minerva.models.nets

Submodules

Classes

DeepLabV3

A DeeplabV3 with a ResNet50 backbone

MLP

A flexible multilayer perceptron (MLP) implemented as a subclass of nn.Sequential.

SETR_PUP

A modular Lightning model wrapper for supervised learning tasks.

SimpleSupervisedModel

A modular Lightning model wrapper for supervised learning tasks.

UNet

This class is a simple implementation of the U-Net model, which is a

WiseNet

A modular Lightning model wrapper for supervised learning tasks.

Package Contents

class minerva.models.nets.DeepLabV3(backbone=None, pred_head=None, loss_fn=None, learning_rate=0.001, num_classes=6, pretrained=False, weights_path=None, train_metrics=None, val_metrics=None, test_metrics=None, optimizer=torch.optim.Adam, optimizer_kwargs=None, lr_scheduler=None, lr_scheduler_kwargs=None, output_shape=None, freeze_backbone=False, interpolate_mode='bilinear', flatten=False, loss_squeeze=True, loss_long=True)[source]

Bases: minerva.models.nets.base.SimpleSupervisedModel

A DeeplabV3 with a ResNet50 backbone

References

Liang-Chieh Chen, George Papandreou, Florian Schroff, Hartwig Adam. “Rethinking Atrous Convolution for Semantic Image Segmentation”, 2017

Initializes a DeepLabV3 model.

Parameters

backbone: Optional[nn.Module]

The backbone network. Defaults to None, which will use a ResNet50 backbone.

pred_head: Optional[nn.Module]

The prediction head network. Defaults to None, which will use a DeepLabV3PredictionHead with specified number of classes.

loss_fn: Optional[nn.Module]

The loss function. Defaults to None, which will use a CrossEntropyLoss.

learning_rate: float

The learning rate for the optimizer. Defaults to 0.001.

num_classes: int

The number of classes for prediction. Defaults to 6.

pretrained: bool

Whether to use pretrained weights. Defaults to False.

weights_path: Optional[str]

Path to local pretrained weights file. If provided with pretrained=True, loads weights from this path instead of downloading. Defaults to None.

train_metrics: Optional[Dict[str, Metric]]

The metrics to be computed during training. Defaults to None.

val_metrics: Optional[Dict[str, Metric]]

The metrics to be computed during validation. Defaults to None.

test_metrics: Optional[Dict[str, Metric]]

The metrics to be computed during testing. Defaults to None.

optimizer: type

Optimizer class to be instantiated. By default, it is set to torch.optim.Adam. Should be a subclass of torch.optim.Optimizer (e.g., torch.optim.SGD).

optimizer_kwargsdict, optional

Additional kwargs passed to the optimizer constructor.

lr_schedulertype, optional

Learning rate scheduler class to be instantiated. By default, it is set to None, which means no scheduler will be used. Should be a subclass of torch.optim.lr_scheduler.LRScheduler (e.g., torch.optim.lr_scheduler.StepLR).

lr_scheduler_kwargsdict, optional

Additional kwargs passed to the scheduler constructor.

output_shape: Optional[Tuple[int, …]]

The output shape of the model. If None, the output shape will be the same as the input shape. Defaults to None. This is useful for models that require a specific output shape, that is different from the input shape.

freeze_backbone: bool

Whether to freeze the backbone weights during training. Defaults to False.

interpolate_mode: Optional[str]

The interpolation mode to use when upscaling the output to the desired output shape. Defaults to “bilinear”. Other options include “nearest”, “bicubic”, etc. See PyTorch documentation for torch.nn.functional.interpolate for all options. Use None to disable upscaling.

flatten: bool

Whether to flatten the output of the backbone before passing it to the prediction head. Defaults to False. Set to True for classification tasks where the prediction head is a fully connected layer.

loss_squeeze: bool

Whether to squeeze the target tensor in the loss function. Defaults to True. This is useful for segmentation tasks where the target tensor has a singleton channel dimension (e.g., shape (B, 1, H, W)) and the loss function expects shape (B, H, W).

loss_long: bool

Whether to convert the target tensor to long type in the loss function. Defaults to True. This is useful for classification tasks where the target tensor is of integer type.

_loss_func(y_hat, y)[source]

Computes the loss between predictions and ground truth.

Parameters

y_hatTensor

Predicted tensor of shape (batch_size, num_classes, height, width)

yTensor

Ground truth tensor of shape (batch_size, 1, height, width)

Parameters:

  • y_hat (torch.Tensor)

  • y (torch.Tensor)

Return type:

torch.Tensor

forward(x)[source]

Performs the forward pass of the DeepLabV3 model.

