Dataset in GPU memory: updating specific elements in existing tf.Variable

[tomrunia]

My machine has a Titan X card so I would like to use the memory as efficient as possible to avoid expensive data transfer between the CPU and GPU. Therefore, I want to have as many images as possible (let's say 5 GB) on the GPU memory inside a tf.Variable or tf.constant. However, I also want to update these data tensors after a number of iterations (e.g. a replay memory). For training, I then only need to send sample indices to the GPU and use tf.slice to generate a training batch.

My question is what the correct way is to update specific elements in an existing tf.Variable? I already found methods using tf.scatter_update and tf.slice:

http://stackoverflow.com/questions/34685947/adjust-single-value-within-tensor-tensorflow
http://stackoverflow.com/questions/37593960/set-k-largest-elements-of-a-tensor-to-zero-in-tensorflow

But these methods seem cumbersome and tricky. Are there better ways of updating existing variables?

[girving]

Some of this is covered #206. For single variables and unordered sets of elements tf.scatter_update is the right thing, and #206 proposes tf.assign_slice_* which cover what you want. Our hope is to make most of these available with something like x[i].assign(...), but there are a few steps before we get there.

If your satisfied with #206 I'll close as duplicate, but let me know if not!

[tomrunia]

Hi Geoffrey, thanks for your reply. I am satisfied with the conversation and proposed solutions in #206. You can close this one. Hopefully the assign operations will arrive soon!