Single GPU ImageNet training: is there a big difference between Lasagne vs Torch vs Caffe?

[IAI]

I plan on playing around with ImageNet on a single 4GB gpu. I am not looking to reproduce state of the art results by any stretch. I am looking to train and just get good enough results (whatever that may mean, I am not familiar with the top benchmarks at the moment).

I'm very comfortable with Lasagne and Theano in general. From what I understand Caffe has multi-GPU support and this is one of the main reasons benchmarkers prefer Caffe over a Theano flavored implementation. I believe Torch also offers some multi-GPU support?

Since the multi-GPU support is a moot point for me, I'm wondering if anyone can comment on the differences between Caffe, Lasagne (or Theano in general perhaps) and Torch for single GPU ImageNet training?

Also since I'm not looking to get the very very best performance, is it feasible to train ImageNet on a single 4GB card?

I really appreciate any experience you may have had.

[Sander Dieleman]

It should be fine if your model is not too large, and if you keep the batch size relatively low. You could also consider simply downsampling the input images to 128x128 or something, the performance will be almost as good and your experiments will be approximately 4 times faster :)

Sander

[Lasagne-Eater]

Hey Sander, thank you for the reply. Oddly enough after I read your reply, I read a Kaggle interview with Yann LeCun and he apparently pushed for ImageNet 128x128 being the next benchmark dataset to take the place of the CIFAR-10/100. But I guess because the ImageNet dataset is not publicly available and researchers are interested in benchmarking and keeping experimental conditions consistent to allow for comparison of methodology, they stayed with the standard 256x256 configurations?

[Sander Dieleman]

There is no real standard size for ImageNet, as the source images are all different sizes. I believe the latest inception models use 299x299 input.

Even if you stick to the AlexNet 'default' of 256x256, there are still many ways in which you could rescale the source images to that size (different aspect ratios, cropping, offsets, etc.), so this is hard to standardize.

Sander