Parameters

xTensor

Input tensor of shape (batch_size, channels, height, width)

Returns

Tensor

Output tensor of shape (batch_size, num_classes, height, width)

Parameters:

x (torch.Tensor)

Return type:

torch.Tensor

interpolate_mode = 'bilinear'
loss_long = True
output_shape = None
squeeze_loss = True
Parameters:

  • backbone (Optional[torch.nn.Module])

  • pred_head (Optional[torch.nn.Module])

  • loss_fn (Optional[torch.nn.Module])

  • learning_rate (float)

  • num_classes (int)

  • pretrained (bool)

  • weights_path (Optional[str])

  • train_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • val_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • test_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • optimizer (type)

  • optimizer_kwargs (Optional[Dict[str, Any]])

  • lr_scheduler (Optional[type])

  • lr_scheduler_kwargs (Optional[Dict[str, Any]])

  • output_shape (Optional[Tuple[int, Ellipsis]])

  • freeze_backbone (bool)

  • interpolate_mode (Optional[str])

  • flatten (bool)

  • loss_squeeze (bool)

  • loss_long (bool)

class minerva.models.nets.MLP(layer_sizes, activation_cls=nn.ReLU, intermediate_ops=None, final_op=None, *args, **kwargs)[source]

Bases: torch.nn.Sequential

A flexible multilayer perceptron (MLP) implemented as a subclass of nn.Sequential.

This class allows you to quickly build an MLP with: - Custom layer sizes - Configurable activation functions - Optional intermediate operations (e.g., BatchNorm, Dropout) after each linear layer - An optional final operation (e.g., normalization, final activation)

Parameters

layer_sizesSequence[int]

A list of integers specifying the sizes of each layer. Must contain at least two values: the input and output dimensions.

activation_clstype, optional

The activation function class (must inherit from nn.Module) to use between layers. Defaults to nn.ReLU.

intermediate_opsOptional[List[Optional[nn.Module]]], optional

A list of modules (e.g., nn.BatchNorm1d, nn.Dropout) to apply after each linear layer and before the activation. Each item corresponds to one linear layer. Use None to skip an operation for that layer. Must be the same length as the number of linear layers.

final_opOptional[nn.Module], optional

A module to apply after the last layer (e.g., a final activation or normalization).

*args, **kwargs :

Additional arguments passed to the activation function constructor.

Example

>>> from torch import nn
>>> mlp = MLP(
...     [128, 256, 64, 10],
...     activation_cls=nn.ReLU,
...     intermediate_ops=[nn.BatchNorm1d(256), nn.BatchNorm1d(64), None],
...     final_op=nn.Sigmoid()
... )
>>> print(mlp)
MLP(
    (0): Linear(in_features=128, out_features=256, bias=True)
    (1): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
    (3): Linear(in_features=256, out_features=64, bias=True)
    (4): BatchNorm1d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (5): ReLU()
    (6): Linear(in_features=64, out_features=10, bias=True)
    (7): Sigmoid()
)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

  • layer_sizes (Sequence[int])

  • activation_cls (type)

  • intermediate_ops (Optional[List[Optional[torch.nn.Module]]])

  • final_op (Optional[torch.nn.Module])

class minerva.models.nets.SETR_PUP(original_resolution=None, img_size=(512, 512), patch_size=16, in_channels=3, embed_dims=1024, num_layers=24, num_heads=16, out_indices=(9, 14, 19, 23), encoder_stride=None, patch_norm=False, dilatation=1, bias=True, padding_type='corner', mlp_ratio=4, attn_drop_rate=0.0, drop_path_rate=0.0, num_fcs=2, qkv_bias=True, output_cls_token=False, act_type=nn.GELU, with_cp=False, encoder_dropout=0.0, encoder_norm_type=nn.LayerNorm, dropout_type=MMDropPath, cls_token=True, interpolate_mode='bilinear', act_params=None, dropout_params=None, encoder_norm_params=None, decoder_channels=256, decoder_in_index=3, num_classes=6, decoder_dropout=0.0, decoder_norm_type=nn.SyncBatchNorm, decoder_num_convs=4, decoder_up_scale=2, decoder_kernel_size=3, decoder_align_corners=False, decoder_norm_params=None, aux_heads_in_index=(0, 1, 2), aux_head_num_convs=2, aux_head_up_scale=4, aux_weights=None, loss_fn=None, optimizer=Adam, optimizer_kwargs=None, train_metrics=None, val_metrics=None, test_metrics=None, freeze_backbone=False, learning_rate=0.001, loss_weights=None, lr_scheduler=None, lr_scheduler_kwargs=None, head_lr_factor=1.0, use_sliding_inference=True, sliding_window_stride=(341, 341))[source]

Bases: minerva.models.nets.base.SimpleSupervisedModel

A modular Lightning model wrapper for supervised learning tasks.

This class enables the construction of supervised models by combining a backbone (feature extractor), an optional adapter, and a fully connected (FC) head. It provides a clean interface for setting up custom training, validation, and testing pipelines with pluggable loss functions, metrics, optimizers, and learning rate schedulers.

The architecture is structured as follows:

Backbone Model

v

Adapter (Optional)

(Flatten if needed)

v

Fully Connected Head

v

Loss Function

Training and validation steps comprise the following steps:

  1. Forward pass input through the backbone.

  2. Pass through adapter (if provided).

  3. Flatten the output (if flatten is True) before the FC head.

  4. Forward through the FC head.

  5. Compute loss with respect to targets.

  6. Backpropagate and update parameters.

  7. Compute metrics and log them.

  8. Return loss. train_loss, val_loss, and test_loss are always logged, along with any additional metrics specified in the train_metrics, val_metrics, and test_metrics dictionaries.

This wrapper is especially useful to quickly set up supervised models for various tasks, such as image classification, object detection, and segmentation. It is designed to be flexible and extensible, allowing users to easily swap out components like the backbone, adapter, and FC head as needed. The model is built with a focus on simplicity and modularity, making it easy to adapt to different use cases and requirements. The model is designed to be used with PyTorch Lightning and is compatible with its training loop.

