_shape)

Specifically, on line 348: dst_shape = list(base_shape) the assignment is [1,4,4] i.e., self.memristors.shape, but then the next line assigns it [1,1,4] because dst.minibatches is True.
So when we get to line 367: result = tf.tensor_scatter_nd_update(var, dst.tf_indices_nd, val) the dimensions of var and val don’t line up.

So my guess is that I should drop the reshaping of self.memristors in the init() of SimmPESBuilder (https://github.com/Tioz90/MemristorLearning/blob/b6eaf7a5627f25bbb999ea3f3444b3ad68eeac3d/memristor_nengo/learning_rules.py#L279) in order for the TensorSignal to have shape (4, 4) but then I would have to modify all my learning rule logic, which expects all Tensors to have shape (1, 1, 4, 4), exactly as the built-in implementation of PES does (https://github.com/nengo/nengo-dl/blob/72242a2e826c172ddfe57cd5731ef7b6070315a2/nengo_dl/learning_rule_builders.py#L268).

What would your suggestions be?

2

I made a new branch of my code (https://github.com/Tioz90/MemristorLearning/blob/a8d222857f26457a70a504faebc8019e2e209231/memristor_nengo/learning_rules.py#L270) because I somehow had a feeling that Building a custom learning rule operator for Nengo DL may have helped with my issue.
Why? Because then I can safely remove the reshape() on my memristor TensorSignals needed to add the extra dimension to keep track of the number of `ops .

And, lo and behold, at first glance it seems as if the issue with .scatter() has been resolved without changing anything in build_step()!

I’m not sure if how I enforce the single op in init() is the “proper” way and I would obviously like to keep the operator merging as a possibility, in order to have all the performance possible.
Do you think the behaviour of scatter() can be improved to better deal with a situation like mine? Or is it the case of changing something else in my code?

3

I think just reshaping memristors back to the correct shape ((4, 4)) before the scatter should work won’t it? Like signals.scatter(self.memristors.reshape((4, 4)), memristors)

4

Thanks! The code runs perfectly now! Sometimes the simplest solution is the last we think of :slight_smile:

Would operator merging still work correctly with the .reshape() as you suggested?

5

Yeah, all the merging has already happened before we get to the build_step function, so doing reshapes or anything like that within your build process can’t make things un-mergeable (we just have to make sure that all our reshaping is consistent with the merged data we’re getting as input).

1 Like