Theano shared updating last element in python

Question

I have a shared variable persistent_vis_chain which is being updated by a theano function where it gets its function from a theano.scan, But thats not the problem just back story.

My shared variable looks like D = [image1, ... , imageN] where each images is [x1,x2,...,x784].

What I want to do is take the average of all the images and put them into the last imageN. That is I want to sum all the values in each image except the last 1, which will result in [s1,s2,...,s784] then I want to set imageN = [s1/len(D),s2/len(D),...s784/len(D)]

So my problem is I do not know how to do this with theano.shared and may be with my understanding of theano functions and doing this computation with symbolic variables. Any help would be greatly appreciated.

Answer 1

If you have N images, each of shape 28x28=784 then, presumably, your shared variable has shape (N,28,28) or (N,784)? This method should work with either shape.

Given D is your shared variable containing your image data. If you want to get the average image then D.mean(keepdims=True) will give it to you symbolically.

It's unclear if you want to change the final image to equal the mean image (sounds like a strange thing to do), or if you want to add a further N+1'th image to the shared variable. For the former you could do something like this:

D = theano.shared(load_D_data())
D_update_expression = do_something_with_scan_to_get_D_update_expression(D)
updates = [(D, T.concatenate(D_update_expression[:-1],
                             D_update_expression.mean(keepdims=True)))]
f = theano.function(..., updates=updates)

If you want to do the latter (add an additional image), change the updates line as follows:

updates = [(D, T.concatenate(D_update_expression,
                             D_update_expression.mean(keepdims=True)))]

Note that this code is intended as a guide. It may not work as it stands (e.g. you may need to mess with the axis= parameter in the T.concatenate command).

The point is that you need to construct a symbolic expression explaining what the new value for D looks like. You want it to be a combination of the updates from scan plus this additional average thing. T.concatenate allows you to combine those two parts together.