search

shrubb / box-convolutions

PyTorch code for the "Deep Neural Networks with Box Convolutions" paper
Apache License 2.0
509 stars 35 forks source link

Did you try you impl BoxConv on Unet(Resnet encoder)? #4

Closed aidonchuk closed 5 years ago

aidonchuk commented 5 years ago

subj How to set max height, max width? Single-process box gen takes too long.

shrubb commented 5 years ago

I didn't, but ENet and ERFNet in our paper are very very similar. They are all typical encoder-decoders of residual blocks with max-poolings and upconvolutions. Anyway, I'll soon upload example model definitions I used in the experiments.

Set max_input_h and max_input_w to maximum height/width of that module's input tensors.

What is box gen? What exactly do you run?

aidonchuk commented 5 years ago

320х320 WxH, box gen method is def reset_parameters(self) box_convolution_module.py

shrubb commented 5 years ago

Right, it's currently not very fast, I will speed it up. Still, I don't expect that you will run reset_parameters() very often. Maybe just once before training.

aidonchuk commented 5 years ago

It's work ok on small HxW with one layer. Pls try add it on UNET-like architecture. Thx for repo it's nice!

shrubb commented 5 years ago

Here's an example of application: ENet. As far as I know, it's UNet-like, except skip connections.

aidonchuk commented 5 years ago

Thx a lot. I'll try it.