minerva.models.nets.image.unetplusplus_resnet50¶

Classes¶

`ConvBlock`	Base class for all neural network modules.
`DeepLabV3ResNet50Backbone`	Base class for all neural network modules.
`LitUNetPlusPlusDeepLabV3`	Base class for all neural network modules.
`UNetPlusPlusDeepLabV3`	UNet++ with DeepLabV3 ResNet50 backbone for semantic segmentation.

Module Contents¶

class minerva.models.nets.image.unetplusplus_resnet50.ConvBlock(in_channels, out_channels)[source]¶

Bases: torch.nn.Module

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes:

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call to(), etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

Variables:

training (bool) – Boolean represents whether this module is in training or evaluation mode.

Parameters:

in_channels (int)
out_channels (int)

Double convolution block for UNet++ decoder.

Applies two consecutive 3x3 convolutions, each followed by batch normalization and ReLU activation, commonly used in U-Net architectures for feature refinement.

Parameters¶

in_channelsint: Number of input channels.
out_channelsint: Number of output channels.

conv¶

forward(x)[source]¶

Forward pass through the double convolution block.

Parameters¶

xtorch.Tensor: Input feature tensor of shape (batch_size, in_channels, H, W).

Parameters:: x (torch.Tensor)
Return type:: torch.Tensor

class minerva.models.nets.image.unetplusplus_resnet50.DeepLabV3ResNet50Backbone(pretrained=True)[source]¶

Bases: torch.nn.Module

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes:

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call to(), etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

Variables:: training (bool) – Boolean represents whether this module is in training or evaluation mode.
Parameters:: pretrained (bool)

Notes¶

The dilation rates are set as follows: - layer3: dilation=2, stride=1 - layer4: dilation=4, stride=1 This preserves the spatial resolution at H/8, W/8 for c3, c4, and c5.

References¶

recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).

bn1¶

conv1¶

forward(x)[source]¶

Forward pass through the ResNet50 backbone.

Parameters¶

xtorch.Tensor: Input tensor of shape (batch_size, 3, H, W).

Parameters:: x (torch.Tensor)
Return type:: Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

layer1¶

layer2¶

layer3¶

layer4¶

maxpool¶

relu¶

class minerva.models.nets.image.unetplusplus_resnet50.LitUNetPlusPlusDeepLabV3(in_channels=3, num_classes=6, deep_supervision=True, lr=0.0003, pretrained=True)[source]¶

Bases: lightning.LightningModule

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes:

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call to(), etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

Variables:

training (bool) – Boolean represents whether this module is in training or evaluation mode.

Parameters:

in_channels (int)
num_classes (int)
deep_supervision (bool)
lr (float)
pretrained (bool)

PyTorch Lightning module for UNet++ with DeepLabV3 backbone.

Wraps the UNet++ DeepLabV3 model with training, validation, and testing loops, optimizer configuration, and metrics for multi-class segmentation.

Parameters¶

in_channelsint, optional: Number of input image channels (default is 3).
num_classesint, optional: Number of segmentation classes (default is 6).
deep_supervisionbool, optional: If True, enables deep supervision training (default is True).
lrfloat, optional: Learning rate for the optimizer (default is 3e-4).
pretrainedbool, optional: If True, uses ImageNet pre-trained backbone (default is True).

Notes¶

Metrics include accuracy, F1-score, mean IoU, precision, and recall, all computed using torchmetrics.

configure_optimizers()[source]¶

Configure optimizer for training.

Return type:: torch.optim.Optimizer

forward(x)[source]¶

Forward pass through the model.

Parameters¶

xtorch.Tensor: Input batch of images of shape (batch_size, in_channels, H, W).

Parameters:: x (torch.Tensor)
Return type:: Union[torch.Tensor, List[torch.Tensor]]

loss_fn¶

lr = 0.0003¶

model¶

num_classes = 6¶

test_accuracy¶

test_f1¶

test_miou¶

test_precision¶

test_recall¶

test_step(batch, batch_idx)[source]¶

Test step for one batch with metrics evaluation.

Parameters¶

batchtuple of torch.Tensor: Batch containing (images, masks).
batch_idxint: Index of the current batch.

Parameters:

batch (Tuple[torch.Tensor, torch.Tensor])
batch_idx (int)

Return type:

torch.Tensor

training_step(batch, batch_idx)[source]¶

Training step for one batch.

Parameters¶

batchtuple of torch.Tensor: Batch containing (images, masks).
batch_idxint: Index of the current batch.

Parameters:

batch (Tuple[torch.Tensor, torch.Tensor])
batch_idx (int)

Return type:

torch.Tensor

validation_step(batch, batch_idx)[source]¶

Validation step for one batch.

Parameters¶

batchtuple of torch.Tensor: Batch containing (images, masks).
batch_idxint: Index of the current batch.

Parameters:

batch (Tuple[torch.Tensor, torch.Tensor])
batch_idx (int)

Return type:

torch.Tensor

class minerva.models.nets.image.unetplusplus_resnet50.UNetPlusPlusDeepLabV3(in_channels=3, num_classes=6, deep_supervision=True, pretrained=True)[source]¶

Bases: torch.nn.Module

UNet++ with DeepLabV3 ResNet50 backbone for semantic segmentation.

Combines DeepLabV3’s multi-scale feature extraction with UNet++’s nested skip connections for robust semantic segmentation, particularly suited for seismic image segmentation tasks.

Parameters¶

in_channelsint, optional: Number of input channels (default is 3 for RGB images).
num_classesint, optional: Number of segmentation classes (default is 6).
deep_supervisionbool, optional: If True, enables deep supervision with auxiliary losses (default is True).
pretrainedbool, optional: If True, uses ImageNet pre-trained ResNet50 backbone (default is True).

Notes¶

The architecture includes: - DeepLabV3 ResNet50 backbone with dilated convolutions. - UNet++ decoder with nested skip connections for feature refinement. - Bilinear upsampling to restore original input resolution. - Optional deep supervision for improved training stability.

References¶

Unet++: A nested u-net architecture for medical image segmentation. In Deep learning in medical image analysis and multimodal learning for clinical decision support (pp. 3-11). Springer. .. [2] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).

_safe_upsample(x, target)[source]¶

Upsample tensor to match target spatial dimensions.

Handles cases where spatial dimensions may not align due to rounding errors in pooling or upsampling operations.

Parameters¶

xtorch.Tensor: Tensor to be upsampled.
targettorch.Tensor: Reference tensor with target spatial dimensions.

Parameters:

x (torch.Tensor)
target (torch.Tensor)

Return type:

torch.Tensor

backbone¶

conv0_1¶

conv0_2¶

conv0_3¶

conv1_1¶

conv1_2¶

conv2_1¶

deep_supervision = True¶

forward(x)[source]¶

Forward pass through the UNet++ DeepLabV3 network.

Parameters¶

xtorch.Tensor: Input tensor of shape (batch_size, in_channels, H, W).

Parameters:: x (torch.Tensor)
Return type:: Union[torch.Tensor, List[torch.Tensor]]

proj0¶

proj1¶

proj2¶

proj3¶

up¶

up_conv¶

Parameters:

in_channels (int)
num_classes (int)
deep_supervision (bool)
pretrained (bool)