Losses¶

class classy_vision.losses.BarronLoss(alpha, size_average, c)¶

This implements the Barron loss.

__init__(alpha, size_average, c)¶

Constructor for ClassyLoss.

forward(prediction, target)¶

Compute the loss for the provided sample.

Refer to torch.nn.Module for more details.

classmethod from_config(config: Dict[str, Any]) → classy_vision.losses.barron_loss.BarronLoss¶

Instantiates a BarronLoss from a configuration.

Parameters

config – A configuration for a BarronLoss. See __init__() for parameters expected in the config.

Returns

A BarronLoss instance.

class classy_vision.losses.ClassyLoss¶

Base class to calculate the loss during training.

This implementation of torch.nn.Module allows building the loss object from a configuration file.

__init__()¶

Constructor for ClassyLoss.

forward(output, target)¶

Compute the loss for the provided sample.

Refer to torch.nn.Module for more details.

classmethod from_config(config: Dict[str, Any]) → classy_vision.losses.classy_loss.ClassyLoss¶

Instantiates a ClassyLoss from a configuration.

Parameters

config – A configuration for a ClassyLoss.

Returns

A ClassyLoss instance.

get_classy_state() → Dict[str, Any]¶

Get the state of the ClassyLoss.

The returned state is used for checkpointing. Note that most losses are stateless and do not need to save any state.

Returns

A state dictionary containing the state of the loss.

has_learned_parameters() → bool¶

Does this loss have learned parameters?

set_classy_state(state: Dict[str, Any]) → None¶

Set the state of the ClassyLoss.

Parameters

state_dict – The state dictionary. Must be the output of a call to get_classy_state().

This is used to load the state of the loss from a checkpoint. Note that most losses are stateless and do not need to load any state.

class classy_vision.losses.LabelSmoothingCrossEntropyLoss(ignore_index=- 100, reduction='mean', smoothing_param=None)¶

__init__(ignore_index=- 100, reduction='mean', smoothing_param=None)¶

Intializer for the label smoothed cross entropy loss. This decreases gap between output scores and encourages generalization. Targets provided to forward can be one-hot vectors (NxC) or class indices (Nx1).

This normalizes the targets to a sum of 1 based on the total count of positive targets for a given sample before applying label smoothing.

Parameters

• ignore_index – sample should be ignored for loss if the class is this value

• reduction – specifies reduction to apply to the output

• smoothing_param – value to be added to each target entry

compute_valid_targets(target, classes)¶

This function takes one-hot or index target vectors and computes valid one-hot target vectors, based on ignore index value

forward(output, target)¶

Compute the loss for the provided sample.

Refer to torch.nn.Module for more details.

classmethod from_config(config: Dict[str, Any]) → classy_vision.losses.label_smoothing_loss.LabelSmoothingCrossEntropyLoss¶

Instantiates a LabelSmoothingCrossEntropyLoss from a configuration.

Parameters

config – A configuration for a LabelSmoothingCrossEntropyLoss. See __init__() for parameters expected in the config.

Returns

A LabelSmoothingCrossEntropyLoss instance.

smooth_targets(valid_targets, classes)¶

This function takes valid (No ignore values present) one-hot target vectors and computes smoothed target vectors (normalized) according to the loss’s smoothing parameter

class classy_vision.losses.MultiOutputSumLoss(loss)¶

Applies the provided loss to the list of outputs (or single output) and sums up the losses.

__init__(loss) → None¶

Constructor for ClassyLoss.

forward(output, target)¶

Compute the loss for the provided sample.

Refer to torch.nn.Module for more details.

classmethod from_config(config: Dict[str, Any]) → classy_vision.losses.multi_output_sum_loss.MultiOutputSumLoss¶

Instantiates a MultiOutputSumLoss from a configuration.

Parameters

config – A configuration for a MultiOutpuSumLoss. See __init__() for parameters expected in the config.

Returns

A MultiOutputSumLoss instance.

class classy_vision.losses.SoftTargetCrossEntropyLoss(ignore_index=- 100, reduction='mean', normalize_targets=True)¶

__init__(ignore_index=- 100, reduction='mean', normalize_targets=True)¶

Intializer for the soft target cross-entropy loss loss. This allows the targets for the cross entropy loss to be multilabel

Parameters

• ignore_index – sample should be ignored for loss if the class is this value

• reduction – specifies reduction to apply to the output

• normalize_targets – whether the targets should be normalized to a sum of 1 based on the total count of positive targets for a given sample

forward(output, target)¶

for N examples and C classes - output: N x C these are raw outputs (without softmax/sigmoid) - target: N x C or N corresponding targets

Target elements set to ignore_index contribute 0 loss.

Samples where all entries are ignore_index do not contribute to the loss reduction.

classmethod from_config(config: Dict[str, Any]) → classy_vision.losses.soft_target_cross_entropy_loss.SoftTargetCrossEntropyLoss¶

Instantiates a SoftTargetCrossEntropyLoss from a configuration.

Parameters

config – A configuration for a SoftTargetCrossEntropyLoss. See __init__() for parameters expected in the config.

Returns

A SoftTargetCrossEntropyLoss instance.

class classy_vision.losses.SumArbitraryLoss(losses: List[classy_vision.losses.classy_loss.ClassyLoss], weights: Optional[torch.Tensor] = None)¶

Sums a collection of (weighted) torch.nn losses.

NOTE: this applies all the losses to the same output and does not support taking a list of outputs as input.

__init__(losses: List[classy_vision.losses.classy_loss.ClassyLoss], weights: Optional[torch.Tensor] = None) → None¶

Constructor for ClassyLoss.

forward(prediction, target)¶

Compute the loss for the provided sample.

Refer to torch.nn.Module for more details.

classmethod from_config(config: Dict[str, Any]) → classy_vision.losses.sum_arbitrary_loss.SumArbitraryLoss¶

Instantiates a SumArbitraryLoss from a configuration.

Parameters

config – A configuration for a SumArbitraryLoss. See __init__() for parameters expected in the config.

Returns

A SumArbitraryLoss instance.

classy_vision.losses.build_loss(config)¶

Builds a ClassyLoss from a config.

This assumes a ‘name’ key in the config which is used to determine what model class to instantiate. For instance, a config {“name”: “my_loss”, “foo”: “bar”} will find a class that was registered as “my_loss” (see register_loss()) and call .from_config on it.

In addition to losses registered with register_loss(), we also support instantiating losses available in the torch.nn.modules.loss module. Any keys in the config will get expanded to parameters of the loss constructor. For instance, the following call will instantiate a torch.nn.modules.CrossEntropyLoss:

build_loss({"name": "CrossEntropyLoss", "reduction": "sum"})

classy_vision.losses.register_loss(name, bypass_checks=False)¶

Registers a ClassyLoss subclass.

This decorator allows Classy Vision to instantiate a subclass of ClassyLoss from a configuration file, even if the class itself is not part of the Classy Vision framework. To use it, apply this decorator to a ClassyLoss subclass, like this:

@register_loss("my_loss")
class MyLoss(ClassyLoss):
     ...

To instantiate a loss from a configuration file, see build_loss().