Data augmentation on CPU？

[jiujing23333]

From my observation, you modified the data augmentation code from the caffe version. But it still cannot run on GPU. How can I do it by GPU or multi-CPU? Thanks.

[fperezgamonal]

Hello,

I am running into the same problem. Have you found a solution? As I am working in a cluster, I have found that I have to wait a lot to allocate the enough amount of CPUs (+memory) and the GPU to successfully run data augmentation without running out of memory.

When I try to run it in the GPU ('/gpu:0') in debug mode I get: Chromatic transform not yet implemented on GPU multiple times. I'll try to find to exactly which transformation chromatic transform refers to and try disabling them I re-running on the GPU.

Cheers.

[fperezgamonal]

I tried to disable all transformation done regarding colour (dataset_configs.py from 'image_b' downwards) to no avail. For now I use Data augmentation on the CPU successfully but it requires a high number of resources and that means I have to wait longer for my job to enter the server's queue.

I'll try to take a look at the PyTorch implementation to see if they perform the augmentation on the GPU and try to "port it". I'll update you if I can manage to do it.