# SPDX-FileCopyrightText: Copyright © 2023 Idiap Research Institute <contact@idiap.ch>
#
# SPDX-License-Identifier: GPL-3.0-or-later
""":py:mod:`DRIU network architecture <.segment.models.driu>`, with :py:class:`torch.nn.PixelShuffle` instead of ConvTrans2D."""
import logging
import typing
import torch
import torch.nn
import torch.utils.data
from ...data.typing import TransformSequence
from .backbones.vgg import vgg16_for_segmentation
from .driu import ConcatFuseBlock
from .losses import SoftJaccardBCELogitsLoss
from .make_layers import UpsampleCropBlock
from .model import Model
logger = logging.getLogger(__name__)
[docs]
class DRIUPIXHead(torch.nn.Module):
"""DRIUPIX head module. DRIU with pixelshuffle instead of ConvTrans2D.
Parameters
----------
in_channels_list : list
Number of channels for each feature map that is returned from backbone.
"""
def __init__(self, in_channels_list):
super().__init__()
(
in_conv_1_2_16,
in_upsample2,
in_upsample_4,
in_upsample_8,
) = in_channels_list
self.conv1_2_16 = torch.nn.Conv2d(in_conv_1_2_16, 16, 3, 1, 1)
# Upsample layers
self.upsample2 = UpsampleCropBlock(in_upsample2, 16, 4, 2, 0, pixelshuffle=True)
self.upsample4 = UpsampleCropBlock(
in_upsample_4, 16, 8, 4, 0, pixelshuffle=True
)
self.upsample8 = UpsampleCropBlock(
in_upsample_8, 16, 16, 8, 0, pixelshuffle=True
)
# Concat and Fuse
self.concatfuse = ConcatFuseBlock()
[docs]
def forward(self, x):
hw = x[0]
conv1_2_16 = self.conv1_2_16(x[1]) # conv1_2_16
upsample2 = self.upsample2(x[2], hw) # side-multi2-up
upsample4 = self.upsample4(x[3], hw) # side-multi3-up
upsample8 = self.upsample8(x[4], hw) # side-multi4-up
return self.concatfuse(conv1_2_16, upsample2, upsample4, upsample8)
[docs]
class DRIUPix(Model):
""":py:mod:`DRIU network architecture <.segment.models.driu>`, with :py:class:`torch.nn.PixelShuffle` instead of ConvTrans2D.
Parameters
----------
loss_type
The loss to be used for training and evaluation.
.. warning::
The loss should be set to always return batch averages (as opposed
to the batch sum), as our logging system expects it so.
loss_arguments
Arguments to the loss.
optimizer_type
The type of optimizer to use for training.
optimizer_arguments
Arguments to the optimizer after ``params``.
scheduler_type
The type of scheduler to use for training.
scheduler_arguments
Arguments to the scheduler after ``params``.
model_transforms
An optional sequence of torch modules containing transforms to be
applied on the input **before** it is fed into the network.
augmentation_transforms
An optional sequence of torch modules containing transforms to be
applied on the input **before** it is fed into the network.
num_classes
Number of outputs (classes) for this model.
pretrained
If True, will use VGG16 pretrained weights.
"""
def __init__(
self,
loss_type: type[torch.nn.Module] = SoftJaccardBCELogitsLoss,
loss_arguments: dict[str, typing.Any] = {},
optimizer_type: type[torch.optim.Optimizer] = torch.optim.Adam,
optimizer_arguments: dict[str, typing.Any] = {},
scheduler_type: type[torch.optim.lr_scheduler.LRScheduler] | None = None,
scheduler_arguments: dict[str, typing.Any] = {},
model_transforms: TransformSequence = [],
augmentation_transforms: TransformSequence = [],
num_classes: int = 1,
pretrained: bool = False,
):
super().__init__(
name="driu-pix",
loss_type=loss_type,
loss_arguments=loss_arguments,
optimizer_type=optimizer_type,
optimizer_arguments=optimizer_arguments,
scheduler_type=scheduler_type,
scheduler_arguments=scheduler_arguments,
model_transforms=model_transforms,
augmentation_transforms=augmentation_transforms,
num_classes=num_classes,
)
self.pretrained = pretrained
self.backbone = vgg16_for_segmentation(
pretrained=self.pretrained,
return_features=[3, 8, 14, 22],
)
self.head = DRIUPIXHead([64, 128, 256, 512])
[docs]
def forward(self, x):
x = self.normalizer(x)
x = self.backbone(x)
return self.head(x)
[docs]
def set_normalizer(self, dataloader: torch.utils.data.DataLoader) -> None:
"""Initialize the normalizer for the current model.
This function is NOOP if ``pretrained = True`` (normalizer set to
imagenet weights, during contruction).
Parameters
----------
dataloader
A torch Dataloader from which to compute the mean and std.
Will not be used if the model is pretrained.
"""
if self.pretrained:
from .normalizer import make_imagenet_normalizer
logger.warning(
f"ImageNet pre-trained {self.name} model - NOT "
f"computing z-norm factors from train dataloader. "
f"Using preset factors from torchvision.",
)
self.normalizer = make_imagenet_normalizer()
else:
super().set_normalizer(dataloader)