Tutorial 3: Learn about criterions

In this folder “fgvclib/criterions” we provide different loss functions for the fgvclib.

We provide four loss functions, cross_entropy_loss, binary_cross_entropy_loss, mean_square_error_loss and mutual_channel_loss

You can choose which loss function you want to use, and you should set it in the “./configs”. For more details about the configs please see FGVC Configs

Loss functions	Name
cross entropy loss	cross_entropy_loss
binary cross entropy loss	binary_cross_entropy_loss
mean square error loss	mean_square_error_loss
mutual channel loss	mutual_channel_loss

Base loss function

cross_entropy_loss, binary_cross_entropy_loss, mean_square_error_loss are the base loss functions, and they are from PyTorch. “fgvclib/criterions/base_loss.py”: provides the base loss functions.

cross_entropy_loss: Build the cross entropy loss function.

• Args:

  cfg (CfgNode): The root node of config.

• Return:

  nn.Module: The loss function.

binary_cross_entropy_loss:Build the binary cross entropy loss function.

• Args:

  cfg (CfgNode): The root node of config.

• Return:

  nn.Module: The loss function.

mean_square_error_loss: Build the mean square error loss function.

• Args:

  cfg (CfgNode): The root node of config.

• Return:

  nn.Module: The loss function.

Mutual channel loss

“fgvclib/criterions/mutual_channel_loss.py”: provides the mutual channel loss function which was proposed on “The Devil is in the Channels: Mutual-Channel Loss for Fine-Grained Image Classification”.

class MutualChannelLoss: The mutual channel loss function.

• Args:

  height (int): The kernel size of average pooling. cnum (int): Channel numbers per class. div_weight (float): The weight for diversity part loss. dis_weight (float): The weight for discriminality part loss.

Utils

In the “fgvclib/criterions/utils.py”, we design a class named LossItem, and two functions, compute_loss_value and detach_loss_value.

LossItem: A dataclass object for store training loss

• Args:

  name (string): The loss item name. value (torch.Tensor): The value of loss. weight (float, optional): The weight of current loss item, default is 1.0.

compute_loss_value: A dataclass object for store training loss

• Args:

  items (List[LossItem]): The loss items.

• Return:

  Tensor: The total loss value.

detach_loss_value: Detach loss value from GPU.

• Args:

  items (List[LossItem]): The loss items.

• Return:

  Dict: A loss information dict whose key is loss name, value is loss value.

The example of using the criterions

Build loss functions for training.

In the “fgvclib/apis/build.py”, use the “fgvclib.criterions” to build loss functions for training. You can choose the loss function name criterion_cfg['name'] from cross_entropy_loss, cross_entropy_loss, mean_square_error_loss and mutual_channel_loss.

from fgvclib.criterions import get_criterion

def build_criterion(criterion_cfg: CfgNode) -> nn.Module:
    criterion_builder = get_criterion(criterion_cfg['name'])
    criterion = criterion_builder(cfg=tltd(criterion_cfg['args']))
    return criterion

Calculate loss functions.

Following is about how to calculate the loss, and you can replace the loss functions.

from fgvclib.criterions.utils import LossItem

losses = list()
losses.append(LossItem(name='cross_entropy_loss', value=self.criterions['cross_entropy_loss']['fn'](x, targets)))

Define the forward.

Set the ResNet50 for example.

from fgvclib.criterions.utils import LossItem

def forward(self, x, targets=None):
    x = self.infer(x)
    if self.training:
        losses = list()
        osses.append(LossItem(name='cross_entropy_loss', value=self.criterions['cross_entropy_loss']['fn'](x, targets)))
        return x, losses
        
    return x