Note: For more complex architectures that does not follow the above structure should not inherit from this class.

Note: Input batches must be tuples (input_tensor, target_tensor).

LightningModule implementation for SETR_PUP (SEgmentation TRansformer with Progressive UPsampling).

Parameters

original_resolutiontuple

Resolution of the original images used to pretrain the backbone.

img_sizetuple

Input image size (height, width) used during training and patch embedding.

patch_sizeint

Size of each image patch extracted in the ViT encoder.

in_channelsint

Number of input channels (usually 3 for RGB).

embed_dimsint

Embedding dimension for each patch.

num_layersint

Number of transformer encoder layers.

num_headsint

Number of attention heads in each transformer layer.

out_indicestuple, optional

Indices of the encoder layers to use as features for decoding.

encoder_strideint

Stride used in patch embedding.

patch_normbool

Whether to apply normalization to patch embeddings.

dilatationint

Dilation factor for patch embedding.

biasbool

Whether to include bias in the projection layers.

padding_typestr

Padding mode used in patch embedding (“same” or “corner”).

mlp_ratioint

Expansion ratio for the MLP block inside transformer layers.

attn_drop_ratefloat

Dropout rate applied to attention weights.

drop_path_ratefloat

Probability of dropping entire residual paths (stochastic depth).

num_fcsint

Number of linear layers in the feed-forward MLP of the transformer.

qkv_biasbool

Whether to include bias in QKV projections.

output_cls_tokenbool

Whether to include class token in encoder output.

act_typetype

Activation function class to use (e.g., nn.GELU).

with_cpbool

Whether to enable checkpointing to save memory.

encoder_dropoutfloat

Dropout rate after positional embedding in the encoder.

encoder_norm_typetype

Normalization type used in the encoder.

dropout_typetype

Type of stochastic path dropout layer.

cls_tokenbool

Whether to use a class token in the ViT.

interpolate_modestr

Interpolation mode used for resizing positional embeddings.

act_paramsdict, optional

Additional parameters for the activation function.

dropout_paramsdict, optional

Additional parameters for the dropout layer.

encoder_norm_paramsdict, optional

Additional parameters for the encoder normalization layer.

decoder_channelsint

Number of channels in intermediate layers of the decoder.

decoder_in_indexint

Index into encoder outputs to be used as decoder input.

num_classesint

Number of segmentation classes.

decoder_dropoutfloat

Dropout probability in the decoder.

decoder_norm_typetype

Type of normalization in decoder conv blocks.

decoder_num_convsint

Number of conv+upsample blocks in the decoder.

decoder_up_scaleint

Upsample scale factor for decoder blocks.

decoder_kernel_sizeint

Convolution kernel size in decoder blocks.

decoder_align_cornersbool

Whether to align corners when using bilinear interpolation.

decoder_norm_paramsdict, optional

Additional arguments for decoder normalization.

aux_heads_in_indextuple of int

Indices of encoder layers to feed into each auxiliary decoder head.

aux_head_num_convsint

Number of conv blocks in each auxiliary head.

aux_head_up_scaleint

Upsample factor for auxiliary heads.

aux_weightslist of float, optional

Weights for auxiliary losses [aux1, aux2, aux3].

loss_fnnn.Module, optional

Loss function module (defaults to CrossEntropy).

optimizer_typetype, optional

Optimizer class (e.g., torch.optim.Adam).

optimizer_paramsdict, optional

Parameters to pass to the optimizer.

train_metricsdict, optional

Dictionary of training metrics.

val_metricsdict, optional

Dictionary of validation metrics.

test_metricsdict, optional

Dictionary of test metrics.

learning_ratefloat

Learning rate for training.

loss_weightslist of float, optional

Class-wise weights for the loss function.

head_lr_factorfloat

Learning rate multiplier for decoder heads.

lr_schedulertype, optional

Learning rate scheduler class to be instantiated. By default, it is set to None, which means no scheduler will be used. Should be a subclass of torch.optim.lr_scheduler.LRScheduler (e.g., torch.optim.lr_scheduler.StepLR).

lr_scheduler_kwargsdict, optional

Additional kwargs passed to the scheduler constructor.

use_sliding_inferencebool

Whether to use sliding window inference for large images on validation and test.

sliding_window_stridetuple of int

Stride for sliding window inference (height, width).

_compute_metrics(y_hat, y, step_name)[source]

Calculate the metrics for the given step.

Parameters

y_hattorch.Tensor

The output data from the forward pass.

ytorch.Tensor

The input data/label.

step_namestr

Name of the step. It will be used to get the metrics from the self.metrics attribute.

Returns

Dict[str, torch.Tensor]

A dictionary with the metrics values.

Parameters:

  • y_hat (torch.Tensor)

  • y (torch.Tensor)

  • step_name (str)

Return type:

Dict[str, torch.Tensor]

_eval_step_with_slide(batch, step_name)[source]
Parameters:

step_name (str)

_loss_func(y_hat, y)[source]

Calculate the loss between the output and the input data.

Parameters

y_hattorch.Tensor

The output data from the forward pass.

ytorch.Tensor

The input data/label.

Returns

torch.Tensor

The loss value.

