File size: 6,290 Bytes
d7e58f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
# ------------------------------------------------------------------------------
# Adapted from https://github.com/jiawei-ren/BalancedMSE
# Original licence: Copyright (c) 2022 Jiawei Ren, under the MIT License.
# ------------------------------------------------------------------------------

from typing import Optional, Union

import torch
import torch.distributed as dist
import torch.nn.functional as F
from mmcv.runner import get_dist_info
from torch.nn.modules.loss import _Loss

from .utils import weighted_loss


@weighted_loss
def bmc_loss_md(pred: torch.Tensor, target: torch.Tensor,
                noise_var: torch.Tensor, all_gather: bool,
                loss_mse_weight: float,
                loss_debias_weight: float) -> torch.Tensor:
    """
    Args:
        pred (torch.Tensor): The prediction. Shape should be (N, L).
        target (torch.Tensor): The learning target of the prediction.
        noise_var (torch.Tensor): Noise var of ground truth distribution.
        all_gather (bool): Whether gather tensors across all sub-processes.
            Only used in DDP training scheme.
        loss_mse_weight (float, optional): The weight of the mse term.
        loss_debias_weight (float, optional): The weight of the debiased term.

    Returns:
            torch.Tensor: The calculated loss
    """
    N = pred.shape[0]
    L = pred.shape[1]
    device = pred.device

    loss_mse = F.mse_loss(pred, target, reduction='none').sum(-1)
    loss_mse = loss_mse / noise_var

    if all_gather:
        rank, world_size = get_dist_info()
        bs, length = target.shape
        all_bs = [torch.zeros(1).to(device) for _ in range(world_size)]
        dist.all_gather(all_bs, torch.Tensor([bs]).to(device))
        all_bs_int = [int(v.item()) for v in all_bs]
        max_bs_int = max(all_bs_int)
        target_padding = torch.zeros(max_bs_int, length).to(device)
        target_padding[:bs] = target
        all_tensor = []
        for _ in range(world_size):
            all_tensor.append(torch.zeros(max_bs_int, length).type_as(target))
        dist.all_gather(all_tensor, target_padding)
        # remove padding
        for i in range(world_size):
            all_tensor[i] = all_tensor[i][:all_bs_int[i]]
        target = torch.cat(all_tensor, dim=0)

    # Debias term
    target = target.unsqueeze(0).repeat(N, 1, 1)
    pred = pred.unsqueeze(1).expand_as(target)
    debias_term = F.mse_loss(pred, target, reduction='none').sum(-1)
    debias_term = -0.5 * debias_term / noise_var
    loss_debias = torch.logsumexp(debias_term, dim=1).squeeze(-1)
    loss = loss_mse * loss_mse_weight + loss_debias * loss_debias_weight
    # recover loss scale of mse_loss
    loss = loss / L * noise_var.detach()
    return loss


class BMCLossMD(_Loss):
    """Balanced MSE loss, use batch monte-carlo to estimate distribution.
    https://arxiv.org/abs/2203.16427.

    Args:
        init_noise_sigma (float, optional): The initial value of noise sigma.
            This sigma is used to represent ground truth distribution.
            Defaults to 1.0.
        all_gather (bool, optional): Whether gather tensors across all
            sub-processes. If set True, BMC will have more precise estimation
            with more time cost. Default: False.
        reduction (str, optional): The method that reduces the loss to a
            scalar. Options are "none", "mean" and "sum".
        loss_mse_weight (float, optional): The weight of the mse term.
            Defaults to 1.0.
        loss_debias_weight (float, optional): The weight of the debiased term.
            Defaults to 1.0.
    """
    def __init__(self,
                 init_noise_sigma: Optional[float] = 1.0,
                 all_gather: Optional[bool] = False,
                 reduction: Optional[str] = 'mean',
                 loss_mse_weight: Optional[float] = 1.0,
                 loss_debias_weight: Optional[float] = 1.0):
        super(BMCLossMD, self).__init__()
        self.noise_sigma = torch.nn.Parameter(
            torch.tensor(init_noise_sigma).float())
        self.all_gather = all_gather
        assert reduction in (None, 'none', 'mean', 'sum')
        reduction = 'none' if reduction is None else reduction
        self.reduction = reduction
        self.loss_mse_weight = loss_mse_weight
        self.loss_debias_weight = loss_debias_weight

    def forward(
            self,
            pred: torch.Tensor,
            target: torch.Tensor,
            weight: Optional[Union[torch.Tensor, None]] = None,
            avg_factor: Optional[Union[int, None]] = None,
            reduction_override: Optional[Union[str,
                                               None]] = None) -> torch.Tensor:
        """Forward function of loss.

        Args:
            pred (torch.Tensor): The prediction.
            target (torch.Tensor): The learning target of the prediction.
            weight (torch.Tensor, optional): Weight of the loss for each
                prediction. Defaults to None.
            avg_factor (int, optional): Average factor that is used to average
                the loss. Defaults to None.
            weight (torch.Tensor, optional): Weight of the loss for each
                prediction. Defaults to None.
            reduction_override (str, optional): The reduction method used to
                override the original reduction method of the loss.
                Defaults to None.
        Returns:
            torch.Tensor: The calculated loss
        """
        assert reduction_override in (None, 'none', 'mean', 'sum')
        reduction = (reduction_override
                     if reduction_override else self.reduction)
        noise_var = (self.noise_sigma**2).type_as(pred)
        pred = pred.view(pred.shape[0], -1)
        target = target.view(target.shape[0], -1)
        loss = bmc_loss_md(pred,
                           target,
                           noise_var=noise_var,
                           all_gather=self.all_gather,
                           loss_mse_weight=self.loss_mse_weight,
                           loss_debias_weight=self.loss_debias_weight,
                           weight=weight,
                           reduction=reduction,
                           avg_factor=avg_factor)
        return loss