losses_soft.py

"""
Implements the knowledge distillation loss
"""
from abc import get_cache_token
import torch
from torch.nn import functional as F
from torch.nn.modules.loss import MSELoss, BCEWithLogitsLoss, CrossEntropyLoss
from utils import batch_index_select
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
import math

class DistillationLoss(torch.nn.Module):
    """
    This module wraps a standard criterion and adds an extra knowledge distillation loss by
    taking a teacher model prediction and using it as additional supervision.
    """
    def __init__(self, base_criterion: torch.nn.Module, teacher_model: torch.nn.Module,
                 distillation_type: str, alpha: float, tau: float):
        super().__init__()
        self.base_criterion = base_criterion
        self.teacher_model = teacher_model
        assert distillation_type in ['none', 'soft', 'hard']
        self.distillation_type = distillation_type
        self.alpha = alpha
        self.tau = tau

    def forward(self, inputs, outputs, labels):
        """
        Args:
            inputs: The original inputs that are feed to the teacher model
            outputs: the outputs of the model to be trained. It is expected to be
                either a Tensor, or a Tuple[Tensor, Tensor], with the original output
                in the first position and the distillation predictions as the second output
            labels: the labels for the base criterion
        """
        outputs_kd = None
        if not isinstance(outputs, torch.Tensor):
            # assume that the model outputs a tuple of [outputs, outputs_kd]
            outputs, outputs_kd = outputs
        base_loss = self.base_criterion(outputs, labels)
        if self.distillation_type == 'none':
            return base_loss

        if outputs_kd is None:
            raise ValueError("When knowledge distillation is enabled, the model is "
                             "expected to return a Tuple[Tensor, Tensor] with the output of the "
                             "class_token and the dist_token")
        # don't backprop throught the teacher
        with torch.no_grad():
            teacher_outputs = self.teacher_model(inputs)

        if self.distillation_type == 'soft':
            T = self.tau
            # taken from https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
            # with slight modifications
            distillation_loss = F.kl_div(
                F.log_softmax(outputs_kd / T, dim=1),
                F.log_softmax(teacher_outputs / T, dim=1),
                reduction='sum',
                log_target=True
            ) * (T * T) / outputs_kd.numel()
        elif self.distillation_type == 'hard':
            distillation_loss = F.cross_entropy(outputs_kd, teacher_outputs.argmax(dim=1))

        loss = base_loss * (1 - self.alpha) + distillation_loss * self.alpha
        return loss


class DiffPruningLoss(torch.nn.Module):
    """
    This module wraps a standard criterion and adds an extra knowledge distillation loss by
    taking a teacher model prediction and using it as additional supervision.
    """
    def __init__(self, base_criterion: torch.nn.Module, dynamic=False, ratio_weight=2.0, pruning_loc=[3,6,9], keep_ratio=[0.75, 0.5, 0.25], clf_weight=0, print_mode=True):
        super().__init__()
        self.base_criterion = base_criterion
        self.clf_weight = clf_weight
        self.pruning_loc = pruning_loc
        self.keep_ratio = keep_ratio
        self.count = 0
        self.print_mode = print_mode
        self.cls_loss = 0
        self.ratio_loss = 0

        self.ratio_weight = ratio_weight

        self.dynamic = dynamic

        if self.dynamic:
            print('using dynamic loss')

    def forward(self, inputs, outputs, labels):
        """
        Args:
            inputs: The original inputs that are feed to the teacher model
            outputs: the outputs of the model to be trained. It is expected to be
                either a Tensor, or a Tuple[Tensor, Tensor], with the original output
                in the first position and the distillation predictions as the second output
            labels: the labels for the base criterion
        """

        pred, out_pred_score = outputs

        pred_loss = 0.0
        # ratio = [1.0,] + self.keep_ratio
        # for i, score in enumerate(out_pred_score):
        #     score = score.mean(1)
        #     now_ratio = ratio[i+1] / ratio[i]
        #     pred_loss = pred_loss + ((score - now_ratio) ** 2).mean()

        ratio = self.keep_ratio
        for i, score in enumerate(out_pred_score):
            pos_ratio = score.mean(1)
            pred_loss = pred_loss + ((pos_ratio - ratio[i]) ** 2).mean()

        cls_loss = self.base_criterion(pred, labels)
        # print(cls_loss, pred_loss)
        loss = self.clf_weight * cls_loss + self.ratio_weight * pred_loss / len(self.pruning_loc)