Parameters:

  • y_hat (Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]])

  • y (torch.Tensor)

Return type:

torch.Tensor

_slide_inference(image, crop_size=(512, 512), stride=(341, 341), ori_shape=None)[source]

Realiza inferência por janelamento (sliding window) com reconstrução e resize final opcional.

Parameters:

  • image (numpy.ndarray)

  • ori_shape (Optional[Tuple[int, int]])

aux_head1
aux_head2
aux_head3
aux_weights = None
configure_optimizers()[source]

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.

Return:

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 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.

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.

static create_from_dict(config)[source]
Parameters:

config (Dict)

Return type:

SETR_PUP

decoder_num_classes = 6
encoder_out_indices = (9, 14, 19, 23)
forward(x)[source]

Perform a forward pass with the input data on the backbone model.

Parameters

xtorch.Tensor

The input data.

Returns

torch.Tensor

The output data from the forward pass.

Parameters:

x (torch.Tensor)

head_lr_factor = 1.0
img_size = (512, 512)
load_backbone(path, freeze=False)[source]

Loads pretrained ViT backbone optionally freezing its weights.

Parameters:

  • path (str)

  • freeze (bool)

num_classes = 6
predict_step(batch, batch_idx, dataloader_idx=None)[source]

Step function called during predict(). By default, it calls forward(). Override to add any processing logic.

The predict_step() is used to scale inference on multi-devices.

To prevent an OOM error, it is possible to use BasePredictionWriter callback to write the predictions to disk or database after each batch or on epoch end.

The BasePredictionWriter should be used while using a spawn based accelerator. This happens for Trainer(strategy="ddp_spawn") or training on 8 TPU cores with Trainer(accelerator="tpu", devices=8) as predictions won’t be returned.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

Predicted output (optional).

Example

class MyModel(LightningModule):

    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        return self(batch)

dm = ...
model = MyModel()
trainer = Trainer(accelerator="gpu", devices=2)
predictions = trainer.predict(model, dm)
Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

  • dataloader_idx (Optional[int])

test_step(batch, batch_idx)[source]

Operates on a single batch of data from the test set. In this step you’d normally generate examples or calculate anything of interest such as accuracy.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

  • Tensor - The loss tensor

  • dict - A dictionary. Can include any keys, but must include the key 'loss'.

  • None - Skip to the next batch.

# if you have one test dataloader:
def test_step(self, batch, batch_idx): ...


# if you have multiple test dataloaders:
def test_step(self, batch, batch_idx, dataloader_idx=0): ...

Examples:

# CASE 1: A single test dataset
def test_step(self, batch, batch_idx):
    x, y = batch

    # implement your own
    out = self(x)
    loss = self.loss(out, y)

    # log 6 example images
    # or generated text... or whatever
    sample_imgs = x[:6]
    grid = torchvision.utils.make_grid(sample_imgs)
    self.logger.experiment.add_image('example_images', grid, 0)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs!
    self.log_dict({'test_loss': loss, 'test_acc': test_acc})

If you pass in multiple test dataloaders, test_step() will have an additional argument. We recommend setting the default value of 0 so that you can quickly switch between single and multiple dataloaders.

# CASE 2: multiple test dataloaders
def test_step(self, batch, batch_idx, dataloader_idx=0):
    # dataloader_idx tells you which dataset this is.
    x, y = batch

    # implement your own
    out = self(x)

    if dataloader_idx == 0:
        loss = self.loss0(out, y)
    else:
        loss = self.loss1(out, y)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs separately for each dataloader
    self.log_dict({f"test_loss_{dataloader_idx}": loss, f"test_acc_{dataloader_idx}": acc})
Note:

If you don’t need to test you don’t need to implement this method.

Note:

When the test_step() is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of the test epoch, the model goes back to training mode and gradients are enabled.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

training_step(batch, batch_idx)[source]

Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

  • Tensor - The loss tensor

  • dict - A dictionary which can include any keys, but must include the key 'loss' in the case of automatic optimization.

  • None - In automatic optimization, this will skip to the next batch (but is not supported for multi-GPU, TPU, or DeepSpeed). For manual optimization, this has no special meaning, as returning the loss is not required.

In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.

Example:

def training_step(self, batch, batch_idx):
    x, y, z = batch
    out = self.encoder(x)
    loss = self.loss(out, x)
    return loss

To use multiple optimizers, you can switch to ‘manual optimization’ and control their stepping:

def __init__(self):
    super().__init__()
    self.automatic_optimization = False


# Multiple optimizers (e.g.: GANs)
def training_step(self, batch, batch_idx):
    opt1, opt2 = self.optimizers()

    # do training_step with encoder
    ...
    opt1.step()
    # do training_step with decoder
    ...
    opt2.step()
Note:

When accumulate_grad_batches > 1, the loss returned here will be automatically normalized by accumulate_grad_batches internally.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

use_sliding_inference = True
validation_step(batch, batch_idx)[source]

Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

  • Tensor - The loss tensor

  • dict - A dictionary. Can include any keys, but must include the key 'loss'.

  • None - Skip to the next batch.

