3D classification with different input size

[YingdiZhang]

Thanks for the excellent work.
I am doing 3D classification task using DenseNet. The tutorial resize the input to （96, 96, 96），I wonder if I could input other sizes volume to the DenseNet，because my data mostly with size of （384. 384, 11）

[wyli]

thanks, please use the discussions tab instead. I'm converting this to Discussions.

[Nic-Ma]

Hi @yiheng-wang-nv ,

Could you please help share the support input data shape of our DenseNet?

Thanks in advance.

[rijobro]

Hi @yingdiLee, there are no hard constraints on image size when using DenseNet. (You may need to make sure that your images are divisible by a 2**n factor if convolution strides are two, I can't remember you'll have to check but that should be a simple test.)

If you want to use images of different sizes compared to one another, you could either use batch size of 1, or make use of our pad_list_data_collate collation function in your DataLoader. This will pad all images in your batch to match their sizes.

[YingdiZhang]

Hi @rijobro Thanks for your reply. when i use densenet121, I resize my input volume into （12, 384, 384）and got an error：
File "/home/zyd/Nutstore Files/beifen_xinsuoqu_/MedSeg/MedSeg/train/Task904TrainerDDP.py", line 413, in
trainer.train(max_epochs=max_epochs, device=device)
File "/home/zyd/Nutstore Files/beifen_xinsuoqu_/MedSeg/MedSeg/train/Task904TrainerDDP.py", line 112, in train
outputs = self.model(batch_data["image"].to(device))
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/monai/networks/nets/densenet.py", line 247, in forward
x = self.features(x)
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/torch/nn/modules/container.py", line 119, in forward
input = module(input)
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/torch/nn/modules/container.py", line 119, in forward
input = module(input)
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/zyd/anaconda3/envs/med/lib/python3.9/site-packages/torch/nn/modules/pooling.py", line 700, in forward
return F.avg_pool3d(input, self.kernel_size, self.stride,
RuntimeError: input image (T: 1 H: 48 W: 48) smaller than kernel size (kT: 2 kH: 2 kW: 2)

should I change my input size？

[rijobro]

import torch
from monai.networks.nets import DenseNet121

net = DenseNet121(spatial_dims=3, in_channels=1, out_channels=1).cuda()
net(torch.rand((1, 1, 32, 384, 384)).cuda())  # works
net(torch.rand((1, 1, 16, 384, 384)).cuda())  # doesn't work

Using the default shape of DenseNet121, you can see that the minimum size required is 32 in all dimensions. You can add a DivisblePad to your list of transforms to ensure this.