Add "repeat" op & gradients / tf.image.resizeBilinear gradients about tfjs HOT 5 CLOSED

For TF implementation of the resizeBilinear gradient we will implement:
https://github.com/tensorflow/tensorflow/blob/9832df/tensorflow/core/kernels/resize_bilinear_op.cc#L275

We have a tf.tile() op, does that suffice? What are you looking for with repeat?

from tfjs.

Awesome to hear about resizeBilinear gradient! As for repeat, I was looking for an op that would perform an element-wise repeat (tensorflow/tensorflow#8246), eg: repeating

[[1, 2],
 [3, 4]]

with [2, 2] would yield:

[[1, 1, 2, 2],
 [1, 1, 2, 2],
 [3, 3, 4, 4],
 [3, 3, 4, 4]]

I was going to use that as an upsampling op to match the tf.maxPool()s on the other side of my autoencoder.

Slightly related: I attempted implementing the solution (tile & reshape) posted in the above TensorFlow issue, but hit an exception when attempting to call tf.tile() on rank 5 tensors.

from tfjs.

Are you trying to repeat on a 4D tensor? Just curious, what is the use-case?

We actually haven't done any work with rank-5 Tensors, but we can add support if there's a driving use case here.

from tfjs.

The only reason I would have needed rank 5 tensors would be to implement the makeshift "repeat" solution posted here: tensorflow/tensorflow#8246 , which I only needed in absence of tf.resizeBilinear gradients. I was trying to achieve the same thing: "upsample" a [batchSize, height, width, numChannels] tensor to [batchSize, height * 2, width * 2, numChannels]. Sorry for the confusion.

from tfjs.

Hi @tehZevo , gradient for resizeBilinear was added in tensorflow/tfjs-core#996 . Make sure to use the latest tfjs. I'm going to close this issue, as per its title, we've fixed it, but feel free to reopen if you run into any issues with resizeBilinear or its gradient.

from tfjs.