Spaces:
Sleeping
Sleeping
File size: 3,796 Bytes
207ef6f |
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 |
import time
import numpy as np
import torch
from torch import nn
import torch.nn.functional as F
class AdaptiveSupervisedPatchNCELoss(nn.Module):
def __init__(self, opt):
super().__init__()
self.opt = opt
self.cross_entropy_loss = torch.nn.CrossEntropyLoss(reduction='none')
self.mask_dtype = torch.bool
self.total_epochs = opt.n_epochs + opt.n_epochs_decay
def forward(self, feat_q, feat_k, current_epoch=-1):
num_patches = feat_q.shape[0]
dim = feat_q.shape[1]
feat_k = feat_k.detach()
# pos logit
l_pos = torch.bmm(
feat_q.view(num_patches, 1, -1), feat_k.view(num_patches, -1, 1))
l_pos = l_pos.view(num_patches, 1)
# neg logit
# Should the negatives from the other samples of a minibatch be utilized?
# In CUT and FastCUT, we found that it's best to only include negatives
# from the same image. Therefore, we set
# --nce_includes_all_negatives_from_minibatch as False
# However, for single-image translation, the minibatch consists of
# crops from the "same" high-resolution image.
# Therefore, we will include the negatives from the entire minibatch.
if self.opt.nce_includes_all_negatives_from_minibatch:
# reshape features as if they are all negatives of minibatch of size 1.
batch_dim_for_bmm = 1
else:
batch_dim_for_bmm = self.opt.batch_size
# reshape features to batch size
feat_q = feat_q.view(batch_dim_for_bmm, -1, dim)
feat_k = feat_k.view(batch_dim_for_bmm, -1, dim)
npatches = feat_q.size(1)
l_neg_curbatch = torch.bmm(feat_q, feat_k.transpose(2, 1))
# diagonal entries are similarity between same features, and hence meaningless.
# just fill the diagonal with very small number, which is exp(-10) and almost zero
diagonal = torch.eye(npatches, device=feat_q.device, dtype=self.mask_dtype)[None, :, :]
l_neg_curbatch.masked_fill_(diagonal, -10.0)
l_neg = l_neg_curbatch.view(-1, npatches)
out = torch.cat((l_pos, l_neg), dim=1) / self.opt.nce_T
loss = self.cross_entropy_loss(out, torch.zeros(out.size(0), dtype=torch.long,
device=feat_q.device))
if self.opt.asp_loss_mode == 'none':
return loss
scheduler, lookup = self.opt.asp_loss_mode.split('_')[:2]
# Compute scheduling
t = (current_epoch - 1) / self.total_epochs
if scheduler == 'sigmoid':
p = 1 / (1 + np.exp((t - 0.5) * 10))
elif scheduler == 'linear':
p = 1 - t
elif scheduler == 'lambda':
k = 1 - self.opt.n_epochs_decay / self.total_epochs
m = 1 / (1 - k)
p = m - m * t if t >= k else 1.0
elif scheduler == 'zero':
p = 1.0
else:
raise ValueError(f"Unrecognized scheduler: {scheduler}")
# Weight lookups
w0 = 1.0
x = l_pos.squeeze().detach()
if lookup == 'top':
x = torch.where(x > 0.0, x, torch.zeros_like(x))
w1 = torch.sqrt(1 - (x - 1) ** 2)
elif lookup == 'linear':
w1 = torch.relu(x)
elif lookup == 'bell':
sigma, mu, sc = 1, 0, 4
w1 = 1 / (sigma * np.sqrt(2 * torch.pi)) * torch.exp(-((x - 0.5) * sc - mu) ** 2 / (2 * sigma ** 2))
elif lookup == 'uniform':
w1 = torch.ones_like(x)
else:
raise ValueError(f"Unrecognized lookup: {lookup}")
# Apply weights with schedule
w = p * w0 + (1 - p) * w1
# Normalize
w = w / w.sum() * len(w)
loss = loss * w
return loss
|