MCPcopy Create free account
hub / github.com/TPCD/DCCL / SupConLoss

Class SupConLoss

methods/representation_learning/representation_learning.py:35–123  ·  view source on GitHub ↗

Supervised Contrastive Learning: https://arxiv.org/pdf/2004.11362.pdf. It also supports the unsupervised contrastive loss in SimCLR From: https://github.com/HobbitLong/SupContrast

Source from the content-addressed store, hash-verified

33from project_utils.k_means_utils import test_kmeans_semi_sup
34
35class SupConLoss(torch.nn.Module):
36 """Supervised Contrastive Learning: https://arxiv.org/pdf/2004.11362.pdf.
37 It also supports the unsupervised contrastive loss in SimCLR
38 From: https://github.com/HobbitLong/SupContrast"""
39 def __init__(self, temperature=0.07, contrast_mode='all',
40 base_temperature=0.07):
41 super(SupConLoss, self).__init__()
42 self.temperature = temperature
43 self.contrast_mode = contrast_mode
44 self.base_temperature = base_temperature
45
46 def forward(self, features, labels=None, mask=None):
47 """Compute loss for model. If both `labels` and `mask` are None,
48 it degenerates to SimCLR unsupervised loss:
49 https://arxiv.org/pdf/2002.05709.pdf
50 Args:
51 features: hidden vector of shape [bsz, n_views, ...].
52 labels: ground truth of shape [bsz].
53 mask: contrastive mask of shape [bsz, bsz], mask_{i,j}=1 if sample j
54 has the same class as sample i. Can be asymmetric.
55 Returns:
56 A loss scalar.
57 """
58 device = (torch.device('cuda')
59 if features.is_cuda
60 else torch.device('cpu'))
61
62 if len(features.shape) < 3:
63 raise ValueError('`features` needs to be [bsz, n_views, ...],'
64 'at least 3 dimensions are required')
65 if len(features.shape) > 3:
66 features = features.view(features.shape[0], features.shape[1], -1)
67
68 batch_size = features.shape[0]
69 if labels is not None and mask is not None:
70 raise ValueError('Cannot define both `labels` and `mask`')
71 elif labels is None and mask is None:
72 mask = torch.eye(batch_size, dtype=torch.float32).to(device)
73 elif labels is not None:
74 labels = labels.contiguous().view(-1, 1)
75 if labels.shape[0] != batch_size:
76 raise ValueError('Num of labels does not match num of features')
77 mask = torch.eq(labels, labels.T).float().to(device)
78 else:
79 mask = mask.float().to(device)
80
81 contrast_count = features.shape[1]
82 contrast_feature = torch.cat(torch.unbind(features, dim=1), dim=0)
83 if self.contrast_mode == 'one':
84 anchor_feature = features[:, 0]
85 anchor_count = 1
86 elif self.contrast_mode == 'all':
87 anchor_feature = contrast_feature
88 anchor_count = contrast_count
89 else:
90 raise ValueError('Unknown mode: {}'.format(self.contrast_mode))
91
92 # compute logits

Callers 1

trainFunction · 0.70

Calls

no outgoing calls

Tested by

no test coverage detected