search

fastmachinelearning / hls4ml

Machine learning on FPGAs using HLS
https://fastmachinelearning.org/hls4ml
Apache License 2.0
1.29k stars 418 forks source link

Convert pytorch resnet18 model to HLS #123

Open garyhujingyao opened 5 years ago

garyhujingyao commented 5 years ago

Hi all,

I really appreciate your efforts to provide this excellent tool to convert pytorch model to HLS.

I am trying to convert a resnet18 model to HLS, and I found that the example-models inside pytorch-to-hls folder only convert the linear layers, so I was wondering how I can convert layers such as conv2d, batchnormalization2d, or maxpool2d to the HLS. It would be very helpful if you can give some examples.

Thanks a lot, Jingyao

nhanvtran commented 5 years ago

hi @garyhujingyao

Thanks for your interest! For PyTorch we do not yet have support for conv layers though we do have some (growing) support in Keras.

For larger networks such as ResNet18, there is also very active work in scaling up network sizes. We can keep you posted here on this issue

garyhujingyao commented 5 years ago

Thanks for your reply @nhanvtran.

I will try myself to convert ResNet18 in Keras, though I think it is not easy :-). Please keep me posted if your guys can scale up network sizes to convert models like ResNets.

mohamedelhewehy commented 2 years ago

Is this not supproted yet ? Also I'm masters student where can I find research areas in hls4ml or using it?

bd0100 commented 1 month ago

Is there any new recent developments in this area?

vloncar commented 1 month ago

If you're willing to get your hands dirty it's possible. There's proof-of-concept work done on ResNet on multiple FPGAs (see here). And if you're willing to do heavy pruning and quantization (to 1 or 2 bits) you might fit into a single chip.