algorithms/GCL.py
- 1 F632: use ==/!= to compare constant literals (str, bytes, int, float, tuple)
- 1 F841: local variable 'idx_drop' is assigned to but never used
-
24
W291: trailing whitespace
- 38 # additional logging, copied from deepCORAL,
- 46 Original method operates on
- 59 to directly operate on the node index+edge lists,
- 74 # Noting that this isn't an elegant process because we need to
- 87 new_edge_attr = orig_edge_attr[edge_subselect]
- 106 Current:
- 107 In-place replacement k edges (i->j) with k edges (i'->j')
- 111 Currently the edge_attr of (i->j) in inherited by (i'->j')
- 112 for datasets that include edge_attr, this is not very
- 113 semantically sound, as the joint distribution between
- 114 (i_feat,j_feat,e_attr), i_feat,j_feat in {node_features},
- 118 A non-naive, 'paired' version will also rely on assumption
- 119 that this is a default pytorch geometric data object where
- 120 edge ordering is such that two the directions of a single
- 121 edge occur in pairs. So that we can operate on bidirectional
- 139 edges_to_insert = torch.multinomial(torch.ones(permute_num,node_num), 2, replacement=False).t().cuda()
- 143 raise NotImplementedError
- 187 # 1) In one version of the pretraining reference code,
- 188 # augmentations are applied in the get(idx) method of
- 190 # the pair of data objects, data, data_aug
- 205 # we're going to start by just applying the augmentation here in line,
- 208 # be parallelized by the dataloader workers ... instead it's blocking
- 250 # if not "use_cl" unpack batch of graphs and apply augmentation
- 259 for i in range(batch_size):
-
20
E501: line too long (135 > 79 characters)
(and 15 similar)
-
Line too long (135 > 79 characters):
12 featurizer, projector, classifier, combined_model = initialize_model(config, d_out, is_featurizer=True, include_projector=True) -
Line too long (86 > 79 characters):
71 idx_nondrop = idx_perm[drop_num:].sort().values.cuda() # sort for humans/debug -
Line too long (87 > 79 characters):
78 src_subselect = torch.nonzero(idx_nondrop[...,None] == orig_edge_index[0])[:,1] -
Line too long (88 > 79 characters):
79 dest_subselect = torch.nonzero(idx_nondrop[...,None] == orig_edge_index[1])[:,1] -
Line too long (116 > 79 characters):
80 edge_subselect = src_subselect[torch.nonzero(src_subselect[..., None] == dest_subselect)[:,0]].sort().values -
Line too long (114 > 79 characters):
139 edges_to_insert = torch.multinomial(torch.ones(permute_num,node_num), 2, replacement=False).t().cuda() -
Line too long (104 > 79 characters):
140 insertion_indices = torch.multinomial(torch.ones(edge_num), permute_num, replacement=False) -
Line too long (86 > 79 characters):
145 orig_edge_index[:,insertion_indices] = edges_to_insert # modified in place -
Line too long (99 > 79 characters):
168 # See Appendix A Algo 1 https://yyou1996.github.io/files/neurips2020_graphcl_supplement.pdf -
Line too long (163 > 79 characters):
169 # Code from: https://github.com/Shen-Lab/GraphCL/blob/d857849d51bb168568267e07007c0b0c8bb6d869/transferLearning_MoleculeNet_PPI/bio/pretrain_graphcl.py#L57 -
Line too long (96 > 79 characters):
175 sim_matrix = torch.einsum('ik,jk->ij', z1, z2) / torch.einsum('i,j->ij', z1_abs, z2_abs) -
Line too long (124 > 79 characters):
191 # See: https://github.com/Shen-Lab/GraphCL/blob/e9e598d478d4a4bff94a3e95a078569c028f1d88/unsupervised_TU/aug.py#L203 -
Line too long (84 > 79 characters):
193 # and since they are not shuffled, the orig and augmented graphs are aligned -
Line too long (158 > 79 characters):
194 # See: https://github.com/Shen-Lab/GraphCL/blob/d857849d51bb168568267e07007c0b0c8bb6d869/transferLearning_MoleculeNet_PPI/chem/pretrain_graphcl.py#L80 -
Line too long (82 > 79 characters):
196 # formulation using einsum and subtracting of the diagonal to separate out -
Line too long (82 > 79 characters):
197 # positive pairs from negative pairs, you can compute the contrastive loss -
Line too long (83 > 79 characters):
207 # initial intuition is that this might be pretty inefficient since it can't -
Line too long (82 > 79 characters):
209 # during training, though, the computations are all torch or native python -
Line too long (114 > 79 characters):
248 aug_mask = torch.multinomial(torch.tensor((1-self.aug_prob, self.aug_prob)), batch_size, replacement=True) -
Line too long (89 > 79 characters):
300 return self.loss.compute(results['y_pred'], results['y_true'], return_dict=False)
-
-
9
E231: missing whitespace after ','
-
Missing whitespace after ',' (in 2 places):
78 src_subselect = torch.nonzero(idx_nondrop[...,None] == orig_edge_index[0])[:,1] -
Missing whitespace after ',' (in 2 places):
79 dest_subselect = torch.nonzero(idx_nondrop[...,None] == orig_edge_index[1])[:,1] -
Missing whitespace after ',':
80 edge_subselect = src_subselect[torch.nonzero(src_subselect[..., None] == dest_subselect)[:,0]].sort().values -
Missing whitespace after ',':
82 new_edge_index = orig_edge_index[:,edge_subselect] -
Missing whitespace after ',':
139 edges_to_insert = torch.multinomial(torch.ones(permute_num,node_num), 2, replacement=False).t().cuda() -
Missing whitespace after ',':
145 orig_edge_index[:,insertion_indices] = edges_to_insert # modified in place -
Missing whitespace after ',':
235 n = torch.randint(2,(1,)).item()
-
-
5
E261: at least two spaces before inline comment
- 71 idx_nondrop = idx_perm[drop_num:].sort().values.cuda() # sort for humans/debug
- 131 paired_perms = False # for later, TODO
- 132 if paired_perms: assert edge_num%2 is 0 # undirected g, paired edges
- 145 orig_edge_index[:,insertion_indices] = edges_to_insert # modified in place
- 264 pass # original graph kept
- 3 E222: multiple spaces after operator
- 3 E265: block comment should start with '# '
- 3 W293: blank line contains whitespace
- 1 E221: multiple spaces before operator
- 1 E228: missing whitespace around modulo operator
- 1 E302: expected 2 blank lines, found 1
- 1 E303: too many blank lines (2)
- 1 E701: multiple statements on one line (colon)
- 1 E722: do not use bare 'except'