models.attention package

Subpackages

models.attention.urr package

Submodules

models.attention.lightning_module module

LightningModule wrappers for U-Net with Attention mechanism on residual frames.

class models.attention.lightning_module.ResidualAttentionLightningModule(batch_size: int, metric: Metric | None = None, loss: Module | str | None = None, model_type: ModelType = ModelType.UNET, encoder_name: str = 'resnet34', encoder_depth: int = 5, encoder_weights: str | None = 'imagenet', in_channels: int = 3, classes: int = 1, num_frames: Literal[5, 10, 15, 20, 30] = 5, weights_from_ckpt_path: str | None = None, optimizer: Optimizer | str = 'adamw', optimizer_kwargs: dict[str, Any] | None = None, scheduler: LRScheduler | str = 'gradual_warmup_scheduler', scheduler_kwargs: dict[str, Any] | None = None, multiplier: int = 2, total_epochs: int = 50, alpha: float = 1.0, _beta: float = 0.0, learning_rate: float = 0.0001, dl_classification_mode: ClassificationMode = ClassificationMode.MULTICLASS_MODE, eval_classification_mode: ClassificationMode = ClassificationMode.MULTICLASS_MODE, residual_mode: ResidualMode = ResidualMode.SUBTRACT_NEXT_FRAME, loading_mode: LoadingMode = LoadingMode.RGB, dump_memory_snapshot: bool = False, flat_conv: bool = False, unet_activation: str | None = None, attention_reduction: Literal['sum', 'prod', 'cat', 'weighted', 'weighted_learnable'] = 'sum', attention_only: bool = False, dummy_predict: DummyPredictMode = DummyPredictMode.NONE, temporal_conv_type: TemporalConvolutionalType = TemporalConvolutionalType.ORIGINAL, metric_mode: MetricMode = MetricMode.INCLUDE_EMPTY_CLASS, metric_div_zero: float = 1.0, single_attention_instance: bool = False)

Bases: CommonModelMixin

Attention mechanism-based U-Net.

batch_size: Batch size of dataloader.

in_channels: Number of image channels.

num_frames: Number of frames used.

dump_memory_snapshot: Whether to dump a memory snapshot.

dummy_predict: Whether to simply return the ground truth for visualisation.

residual_mode: Residual frames generation mode.

optimizer: type[Optimizer] | Optimizer | str: Optimizer for training.

optimizer_kwargs: dict[str, Any]: Optimizer kwargs.

scheduler: type[LRScheduler] | LRScheduler | str: Scheduler for training.

scheduler_kwargs: dict[str, Any]: Scheduler kwargs.

loading_mode: Image loading mode.

multiplier: Learning rate multiplier.

total_epochs: int: Number of total epochs for training.

alpha: Loss scaling factor.

learning_rate: Learning rate for training.

classes: int: Number of segmentation classes.

weights_from_ckpt_path: Model checkpoint path to load weights from.

forward(x_img: Tensor, x_res: Tensor) → Tensor: Forward pass of the model.

log_metrics(prefix) → None

Implement shared metric logging epoch end here.

Note: This is to prevent circular imports with the logging module.

training_step(batch: tuple[Tensor, Tensor, Tensor, str], batch_idx: int) → Tensor

Forward pass for the model with dataloader batches.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.

Returns:

Training loss.

Raises:

AssertionError – Prediction shape and ground truth mask shapes are different.

validation_step(batch: tuple[Tensor, Tensor, Tensor, str], batch_idx: int)

Forward pass for the model for one minibatch of a validation epoch.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.

test_step(batch: tuple[Tensor, Tensor, Tensor, str], batch_idx: int)

Forward pass for the model for one minibatch of a test epoch.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.

predict_step(batch: tuple[Tensor, Tensor, Tensor, str | list[str]], batch_idx: int, dataloader_idx: int = 0) → tuple[Tensor, Tensor, str | list[str]]

Forward pass for the model for one minibatch of a test epoch.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.
dataloader_idx – Index of the dataloader.

Returns:

Mask predictions, original images, and filename.

configure_optimizers()

Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple. Optimization with multiple optimizers only works in the manual optimization mode.

Returns:

Any of these 6 options.

Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple lr_scheduler_config).
Dictionary, with an "optimizer" key, and (optionally) a "lr_scheduler" key whose value is a single LR scheduler or lr_scheduler_config.
None - Fit will run without any optimizer.