# if you have one val dataloader:
def validation_step(self, batch, batch_idx): ...


# if you have multiple val dataloaders:
def validation_step(self, batch, batch_idx, dataloader_idx=0): ...

Examples:

# CASE 1: A single validation dataset
def validation_step(self, batch, batch_idx):
    x, y = batch

    # implement your own
    out = self(x)
    loss = self.loss(out, y)

    # log 6 example images
    # or generated text... or whatever
    sample_imgs = x[:6]
    grid = torchvision.utils.make_grid(sample_imgs)
    self.logger.experiment.add_image('example_images', grid, 0)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs!
    self.log_dict({'val_loss': loss, 'val_acc': val_acc})

If you pass in multiple val dataloaders, validation_step() will have an additional argument. We recommend setting the default value of 0 so that you can quickly switch between single and multiple dataloaders.

# CASE 2: multiple validation dataloaders
def validation_step(self, batch, batch_idx, dataloader_idx=0):
    # dataloader_idx tells you which dataset this is.
    x, y = batch

    # implement your own
    out = self(x)

    if dataloader_idx == 0:
        loss = self.loss0(out, y)
    else:
        loss = self.loss1(out, y)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs separately for each dataloader
    self.log_dict({f"val_loss_{dataloader_idx}": loss, f"val_acc_{dataloader_idx}": acc})
Note:

If you don’t need to validate you don’t need to implement this method.

Note:

When the validation_step() is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

Parameters:

  • original_resolution (Optional[tuple])

  • img_size (tuple)

  • patch_size (int)

  • in_channels (int)

  • embed_dims (int)

  • num_layers (int)

  • num_heads (int)

  • out_indices (Optional[tuple])

  • encoder_stride (Optional[int])

  • patch_norm (bool)

  • dilatation (int)

  • bias (bool)

  • padding_type (str)

  • mlp_ratio (int)

  • attn_drop_rate (float)

  • drop_path_rate (float)

  • num_fcs (int)

  • qkv_bias (bool)

  • output_cls_token (bool)

  • act_type (type)

  • with_cp (bool)

  • encoder_dropout (float)

  • encoder_norm_type (type)

  • dropout_type (type)

  • cls_token (bool)

  • interpolate_mode (str)

  • act_params (Optional[dict])

  • dropout_params (Optional[dict])

  • encoder_norm_params (Optional[dict])

  • decoder_channels (int)

  • decoder_in_index (int)

  • num_classes (int)

  • decoder_dropout (float)

  • decoder_norm_type (type)

  • decoder_num_convs (int)

  • decoder_up_scale (int)

  • decoder_kernel_size (int)

  • decoder_align_corners (bool)

  • decoder_norm_params (Optional[dict])

  • aux_heads_in_index (tuple[int, int, int])

  • aux_head_num_convs (int)

  • aux_head_up_scale (int)

  • aux_weights (Optional[list[float]])

  • loss_fn (Optional[torch.nn.Module])

  • optimizer (type)

  • optimizer_kwargs (Optional[Dict])

  • train_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • val_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • test_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • freeze_backbone (bool)

  • learning_rate (float)

  • loss_weights (Optional[list[float]])

  • lr_scheduler (Optional[type])

  • lr_scheduler_kwargs (Optional[Dict[str, Any]])

  • head_lr_factor (float)

  • use_sliding_inference (bool)

  • sliding_window_stride (Tuple[int, int])

class minerva.models.nets.SimpleSupervisedModel(backbone, fc, loss_fn, adapter=None, learning_rate=0.001, flatten=True, train_metrics=None, val_metrics=None, test_metrics=None, freeze_backbone=False, optimizer=torch.optim.Adam, optimizer_kwargs=None, lr_scheduler=None, lr_scheduler_kwargs=None)[source]

Bases: lightning.LightningModule

A modular Lightning model wrapper for supervised learning tasks.

This class enables the construction of supervised models by combining a backbone (feature extractor), an optional adapter, and a fully connected (FC) head. It provides a clean interface for setting up custom training, validation, and testing pipelines with pluggable loss functions, metrics, optimizers, and learning rate schedulers.

The architecture is structured as follows:

Backbone Model

v

Adapter (Optional)

(Flatten if needed)

v

Fully Connected Head

v

Loss Function

Training and validation steps comprise the following steps:

  1. Forward pass input through the backbone.

  2. Pass through adapter (if provided).

  3. Flatten the output (if flatten is True) before the FC head.

  4. Forward through the FC head.

  5. Compute loss with respect to targets.

  6. Backpropagate and update parameters.

  7. Compute metrics and log them.

  8. Return loss. train_loss, val_loss, and test_loss are always logged, along with any additional metrics specified in the train_metrics, val_metrics, and test_metrics dictionaries.

This wrapper is especially useful to quickly set up supervised models for various tasks, such as image classification, object detection, and segmentation. It is designed to be flexible and extensible, allowing users to easily swap out components like the backbone, adapter, and FC head as needed. The model is built with a focus on simplicity and modularity, making it easy to adapt to different use cases and requirements. The model is designed to be used with PyTorch Lightning and is compatible with its training loop.

Note: For more complex architectures that does not follow the above structure should not inherit from this class.

Note: Input batches must be tuples (input_tensor, target_tensor).

Initializes the supervised model with training components and configs.

