lucidrains / x-unet Goto Github PK

Hi @lucidrains , currently if I set frame_kernel_size=1 it throws an error. In CT scan we will have single channel and width, height and depth. Anyway I can do this? Thanks in advance.

AssertionError "height and width must be power of two" will be raised if shape of inputs are 384, 224, 192 etc.
It is possible to forward model using well designed nested_unet_depths and removing #587.

import torch
from x_unet import XUnet

unet = XUnet(
    dim = 64,
    dim_mults = (1, 2, 4, 8),
    nested_unet_depths = (4, 3, 2, 1),
    consolidate_upsample_fmaps = True
)

img = torch.randn(1, 3, 384, 384)
out = unet(img)
print(out.size())

Hi! In the original paper, it looks like convnext uses kernel sizes of 7, then 1, then 1. But it looks like the implementation here is using 7-3-3. Is this intentional? Is 7-3-3 known to work better?

Thank you!

How to make the model work with non-square images ?

Hi, @lucidrains, thanks for your great repo, I just wonder if you have the plan to add some benchmark testing results of your x-net,(PSACL VOC, COCO, ADE20K, etc.) or some advice/docs about the performance of x-net.
This is really helpful for everyone to select the correct network.

Many thanks again!

1. Trans-Unet https://arxiv.org/abs/2102.04306

2. U^2-Net https://arxiv.org/abs/2005.09007

3. Restormer https://arxiv.org/abs/2111.09881

Hi there,

thanks a lot for all your great repos and implementations!

I've wanted to try this for a segmentation problem and I've had issues training on colabs 40GB GPU with dimensions 256x256.
The Model I've wanted to use is initialized like so:

gen = XUnet(
        dim = target_shape,
        channels = 3,
        dim_mults = (1, 2, 4, 4),
        nested_unet_depths = (4, 3, 2, 1),     # nested unet depths, from unet-squared paper
        consolidate_upsample_fmaps = True,     # whether to consolidate outputs from all upsample blocks, used in unet-squared paper
).to(device)

Is there a trick or what do you estimate the needed Memory is?
I set pin_memory to false, which improved it a little, but still wasn't able to do a single pass (batch_size = 1).

I also noticed most of the memory is reserved, and not allocated, irrespective of the initial size? (always around 35 - 38 GB).