The lr_scheduler_config is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.

lr_scheduler_config = {
    # REQUIRED: The scheduler instance
    "scheduler": lr_scheduler,
    # The unit of the scheduler's step size, could also be 'step'.
    # 'epoch' updates the scheduler on epoch end whereas 'step'
    # updates it after a optimizer update.
    "interval": "epoch",
    # How many epochs/steps should pass between calls to
    # `scheduler.step()`. 1 corresponds to updating the learning
    # rate after every epoch/step.
    "frequency": 1,
    # Metric to to monitor for schedulers like `ReduceLROnPlateau`
    "monitor": "val_loss",
    # If set to `True`, will enforce that the value specified 'monitor'
    # is available when the scheduler is updated, thus stopping
    # training if not found. If set to `False`, it will only produce a warning
    "strict": True,
    # If using the `LearningRateMonitor` callback to monitor the
    # learning rate progress, this keyword can be used to specify
    # a custom logged name
    "name": None,
}

When there are schedulers in which the .step() method is conditioned on a value, such as the torch.optim.lr_scheduler.ReduceLROnPlateau scheduler, Lightning requires that the lr_scheduler_config contains the keyword "monitor" set to the metric name that the scheduler should be conditioned on.

# The ReduceLROnPlateau scheduler requires a monitor
def configure_optimizers(self):
    optimizer = Adam(...)
    return {
        "optimizer": optimizer,
        "lr_scheduler": {
            "scheduler": ReduceLROnPlateau(optimizer, ...),
            "monitor": "metric_to_track",
            "frequency": "indicates how often the metric is updated",
            # If "monitor" references validation metrics, then "frequency" should be set to a
            # multiple of "trainer.check_val_every_n_epoch".
        },
    }


# In the case of two optimizers, only one using the ReduceLROnPlateau scheduler
def configure_optimizers(self):
    optimizer1 = Adam(...)
    optimizer2 = SGD(...)
    scheduler1 = ReduceLROnPlateau(optimizer1, ...)
    scheduler2 = LambdaLR(optimizer2, ...)
    return (
        {
            "optimizer": optimizer1,
            "lr_scheduler": {
                "scheduler": scheduler1,
                "monitor": "metric_to_track",
            },
        },
        {"optimizer": optimizer2, "lr_scheduler": scheduler2},
    )

Metrics can be made available to monitor by simply logging it using self.log('metric_to_track', metric_val) in your LightningModule.

Note

Some things to know:

Lightning calls .backward() and .step() automatically in case of automatic optimization.
If a learning rate scheduler is specified in configure_optimizers() with key "interval" (default “epoch”) in the scheduler configuration, Lightning will call the scheduler’s .step() method automatically in case of automatic optimization.
If you use 16-bit precision (precision=16), Lightning will automatically handle the optimizer.
If you use torch.optim.LBFGS, Lightning handles the closure function automatically for you.
If you use multiple optimizers, you will have to switch to ‘manual optimization’ mode and step them yourself.
If you need to control how often the optimizer steps, override the optimizer_step() hook.

models.attention.model module

Implementation of residual frames-based attention layers.

class models.attention.model.AttentionLayer(embed_dim: int, num_frames: int, num_heads: int = 1, key_embed_dim: int | None = None, value_embed_dim: int | None = None, need_weights: bool = False, reduce: Literal['sum', 'prod', 'cat', 'weighted', 'weighted_learnable'] = 'sum', one_instance: bool = False)

Bases: Module

Attention mechanism between spatio-temporal and spatial embeddings.

As the spatial dimensions of the image can be considered the sequence to be processed, the channel dimension must be the embedding dimension for each part of Q, K, and V tensors.

forward(q: Tensor, ks: Tensor, vs: Tensor) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class models.attention.model.SpatialAttentionBlock(temporal_conv: OneD | DilatedOneD | Temporal3DConv | None, attention: AttentionLayer, num_frames: int, reduce: Literal['sum', 'prod', 'cat', 'weighted', 'weighted_learnable'], reduce_dim: int = 0, one_instance: bool = False, _attention_only: bool = False)

Bases: Module

Residual block with attention mechanism between spatio-temporal embeddings.

forward(st_embeddings: Tensor, res_embeddings: Tensor, return_o1: bool = False) → Tensor | tuple[Tensor, Tensor]

Forward pass of the residual block.

Parameters:

st_embeddings – Spatio-temporal embeddings from raw frames.
res_embeddings – Spatial embeddings from residual frames.
return_o1 – Whether to return o1.

models.attention.segmentation_model module

Implementation of residual frames-based U-Net and U-Net++ architectures.

class models.attention.segmentation_model.ResidualAttentionUnet(*args, **kwargs)

Bases: SegmentationModel

U-Net with Attention mechanism on residual frames.

check_input_shape(x): Check if the input shape is divisible by the output stride. If not, raise a RuntimeError.

initialize() → None

property encoder: Get the encoder of the model.

forward(regular_frames: Tensor, residual_frames: Tensor) → Tensor

Forward pass of the model.

