# SPDX-FileCopyrightText: Copyright © 2023 Idiap Research Institute <contact@idiap.ch>
#
# SPDX-License-Identifier: GPL-3.0-or-later
"""Holistically-Nested Edge Detection (HED) network architecture, from :cite:p:`xie_holistically-nested_2015`."""
import logging
import typing
import torch
import torch.nn
import torch.utils.data
from ...data.typing import TransformSequence
from ..losses import MultiLayerSoftJaccardAndBCELogitsLoss
from .backbones.vgg import vgg16_for_segmentation
from .make_layers import UpsampleCropBlock, conv_with_kaiming_uniform
from .model import Model
logger = logging.getLogger(__name__)
[docs]
class ConcatFuseBlock(torch.nn.Module):
"""Take in five feature maps with one channel each, concatenates thems and applies a
1x1 convolution with 1 output channel.
"""
def __init__(self):
super().__init__()
self.conv = conv_with_kaiming_uniform(5, 1, 1, 1, 0)
[docs]
def forward(self, x1, x2, x3, x4, x5):
x_cat = torch.cat([x1, x2, x3, x4, x5], dim=1)
return self.conv(x_cat)
[docs]
class HEDHead(torch.nn.Module):
"""HED head module.
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_upsample_16,
) = in_channels_list
self.conv1_2_16 = torch.nn.Conv2d(in_conv_1_2_16, 1, 3, 1, 1)
# Upsample
self.upsample2 = UpsampleCropBlock(in_upsample2, 1, 4, 2, 0)
self.upsample4 = UpsampleCropBlock(in_upsample_4, 1, 8, 4, 0)
self.upsample8 = UpsampleCropBlock(in_upsample_8, 1, 16, 8, 0)
self.upsample16 = UpsampleCropBlock(in_upsample_16, 1, 32, 16, 0)
# Concat and Fuse
self.concatfuse = ConcatFuseBlock()
[docs]
def forward(self, x):
hw = x[0]
conv1_2_16 = self.conv1_2_16(x[1])
upsample2 = self.upsample2(x[2], hw)
upsample4 = self.upsample4(x[3], hw)
upsample8 = self.upsample8(x[4], hw)
upsample16 = self.upsample16(x[5], hw)
concatfuse = self.concatfuse(
conv1_2_16, upsample2, upsample4, upsample8, upsample16
)
return (upsample2, upsample4, upsample8, upsample16, concatfuse)
[docs]
class HED(Model):
"""Holistically-Nested Edge Detection (HED) network architecture, from :cite:p:`xie_holistically-nested_2015`.
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.
pretrained
If True, will use VGG16 pretrained weights.
"""
def __init__(
self,
loss_type: type[torch.nn.Module] = MultiLayerSoftJaccardAndBCELogitsLoss,
loss_arguments: dict[str, typing.Any] | None = None,
optimizer_type: type[torch.optim.Optimizer] = torch.optim.Adam,
optimizer_arguments: dict[str, typing.Any] | None = None,
scheduler_type: type[torch.optim.lr_scheduler.LRScheduler] | None = None,
scheduler_arguments: dict[str, typing.Any] | None = None,
model_transforms: TransformSequence | None = None,
augmentation_transforms: TransformSequence | None = None,
pretrained: bool = False,
):
super().__init__(
name="hed",
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=1, # fixed at current implementation
)
if pretrained:
from ..normalizer import make_imagenet_normalizer
Model.normalizer.fset(self, make_imagenet_normalizer()) # type: ignore[attr-defined]
self.backbone = vgg16_for_segmentation(
pretrained=pretrained, return_features=[3, 8, 14, 22, 29]
)
self.head = HEDHead([64, 128, 256, 512, 512])
[docs]
def forward(self, x):
x = self.normalizer(x)
x = self.backbone(x)
return self.head(x)