Parameters

backbonetorch.nn.Module or LoadableModule

The backbone (feature extractor) model.

fctorch.nn.Module or LoadableModule

The fully connected head. Use nn.Identity() if not required.

loss_fntorch.nn.Module

Loss function to optimize during training.

adapterCallable, optional

Function to transform backbone outputs before feeding into fc.

learning_ratefloat, default=1e-3

Learning rate used for optimization.

flattenbool, default=True

If True, flattens backbone outputs before fc.

train_metricsdict, optional

TorchMetrics dictionary for training evaluation.

val_metricsdict, optional

TorchMetrics dictionary for validation evaluation.

test_metricsdict, optional

TorchMetrics dictionary for test evaluation.

freeze_backbonebool, default=False

If True, backbone parameters are frozen during training.

optimizer: type

Optimizer class to be instantiated. By default, it is set to torch.optim.Adam. Should be a subclass of torch.optim.Optimizer (e.g., torch.optim.SGD).

optimizer_kwargsdict, optional

Additional kwargs passed to the optimizer constructor.

lr_schedulertype, optional

Learning rate scheduler class to be instantiated. By default, it is set to None, which means no scheduler will be used. Should be a subclass of torch.optim.lr_scheduler.LRScheduler (e.g., torch.optim.lr_scheduler.StepLR).

lr_scheduler_kwargsdict, optional

Additional kwargs passed to the scheduler constructor.

_compute_metrics(y_hat, y, step_name)[source]

Calculate the metrics for the given step.

Parameters

y_hattorch.Tensor

The output data from the forward pass.

ytorch.Tensor

The input data/label.

step_namestr

Name of the step. It will be used to get the metrics from the self.metrics attribute.

Returns

Dict[str, torch.Tensor]

A dictionary with the metrics values.

Parameters:

  • y_hat (torch.Tensor)

  • y (torch.Tensor)

  • step_name (str)

Return type:

Dict[str, torch.Tensor]

_loss_func(y_hat, y)[source]

Calculate the loss between the output and the input data.

Parameters

y_hattorch.Tensor

The output data from the forward pass.

ytorch.Tensor

The input data/label.

Returns

torch.Tensor

The loss value.

Parameters:

  • y_hat (torch.Tensor)

  • y (torch.Tensor)

Return type:

torch.Tensor

_set_trainable_params()[source]

Freeze the parameters of the backbone model if freeze_backbone is set to True.

_single_step(batch, batch_idx, step_name)[source]

Perform a single train/validation/test step. It consists in making a forward pass with the input data on the backbone model, computing the loss between the output and the input data, and logging the loss.

Parameters

batchtorch.Tensor

The input data. It must be a 2-element tuple of tensors, where the first tensor is the input data and the second tensor is the mask.

batch_idxint

The index of the batch.

step_namestr

The name of the step. It will be used to log the loss. The possible values are: “train”, “val” and “test”. The loss will be logged as “{step_name}_loss”.

Returns

torch.Tensor

A tensor with the loss value.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

  • step_name (str)

Return type:

torch.Tensor

adapter = None
backbone
configure_optimizers()[source]

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.

Return:

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 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.

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.

fc
flatten = True
forward(x)[source]

Perform a forward pass with the input data on the backbone model.

Parameters

xtorch.Tensor

The input data.

Returns

torch.Tensor

The output data from the forward pass.

Parameters:

x (torch.Tensor)

Return type:

torch.Tensor

freeze_backbone = False
learning_rate = 0.001
loss_fn
lr_scheduler = None
lr_scheduler_kwargs
metrics
optimizer
optimizer_kwargs
predict_step(batch, batch_idx, dataloader_idx=None)[source]

Step function called during predict(). By default, it calls forward(). Override to add any processing logic.

The predict_step() is used to scale inference on multi-devices.

To prevent an OOM error, it is possible to use BasePredictionWriter callback to write the predictions to disk or database after each batch or on epoch end.

The BasePredictionWriter should be used while using a spawn based accelerator. This happens for Trainer(strategy="ddp_spawn") or training on 8 TPU cores with Trainer(accelerator="tpu", devices=8) as predictions won’t be returned.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

Predicted output (optional).

Example

class MyModel(LightningModule):

    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        return self(batch)

dm = ...
model = MyModel()
trainer = Trainer(accelerator="gpu", devices=2)
predictions = trainer.predict(model, dm)
test_step(batch, batch_idx)[source]

Operates on a single batch of data from the test set. In this step you’d normally generate examples or calculate anything of interest such as accuracy.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

  • Tensor - The loss tensor

  • dict - A dictionary. Can include any keys, but must include the key 'loss'.

  • None - Skip to the next batch.

# if you have one test dataloader:
def test_step(self, batch, batch_idx): ...


# if you have multiple test dataloaders:
def test_step(self, batch, batch_idx, dataloader_idx=0): ...

Examples:

# CASE 1: A single test dataset
def test_step(self, batch, batch_idx):
    x, y = batch

    # implement your own
    out = self(x)
    loss = self.loss(out, y)

    # log 6 example images
    # or generated text... or whatever
    sample_imgs = x[:6]
    grid = torchvision.utils.make_grid(sample_imgs)
    self.logger.experiment.add_image('example_images', grid, 0)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs!
    self.log_dict({'test_loss': loss, 'test_acc': test_acc})

If you pass in multiple test dataloaders, test_step() will have an additional argument. We recommend setting the default value of 0 so that you can quickly switch between single and multiple dataloaders.

