File size: 2,385 Bytes
00fc29f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import List, Optional, Union

import torch
import torch.nn as nn
from omegaconf import ListConfig

from ...util import append_dims, instantiate_from_config
from ...modules.autoencoding.lpips.loss.lpips import LPIPS


class StandardDiffusionLoss(nn.Module):
    def __init__(

        self,

        sigma_sampler_config,

        type="l2",

        offset_noise_level=0.0,

        batch2model_keys: Optional[Union[str, List[str], ListConfig]] = None,

    ):
        super().__init__()

        assert type in ["l2", "l1", "lpips"]

        self.sigma_sampler = instantiate_from_config(sigma_sampler_config)

        self.type = type
        self.offset_noise_level = offset_noise_level

        if type == "lpips":
            self.lpips = LPIPS().eval()

        if not batch2model_keys:
            batch2model_keys = []

        if isinstance(batch2model_keys, str):
            batch2model_keys = [batch2model_keys]

        self.batch2model_keys = set(batch2model_keys)

    def __call__(self, network, denoiser, conditioner, input, batch):
        cond = conditioner(batch)
        additional_model_inputs = {
            key: batch[key] for key in self.batch2model_keys.intersection(batch)
        }

        sigmas = self.sigma_sampler(input.shape[0]).to(input.device)
        noise = torch.randn_like(input)
        if self.offset_noise_level > 0.0:
            noise = noise + self.offset_noise_level * append_dims(
                torch.randn(input.shape[0], device=input.device), input.ndim
            )
        noised_input = input + noise * append_dims(sigmas, input.ndim)
        model_output = denoiser(
            network, noised_input, sigmas, cond, **additional_model_inputs
        )
        w = append_dims(denoiser.w(sigmas), input.ndim)
        return self.get_loss(model_output, input, w)

    def get_loss(self, model_output, target, w):
        if self.type == "l2":
            return torch.mean(
                (w * (model_output - target) ** 2).reshape(target.shape[0], -1), 1
            )
        elif self.type == "l1":
            return torch.mean(
                (w * (model_output - target).abs()).reshape(target.shape[0], -1), 1
            )
        elif self.type == "lpips":
            loss = self.lpips(model_output, target).reshape(-1)
            return loss