Do you use this model on other downstream work? (like semantic segmentation ) about convmixer HOT 3 CLOSED

We haven't tried it on any tasks other than what was mentioned in the paper (i.e., image classification on CIFAR-10 and ImageNet-1k).

Feel free to reopen this issue or open another if you have more questions.

from convmixer.

We haven't tried it on any tasks other than what was mentioned in the paper (i.e., image classification on CIFAR-10 and ImageNet-1k).

Feel free to reopen this issue or open another if you have more questions.

I try it on semantic segmentation yesterday, but it's cost to much gpu memory(because i change the patch size, patch_size=4, I also change the image resolution, network depth, but it's not decrease too much memory when trining ). so limit of the gpu memory, I have to stop the experiment .

from convmixer.

ConvMixers are relatively compute- and memory-hungry due to the large-kernel depthwise convolutions. We believe these are poorly-optimized by deep learning frameworks, and we are currently investigating low-level optimizations to improve this.

That said, if you want to make ConvMixer less expensive, you should increase the patch size (e.g., try patch_size=14). (I'm assuming that your inputs are quite large for semantic segmentation.)

from convmixer.