# CASE 2: multiple test dataloaders
def test_step(self, batch, batch_idx, dataloader_idx=0):
    # dataloader_idx tells you which dataset this is.
    x, y = batch

    # implement your own
    out = self(x)

    if dataloader_idx == 0:
        loss = self.loss0(out, y)
    else:
        loss = self.loss1(out, y)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs separately for each dataloader
    self.log_dict({f"test_loss_{dataloader_idx}": loss, f"test_acc_{dataloader_idx}": acc})
Note:

If you don’t need to test you don’t need to implement this method.

Note:

When the test_step() is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of the test epoch, the model goes back to training mode and gradients are enabled.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

training_step(batch, batch_idx)[source]

Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

  • Tensor - The loss tensor

  • dict - A dictionary which can include any keys, but must include the key 'loss' in the case of automatic optimization.

  • None - In automatic optimization, this will skip to the next batch (but is not supported for multi-GPU, TPU, or DeepSpeed). For manual optimization, this has no special meaning, as returning the loss is not required.

In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.

Example:

def training_step(self, batch, batch_idx):
    x, y, z = batch
    out = self.encoder(x)
    loss = self.loss(out, x)
    return loss

To use multiple optimizers, you can switch to ‘manual optimization’ and control their stepping:

def __init__(self):
    super().__init__()
    self.automatic_optimization = False


# Multiple optimizers (e.g.: GANs)
def training_step(self, batch, batch_idx):
    opt1, opt2 = self.optimizers()

    # do training_step with encoder
    ...
    opt1.step()
    # do training_step with decoder
    ...
    opt2.step()
Note:

When accumulate_grad_batches > 1, the loss returned here will be automatically normalized by accumulate_grad_batches internally.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

validation_step(batch, batch_idx)[source]

Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

  • Tensor - The loss tensor

  • dict - A dictionary. Can include any keys, but must include the key 'loss'.

  • None - Skip to the next batch.

# if you have one val dataloader:
def validation_step(self, batch, batch_idx): ...


# if you have multiple val dataloaders:
def validation_step(self, batch, batch_idx, dataloader_idx=0): ...

Examples:

# CASE 1: A single validation dataset
def validation_step(self, batch, batch_idx):
    x, y = batch

    # implement your own
    out = self(x)
    loss = self.loss(out, y)

    # log 6 example images
    # or generated text... or whatever
    sample_imgs = x[:6]
    grid = torchvision.utils.make_grid(sample_imgs)
    self.logger.experiment.add_image('example_images', grid, 0)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs!
    self.log_dict({'val_loss': loss, 'val_acc': val_acc})

If you pass in multiple val dataloaders, validation_step() will have an additional argument. We recommend setting the default value of 0 so that you can quickly switch between single and multiple dataloaders.

# CASE 2: multiple validation dataloaders
def validation_step(self, batch, batch_idx, dataloader_idx=0):
    # dataloader_idx tells you which dataset this is.
    x, y = batch

    # implement your own
    out = self(x)

    if dataloader_idx == 0:
        loss = self.loss0(out, y)
    else:
        loss = self.loss1(out, y)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs separately for each dataloader
    self.log_dict({f"val_loss_{dataloader_idx}": loss, f"val_acc_{dataloader_idx}": acc})
Note:

If you don’t need to validate you don’t need to implement this method.

Note:

When the validation_step() is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

Parameters:

  • backbone (Union[torch.nn.Module, minerva.models.loaders.LoadableModule])

  • fc (Union[torch.nn.Module, minerva.models.loaders.LoadableModule])

  • loss_fn (torch.nn.Module)

  • adapter (Optional[Callable[[torch.Tensor], torch.Tensor]])

  • learning_rate (float)

  • flatten (bool)

  • train_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • val_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • test_metrics (Optional[Dict[str, torchmetrics.Metric]])

  • freeze_backbone (bool)

  • optimizer (type)

  • optimizer_kwargs (Optional[Dict[str, Any]])

  • lr_scheduler (Optional[type])

  • lr_scheduler_kwargs (Optional[Dict[str, Any]])

class minerva.models.nets.UNet(n_channels=1, bilinear=False, learning_rate=0.001, loss_fn=None, **kwargs)[source]

Bases: minerva.models.nets.base.SimpleSupervisedModel

This class is a simple implementation of the U-Net model, which is a convolutional neural network used for image segmentation. The model consists of a contracting path (encoder) and an expansive path (decoder). The contracting path follows the typical architecture of a convolutional neural network, with repeated applications of convolutions and max pooling layers. The expansive path consists of up-convolutions and concatenation of feature maps from the contracting path. The model also has skip connections, which allows the expansive path to use information from the contracting path at multiple resolutions. The U-Net model was originally proposed by Ronneberger, Fischer, and Brox in 2015.

This architecture, handles arbitrary input sizes, and returns an output of the same size as the input. The expected input size is (N, C, H, W), where N is the batch size, C is the number of channels, H is the height of the input image, and W is the width of the input image.

Note that, for this implementation, the input batch is a single tensor and not a tuple of tensors (e.g., data and label).

Note that this class wrappers the _UNet class, which is the actual implementation of the U-Net model, into a SimpleReconstructionNet class, which is a simple autoencoder pipeline for reconstruction tasks.

References

Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. “U-net: Convolutional networks for biomedical image segmentation.” Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015.

Wrapper implementation of the U-Net model.

