Add MedNext implementation #8004
Conversation
Signed-off-by: Suraj Pai <bspai@bwh.harvard.edu>
| return self.conv_out(x) | ||
|
|
||
|
|
||
| class LayerNorm(nn.Module): |
There was a problem hiding this comment.
Do we need this LayerNorm implementation? Or can we just use the torch.nn.LayerNorm? I don't see the channels_last used anywhere.
There was a problem hiding this comment.
This is a copy from the original codebase and can probably be replaced with torch.nn.LayerNorm as 'channels_first' is always assumed.
| self.grn_beta = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1, 1), requires_grad=True) | ||
| self.grn_gamma = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1, 1), requires_grad=True) | ||
|
|
||
| def forward(self, x, dummy_tensor=None): |
There was a problem hiding this comment.
Do we know what the dummy_tensor is used for? I don't see it being used anywhere. This applies to the other forward functions as well
There was a problem hiding this comment.
This also come from a copy from the original codebase and should be removed as it's never supposed to be used.
| self, | ||
| spatial_dims: int = 3, | ||
| init_filters: int = 32, | ||
| in_channels: int = 1, |
There was a problem hiding this comment.
How about setting this to None per default and using LazyConv if the in_channels are set to None?
| blocks_down: list = [2, 2, 2, 2], | ||
| blocks_bottleneck: int = 2, | ||
| blocks_up: list = [2, 2, 2, 2], | ||
| norm_type: str = "group", |
There was a problem hiding this comment.
Should we make this a StrEnum? Makes it easy to see the options. Ofc also applies to the Block
There was a problem hiding this comment.
@johnzielke Can you give me an exemple of its usage in one of network implementation. Because it's not straighforward for me how to use it.
monai/networks/nets/mednext.py
Outdated
| enc_exp_r: int = 2, | ||
| dec_exp_r: int = 2, | ||
| bottlenec_exp_r: int = 2, |
There was a problem hiding this comment.
IMO encoder_expansion_ratio (the same for decoder and bottleneck) is better here. It's a public API, makes it easier to see what it refers to without having to look at the docs and the few letters more don't matter that much
There was a problem hiding this comment.
I've kept this convention similar to the original implementation which uses a joint exp_r argument. Happy to rename it. Agree with the comment
monai/networks/nets/mednext.py
Outdated
| bottlenec_exp_r: int = 2, | ||
| kernel_size: int = 7, | ||
| deep_supervision: bool = False, | ||
| do_res: bool = False, |
There was a problem hiding this comment.
maybe use_residual_connections?
There was a problem hiding this comment.
I think use_res would probably suffice? Similar to args in other APIs like SegResNet such as use_conv_final
There was a problem hiding this comment.
Yeah, my only concern is that res can be sth. like resolution or result.
monai/networks/nets/mednext.py
Outdated
| kernel_size: int = 7, | ||
| deep_supervision: bool = False, | ||
| do_res: bool = False, | ||
| do_res_up_down: bool = False, |
There was a problem hiding this comment.
maybe use_residual_connections_up_down_blocks?
There was a problem hiding this comment.
This sounds very verbose to me and may not be necessary. I think the contraction res is easily understood. If playing with these params, users are encouraged to read the code nonetheless to figure out where it goes.
There was a problem hiding this comment.
How about use_residual_up_down ?
| if dim == "2d": | ||
| self.grn_beta = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1), requires_grad=True) | ||
| self.grn_gamma = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1), requires_grad=True) | ||
| else: | ||
| self.grn_beta = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1, 1), requires_grad=True) | ||
| self.grn_gamma = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1, 1), requires_grad=True) | ||
|
|
There was a problem hiding this comment.
I think this is easier to understand, but the current one is fine as well.
| if dim == "2d": | |
| self.grn_beta = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1), requires_grad=True) | |
| self.grn_gamma = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1), requires_grad=True) | |
| else: | |
| self.grn_beta = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1, 1), requires_grad=True) | |
| self.grn_gamma = nn.Parameter(torch.zeros(1, exp_r * in_channels, 1, 1, 1), requires_grad=True) | |
| if dim == "2d": | |
| grn_parameter_shape = (1,1) | |
| elif dim == "3d": | |
| grn_parameter_shape = (1,1,1) | |
| else: | |
| raise ValueError() | |
| grn_parameter_shape = (1, exp_r * in_channels,) + grn_parameter_shape | |
| self.grn_beta = nn.Parameter(torch.zeros(grn_parameter_shape), requires_grad=True) | |
| self.grn_gamma = nn.Parameter(torch.zeros(grn_parameter_shape), requires_grad=True) | |
|
Can't it be usefull to adapt the factory functions of create_mednext_v1 script inside your |
…edNext variants (S, B, M, L) + integration of remarks from @johnzilke (Project-MONAI#8004 (review)) for renaming class arguments - removal of self defined LayerNorm - linked residual connection for encoder and decoder Signed-off-by: Robin CREMESE <robin.cremese@pasteur.fr>
|
Thanks for the comments @johnzielke and thank you for the updates @rcremese. I'll update my branch to dev and look through the changes in your PR asap. |
…edNext variants (S, B, M, L) + integration of remarks from @johnzilke (Project-MONAI#8004 (review)) for renaming class arguments - removal of self defined LayerNorm - linked residual connection for encoder and decoder Signed-off-by: Robin CREMESE <robin.cremese@pasteur.fr>
…edNext variants (S, B, M, L) + integration of remarks from @johnzilke (Project-MONAI#8004 (review)) for renaming class arguments - removal of self defined LayerNorm - linked residual connection for encoder and decoder Signed-off-by: Robin CREMESE <robin.cremese@pasteur.fr>
Signed-off-by: Suraj Pai <bspai@bwh.harvard.edu>
for more information, see https://pre-commit.ci
Signed-off-by: Suraj Pai <bspai@bwh.harvard.edu>
|
@KumoLiu This implementation should be ready now. Please let me know if you have any comments |
Signed-off-by: Suraj Pai <bspai@bwh.harvard.edu>
|
@KumoLiu Added. Do you think there would be interest to add this as a candidate for Auto3DSeg? I refer to this paper for its performance benchmarking: https://arxiv.org/abs/2404.09556 |
Thank you for bringing this up, it's an interesting suggestion. I believe it could be worthwhile to consider this as a potential candidate for Auto3DSeg. However, before moving forward, I would appreciate hearing others' thoughts and insights on whether this aligns with the current goals and roadmap for Auto3DSeg. cc @mingxin-zheng @dongyang0122 @Nic-Ma @myron |
Fixes #7786
Description
Added MedNext architectures implementation for MONAI.
Since a lot of the code is heavily sourced from the original MedNext repo, https://github.com/MIC-DKFZ/MedNeXt, I wanted to check if there is an attribution policy with regarded to borrowed source code. I've added a derivative notice bellow the monai copyright comment. Let me know if this needs to be changed.
The blocks have been taken almost as is but the network implementation has been changed largely to allow flexible blocks and follow MONAI segresnet styling.
Types of changes
./runtests.sh -f -u --net --coverage../runtests.sh --quick --unittests --disttests.make htmlcommand in thedocs/folder.