        if self.print_mode:
            self.cls_loss += cls_loss.item()
            self.ratio_loss += pred_loss.item()
            self.count += 1
            if self.count == 100:
                print('loss info: cls_loss=%.4f, ratio_loss=%.4f' % (self.cls_loss / 100, self.ratio_loss / 100))
                self.count = 0
                self.cls_loss = 0
                self.ratio_loss = 0
        return loss


class DistillDiffPruningLoss(torch.nn.Module):
    """
    This module wraps a standard criterion and adds an extra knowledge distillation loss by
    taking a teacher model prediction and using it as additional supervision.
    """
    def __init__(self, teacher_model, base_criterion: torch.nn.Module, ratio_weight=2.0, distill_weight=0.5, dynamic=False, pruning_loc=[3,6,9], keep_ratio=[0.75, 0.5, 0.25], clf_weight=0, mse_token=False, print_mode=True):
        super().__init__()
        self.teacher_model = teacher_model
        self.base_criterion = base_criterion
        self.clf_weight = clf_weight
        self.pruning_loc = pruning_loc
        self.keep_ratio = keep_ratio
        self.count = 0
        self.print_mode = print_mode
        self.cls_loss = 0
        self.ratio_loss = 0
        self.cls_distill_loss = 0
        self.token_distill_loss = 0
        self.mse_token = mse_token
        self.dynamic = dynamic

        self.ratio_weight = ratio_weight
        self.distill_weight = distill_weight

        print('ratio_weight=', ratio_weight, 'distill_weight', distill_weight)


        if dynamic:
            print('using dynamic loss')

    def forward(self, inputs, outputs, labels):
        """
        Args:
            inputs: The original inputs that are feed to the teacher model
            outputs: the outputs of the model to be trained. It is expected to be
                either a Tensor, or a Tuple[Tensor, Tensor], with the original output
                in the first position and the distillation predictions as the second output
            labels: the labels for the base criterion
        """

        pred, token_pred, mask, out_pred_score = outputs


        pred_loss = 0.0
        for i, score in enumerate(out_pred_score):
            pred_loss += score.abs().mean() * self.keep_ratio[i]

        cls_loss = self.base_criterion(pred, labels)

        with torch.no_grad():
            cls_t, token_t = self.teacher_model(inputs)


        cls_kl_loss = F.kl_div(
                F.log_softmax(pred, dim=-1),
                F.log_softmax(cls_t, dim=-1),
                reduction='batchmean',
                log_target=True
            )

        B, N, C = token_pred.size()
        assert mask.numel() == B * N

        # print(mask)

        # bool_mask = mask.reshape(B*N) > 0.5
        # print('====================')
        # print(mask.size())
        bool_mask = mask.repeat(1,1,C).reshape(B*N, C)
        # print(bool_mask.size())
        # print('------')


        token_pred = token_pred.reshape(B*N, C)
        token_t = token_t.reshape(B*N, C)
        # print(token_t.size())
        # print(token_pred.size()) 维度都是对对
        # print('====================')

        if mask.sum() < 0.1:
            token_kl_loss = token_pred.new(1,).fill_(0.0)
        else:
            token_t = token_t
            token_pred = token_pred
            if self.mse_token:
                kl_tm = torch.pow(token_pred - token_t, 2)*bool_mask
                token_kl_loss = kl_tm.mean()
            else:
                token_kl_loss = F.kl_div(
                        F.log_softmax(token_pred, dim=-1),
                        F.log_softmax(token_t, dim=-1),
                        reduction='batchmean',
                        log_target=True
                    )

        loss = self.clf_weight * cls_loss + self.ratio_weight * pred_loss/len(self.pruning_loc) + self.distill_weight * cls_kl_loss + self.distill_weight * token_kl_loss
        # if self.count % 20:
        #     print('loss info: cls_loss=%.4f, ratio_loss=%.4f, cls_kl=%.4f, token_kl=%.4f' % (cls_loss, pred_loss , cls_kl_loss, token_kl_loss))
        if self.print_mode:
            self.cls_loss += cls_loss.item()
            self.ratio_loss += pred_loss.item()
            self.cls_distill_loss += cls_kl_loss.item()
            self.token_distill_loss += token_kl_loss.item()
            self.count += 1
            if self.count == 100:
                print('loss info: cls_loss=%.4f, ratio_loss=%.4f, cls_kl=%.4f, token_kl=%.4f' % (self.cls_loss / 100, self.ratio_loss / 100, self.cls_distill_loss/ 100, self.token_distill_loss/ 100))
                self.count = 0
                self.cls_loss = 0
                self.ratio_loss = 0
                self.cls_distill_loss = 0
                self.token_distill_loss = 0
        return loss