Parameters:

regular_frames – Regular frames from the sequence.
residual_frames – Residual frames from the sequence.

Returns:

Predicted mask logits.

predict(regular_frames: Tensor, residual_frames: Tensor) → Tensor

Inference method. Switch model to eval mode, call .forward(x) with torch.no_grad()

Parameters:: x – 4D torch tensor with shape (batch_size, channels, height, width)
Returns:: 4D torch tensor with shape (batch_size, classes, height, width)
Return type:: prediction

class models.attention.segmentation_model.ResidualAttentionUnetPlusPlus(*args, **kwargs)

Bases: ResidualAttentionUnet

U-Net++ with Attention mechanism on residual frames.

models.attention.utils module

Helper classes and typedefs for the residual frames-based attention models.

Module contents

Residual frames-based attention U-Net and U-Net++ implementation.

class models.attention.ResidualAttentionLightningModule(batch_size: int, metric: Metric | None = None, loss: Module | str | None = None, model_type: ModelType = ModelType.UNET, encoder_name: str = 'resnet34', encoder_depth: int = 5, encoder_weights: str | None = 'imagenet', in_channels: int = 3, classes: int = 1, num_frames: Literal[5, 10, 15, 20, 30] = 5, weights_from_ckpt_path: str | None = None, optimizer: Optimizer | str = 'adamw', optimizer_kwargs: dict[str, Any] | None = None, scheduler: LRScheduler | str = 'gradual_warmup_scheduler', scheduler_kwargs: dict[str, Any] | None = None, multiplier: int = 2, total_epochs: int = 50, alpha: float = 1.0, _beta: float = 0.0, learning_rate: float = 0.0001, dl_classification_mode: ClassificationMode = ClassificationMode.MULTICLASS_MODE, eval_classification_mode: ClassificationMode = ClassificationMode.MULTICLASS_MODE, residual_mode: ResidualMode = ResidualMode.SUBTRACT_NEXT_FRAME, loading_mode: LoadingMode = LoadingMode.RGB, dump_memory_snapshot: bool = False, flat_conv: bool = False, unet_activation: str | None = None, attention_reduction: Literal['sum', 'prod', 'cat', 'weighted', 'weighted_learnable'] = 'sum', attention_only: bool = False, dummy_predict: DummyPredictMode = DummyPredictMode.NONE, temporal_conv_type: TemporalConvolutionalType = TemporalConvolutionalType.ORIGINAL, metric_mode: MetricMode = MetricMode.INCLUDE_EMPTY_CLASS, metric_div_zero: float = 1.0, single_attention_instance: bool = False)

Bases: CommonModelMixin

Attention mechanism-based U-Net.

configure_optimizers()

Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple. Optimization with multiple optimizers only works in the manual optimization mode.

Returns:

Any of these 6 options.

Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple lr_scheduler_config).
Dictionary, with an "optimizer" key, and (optionally) a "lr_scheduler" key whose value is a single LR scheduler or lr_scheduler_config.
None - Fit will run without any optimizer.

The lr_scheduler_config is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.

lr_scheduler_config = {
    # REQUIRED: The scheduler instance
    "scheduler": lr_scheduler,
    # The unit of the scheduler's step size, could also be 'step'.
    # 'epoch' updates the scheduler on epoch end whereas 'step'
    # updates it after a optimizer update.
    "interval": "epoch",
    # How many epochs/steps should pass between calls to
    # `scheduler.step()`. 1 corresponds to updating the learning
    # rate after every epoch/step.
    "frequency": 1,
    # Metric to to monitor for schedulers like `ReduceLROnPlateau`
    "monitor": "val_loss",
    # If set to `True`, will enforce that the value specified 'monitor'
    # is available when the scheduler is updated, thus stopping
    # training if not found. If set to `False`, it will only produce a warning
    "strict": True,
    # If using the `LearningRateMonitor` callback to monitor the
    # learning rate progress, this keyword can be used to specify
    # a custom logged name
    "name": None,
}

When there are schedulers in which the .step() method is conditioned on a value, such as the torch.optim.lr_scheduler.ReduceLROnPlateau scheduler, Lightning requires that the lr_scheduler_config contains the keyword "monitor" set to the metric name that the scheduler should be conditioned on.

