MCPcopy Create free account
hub / github.com/JunlinHan/DCLGAN / PatchNCELoss2

Class PatchNCELoss2

models/patchnce.py:53–100  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

51
52# Used in vanilla CUT
53class PatchNCELoss2(nn.Module):
54 def __init__(self, opt):
55 super().__init__()
56 self.opt = opt
57 self.cross_entropy_loss = torch.nn.CrossEntropyLoss(reduction='none')
58 self.mask_dtype = torch.uint8 if version.parse(torch.__version__) < version.parse('1.2.0') else torch.bool
59
60 def forward(self, feat_q, feat_k):
61 batchSize = feat_q.shape[0]
62 dim = feat_q.shape[1]
63 feat_k = feat_k.detach()
64
65 # pos logit
66 l_pos = torch.bmm(feat_q.view(batchSize, 1, -1), feat_k.view(batchSize, -1, 1))
67 l_pos = l_pos.view(batchSize, 1)
68
69 # neg logit
70
71 # Should the negatives from the other samples of a minibatch be utilized?
72 # In CUT and FastCUT, we found that it's best to only include negatives
73 # from the same image. Therefore, we set
74 # --nce_includes_all_negatives_from_minibatch as False
75 # However, for single-image translation, the minibatch consists of
76 # crops from the "same" high-resolution image.
77 # Therefore, we will include the negatives from the entire minibatch.
78 if self.opt.nce_includes_all_negatives_from_minibatch:
79 # reshape features as if they are all negatives of minibatch of size 1.
80 batch_dim_for_bmm = 1
81 else:
82 batch_dim_for_bmm = self.opt.batch_size
83
84 # reshape features to batch size
85 feat_q = feat_q.view(batch_dim_for_bmm, -1, dim)
86 feat_k = feat_k.view(batch_dim_for_bmm, -1, dim)
87 npatches = feat_q.size(1)
88 l_neg_curbatch = torch.bmm(feat_q, feat_k.transpose(2, 1))
89
90 # diagonal entries are similarity between same features, and hence meaningless.
91 # just fill the diagonal with very small number, which is exp(-10) and almost zero
92 diagonal = torch.eye(npatches, device=feat_q.device, dtype=self.mask_dtype)[None, :, :]
93 l_neg_curbatch.masked_fill_(diagonal, -10.0)
94 l_neg = l_neg_curbatch.view(-1, npatches)
95
96 out = torch.cat((l_pos, l_neg), dim=1) / self.opt.nce_T
97
98 loss = self.cross_entropy_loss(out, torch.zeros(out.size(0), dtype=torch.long,
99 device=feat_q.device))
100 return loss
101

Callers 2

__init__Method · 0.85
__init__Method · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected