d-li14 / lambda.pytorch

PyTorch implementation of Lambda Network and pretrained Lambda-ResNet

lambda.pytorch's Issues

A better speed-accuracy trade-off

In the appendix E.2, the author says using lambda layer in c4 and c5 stage is good enough for the result, while keeping the speed not too slow.

Which dataset did you use for training the model?

What subset of imagenet did you use for training your model? Can you provide me an exact link? Complete imagenet dataset is of more than 1.1 TB. Thanks

Training time

Hi @d-li14 thank you for the repo! I have a couple of questions:

1. How much time did it take for you to train the model? I'm training it now and with 8 v100 is taking ~3h per epoch on Imagenet which seems to be a lot! I obtained the same params and FLOPS as you did. Could you perhaps share the training script you used?

2. In the paper they report 78.4 for LambdaResnet50 while you obtained 78.2, do you think this difference is not significative? Or perhaps there are some differences in the engineering tricks used by the authors?

Thank you!