# The ReduceLROnPlateau scheduler requires a monitor
def configure_optimizers(self):
    optimizer = Adam(...)
    return {
        "optimizer": optimizer,
        "lr_scheduler": {
            "scheduler": ReduceLROnPlateau(optimizer, ...),
            "monitor": "metric_to_track",
            "frequency": "indicates how often the metric is updated",
            # If "monitor" references validation metrics, then "frequency" should be set to a
            # multiple of "trainer.check_val_every_n_epoch".
        },
    }


# In the case of two optimizers, only one using the ReduceLROnPlateau scheduler
def configure_optimizers(self):
    optimizer1 = Adam(...)
    optimizer2 = SGD(...)
    scheduler1 = ReduceLROnPlateau(optimizer1, ...)
    scheduler2 = LambdaLR(optimizer2, ...)
    return (
        {
            "optimizer": optimizer1,
            "lr_scheduler": {
                "scheduler": scheduler1,
                "monitor": "metric_to_track",
            },
        },
        {"optimizer": optimizer2, "lr_scheduler": scheduler2},
    )

Metrics can be made available to monitor by simply logging it using self.log('metric_to_track', metric_val) in your LightningModule.

Note

Some things to know:

Lightning calls .backward() and .step() automatically in case of automatic optimization.
If a learning rate scheduler is specified in configure_optimizers() with key "interval" (default “epoch”) in the scheduler configuration, Lightning will call the scheduler’s .step() method automatically in case of automatic optimization.
If you use 16-bit precision (precision=16), Lightning will automatically handle the optimizer.
If you use torch.optim.LBFGS, Lightning handles the closure function automatically for you.
If you use multiple optimizers, you will have to switch to ‘manual optimization’ mode and step them yourself.
If you need to control how often the optimizer steps, override the optimizer_step() hook.

forward(x_img: Tensor, x_res: Tensor) → Tensor: Forward pass of the model.

log_metrics(prefix) → None

Implement shared metric logging epoch end here.

Note: This is to prevent circular imports with the logging module.

predict_step(batch: tuple[Tensor, Tensor, Tensor, str | list[str]], batch_idx: int, dataloader_idx: int = 0) → tuple[Tensor, Tensor, str | list[str]]

Forward pass for the model for one minibatch of a test epoch.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.
dataloader_idx – Index of the dataloader.

Returns:

Mask predictions, original images, and filename.

test_step(batch: tuple[Tensor, Tensor, Tensor, str], batch_idx: int)

Forward pass for the model for one minibatch of a test epoch.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.

training_step(batch: tuple[Tensor, Tensor, Tensor, str], batch_idx: int) → Tensor

Forward pass for the model with dataloader batches.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.

Returns:

Training loss.

Raises:

AssertionError – Prediction shape and ground truth mask shapes are different.

validation_step(batch: tuple[Tensor, Tensor, Tensor, str], batch_idx: int)

Forward pass for the model for one minibatch of a validation epoch.

Parameters:

batch – Batch of frames, residual frames, masks, and filenames.
batch_idx – Index of the batch in the epoch.

dl_classification_mode: ClassificationMode: Classification mode for the dataloader instances.

eval_classification_mode: ClassificationMode: Classification mode for the evaluation process.

dice_metrics: dict[str, MetricCollection | Metric]: A collection of dice score variants.

other_metrics: dict[str, MetricCollection]: A collection of other metrics (recall, precision, jaccard).

hausdorff_metrics: dict[str, MetricCollection]: Just hausdorff distance metrics.

infarct_metrics: dict[str, MetricCollection]: A collection of infarct-related clinical heuristics.

model: nn.Module: The internal model used.

model_type: ModelType: The architecture of the model, if appropriate.

de_transform: Compose | InverseNormalize: The inverse transformation from augmentation of the samples by the dataloaders.

classes: int: Number of segmentation classes.

optimizer: type[Optimizer] | Optimizer | str: Optimizer for training.

optimizer_kwargs: dict[str, Any]: Optimizer kwargs.

total_epochs: int: Number of total epochs for training.

scheduler: type[LRScheduler] | LRScheduler | str: Scheduler for training.

scheduler_kwargs: dict[str, Any]: Scheduler kwargs.

prepare_data_per_node: bool

allow_zero_length_dataloader_with_multiple_devices: bool

training: bool

batch_size: Batch size of dataloader.

in_channels: Number of image channels.

num_frames: Number of frames used.

dump_memory_snapshot: Whether to dump a memory snapshot.

dummy_predict: Whether to simply return the ground truth for visualisation.

residual_mode: Residual frames generation mode.

loading_mode: Image loading mode.

multiplier: Learning rate multiplier.

alpha: Loss scaling factor.

learning_rate: Learning rate for training.

weights_from_ckpt_path: Model checkpoint path to load weights from.