Cntk: Transforms with a composite MinibatchSource

Created on 6 Sep 2017 · 6Comments · Source: microsoft/CNTK

Hi, I am attempting to use a composite MinibatchSource with some ImageDeserializers and need to do some transforms like random crop and random flip. But in my case I need the transforms to be the same for all images with the same sequenceID.
For example, if the first image (sequenceID = 0) of the first ImageDeserializer is flipped, then the first image (sequenceID = 0) of the second ImageDeserializer also needs to be flipped.

Is that possible? Will the transforms be the same for all ImageDeserializers in a composite reader?

bug

Source

patykov

Most helpful comment

We will try to address it in version 2.3 (version 2.2 will be released very soon)

n17s on 14 Sep 2017

👍2

All 6 comments

To verify this,you can do the following: take the example code from this SO thread, provide identical files to both readers, add some random transforms, to both sides, and create a "network" that just takes the two inputs and computes the squared error of their difference. If the output of the network is zero (or very close to zero) then it means that the same transforms are applied. If not, then let us know so we can fix it.

n17s on 7 Sep 2017

The transforms (random cropping and flipping) will applied to both readers only when minibatch size is 1. For minibatch size > 1, the two readers does not get the same transforms.