Parameters

n_channelsint, optional

The number of channels of the input, by default 1

bilinearbool, optional

If True use bilinear interpolation for upsampling, by default False.

learning_ratefloat, optional

The learning rate to Adam optimizer, by default 1e-3

loss_fntorch.nn.Module, optional

The function used to compute the loss. If None, it will be used the MSELoss, by default None.

kwargsDict

Additional arguments to be passed to the SimpleSupervisedModel class.

Parameters:

  • n_channels (int)

  • bilinear (bool)

  • learning_rate (float)

  • loss_fn (Optional[torch.nn.Module])

class minerva.models.nets.WiseNet(in_channels=1, out_channels=1, loss_fn=None, learning_rate=0.001, **kwargs)[source]

Bases: minerva.models.nets.base.SimpleSupervisedModel

A modular Lightning model wrapper for supervised learning tasks.

This class enables the construction of supervised models by combining a backbone (feature extractor), an optional adapter, and a fully connected (FC) head. It provides a clean interface for setting up custom training, validation, and testing pipelines with pluggable loss functions, metrics, optimizers, and learning rate schedulers.

The architecture is structured as follows:

Backbone Model

v

Adapter (Optional)

(Flatten if needed)

v

Fully Connected Head

v

Loss Function

Training and validation steps comprise the following steps:

  1. Forward pass input through the backbone.

  2. Pass through adapter (if provided).

  3. Flatten the output (if flatten is True) before the FC head.

  4. Forward through the FC head.

  5. Compute loss with respect to targets.

  6. Backpropagate and update parameters.

  7. Compute metrics and log them.

  8. Return loss. train_loss, val_loss, and test_loss are always logged, along with any additional metrics specified in the train_metrics, val_metrics, and test_metrics dictionaries.

This wrapper is especially useful to quickly set up supervised models for various tasks, such as image classification, object detection, and segmentation. It is designed to be flexible and extensible, allowing users to easily swap out components like the backbone, adapter, and FC head as needed. The model is built with a focus on simplicity and modularity, making it easy to adapt to different use cases and requirements. The model is designed to be used with PyTorch Lightning and is compatible with its training loop.

Note: For more complex architectures that does not follow the above structure should not inherit from this class.

Note: Input batches must be tuples (input_tensor, target_tensor).

Initializes the supervised model with training components and configs.

Parameters

backbonetorch.nn.Module or LoadableModule

The backbone (feature extractor) model.

fctorch.nn.Module or LoadableModule

The fully connected head. Use nn.Identity() if not required.

loss_fntorch.nn.Module

Loss function to optimize during training.

adapterCallable, optional

Function to transform backbone outputs before feeding into fc.

learning_ratefloat, default=1e-3

Learning rate used for optimization.

flattenbool, default=True

If True, flattens backbone outputs before fc.

train_metricsdict, optional

TorchMetrics dictionary for training evaluation.

val_metricsdict, optional

TorchMetrics dictionary for validation evaluation.

test_metricsdict, optional

TorchMetrics dictionary for test evaluation.

freeze_backbonebool, default=False

If True, backbone parameters are frozen during training.

optimizer: type

Optimizer class to be instantiated. By default, it is set to torch.optim.Adam. Should be a subclass of torch.optim.Optimizer (e.g., torch.optim.SGD).

optimizer_kwargsdict, optional

Additional kwargs passed to the optimizer constructor.

lr_schedulertype, optional

Learning rate scheduler class to be instantiated. By default, it is set to None, which means no scheduler will be used. Should be a subclass of torch.optim.lr_scheduler.LRScheduler (e.g., torch.optim.lr_scheduler.StepLR).

lr_scheduler_kwargsdict, optional

Additional kwargs passed to the scheduler constructor.

_single_step(batch, batch_idx, step_name)[source]

Perform a single train/validation/test step. It consists in making a forward pass with the input data on the backbone model, computing the loss between the output and the input data, and logging the loss.

Parameters

batchtorch.Tensor

The input data. It must be a 2-element tuple of tensors, where the first tensor is the input data and the second tensor is the mask.

batch_idxint

The index of the batch.

step_namestr

The name of the step. It will be used to log the loss. The possible values are: “train”, “val” and “test”. The loss will be logged as “{step_name}_loss”.

Returns

torch.Tensor

A tensor with the loss value.

Parameters:

  • batch (torch.Tensor)

  • batch_idx (int)

  • step_name (str)

Return type:

torch.Tensor

predict_step(batch, batch_idx, dataloader_idx=None)[source]

Step function called during predict(). By default, it calls forward(). Override to add any processing logic.

The predict_step() is used to scale inference on multi-devices.

To prevent an OOM error, it is possible to use BasePredictionWriter callback to write the predictions to disk or database after each batch or on epoch end.

The BasePredictionWriter should be used while using a spawn based accelerator. This happens for Trainer(strategy="ddp_spawn") or training on 8 TPU cores with Trainer(accelerator="tpu", devices=8) as predictions won’t be returned.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

Predicted output (optional).

Example

class MyModel(LightningModule):

    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        return self(batch)

dm = ...
model = MyModel()
trainer = Trainer(accelerator="gpu", devices=2)
predictions = trainer.predict(model, dm)
Parameters:

  • in_channels (int)

  • out_channels (int)

  • loss_fn (torch.nn.Module)

  • learning_rate (float)