Fixed some criteria formatting and missing flags in multisimilarity.py

Author: Confusezius

criteria/adversarial_separation.py

@@ -21,11 +21,6 @@ def __init__(self, opt):
         super().__init__()
 
         ####
-        self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
-        self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
-        self.REQUIRES_OPTIM      = REQUIRES_OPTIM
-
-        ####        
         self.embed_dim  = opt.embed_dim
         self.proj_dim   = opt.diva_decorrnet_dim
 
@@ -43,6 +38,14 @@ def __init__(self, opt):
         self.lr        = opt.diva_decorrnet_lr
 
 
+        ####
+        self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
+        self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
+        self.REQUIRES_OPTIM      = REQUIRES_OPTIM
+
+
+        
+
     def forward(self, feature_dict):
         #Apply gradient reversal on input embeddings.
         adj_feature_dict = {key:torch.nn.functional.normalize(grad_reverse(features),dim=-1) for key, features in feature_dict.items()}

criteria/angular.py

@@ -24,6 +24,8 @@ def __init__(self, opt, batchminer):
         self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
         self.REQUIRES_OPTIM      = REQUIRES_OPTIM
 
+        
+
     def forward(self, batch, labels, **kwargs):
         ####NOTE: Normalize Angular Loss, but dont normalize npair loss!
         anchors, positives, negatives = self.batchminer(batch, labels)

criteria/arcface.py

@@ -13,11 +13,6 @@ def __init__(self, opt):
         super(Criterion, self).__init__()
         self.par = opt
 
-        ####
-        self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
-        self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
-        self.REQUIRES_OPTIM      = REQUIRES_OPTIM
-
         ####
         self.angular_margin = opt.loss_arcface_angular_margin
         self.feature_scale  = opt.loss_arcface_feature_scale
@@ -30,6 +25,14 @@ def __init__(self, opt):
 
         self.lr    = opt.loss_arcface_lr
 
+        ####
+        self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
+        self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
+        self.REQUIRES_OPTIM      = REQUIRES_OPTIM
+
+
+
+
     def forward(self, batch, labels, **kwargs):
         bs, labels = len(batch), labels.to(self.par.device)
 

criteria/contrastive.py

@@ -22,6 +22,8 @@ def __init__(self, opt, batchminer):
         self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
         self.REQUIRES_OPTIM      = REQUIRES_OPTIM
 
+
+
     def forward(self, batch, labels, **kwargs):
         sampled_triplets = self.batchminer(batch, labels)
 

criteria/lifted.py

@@ -22,6 +22,8 @@ def __init__(self, opt, batchminer):
         self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
         self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
         self.REQUIRES_OPTIM      = REQUIRES_OPTIM
+
+
 
     def forward(self, batch, labels, **kwargs):
         anchors, positives, negatives = self.batchminer(batch, labels)

criteria/margin.py

@@ -34,6 +34,8 @@ def __init__(self, opt, batchminer):
         self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
         self.REQUIRES_OPTIM      = REQUIRES_OPTIM
 
+
+
     def forward(self, batch, labels, **kwargs):
         sampled_triplets = self.batchminer(batch, labels)
 

criteria/multisimilarity.py

@@ -19,6 +19,12 @@ def __init__(self, opt):
 
         self.name           = 'multisimilarity'
 
+        ####
+        self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
+        self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
+        self.REQUIRES_OPTIM      = REQUIRES_OPTIM
+
+
     def forward(self, batch, labels, **kwargs):
         similarity = batch.mm(batch.T)
 

criteria/proxynca.py

@@ -17,11 +17,6 @@ def __init__(self, opt):
         """
         super(Criterion, self).__init__()
 
-        ####
-        self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
-        self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
-        self.REQUIRES_OPTIM      = REQUIRES_OPTIM
-
         ####
         self.num_proxies        = opt.n_classes
         self.embed_dim          = opt.embed_dim
@@ -34,6 +29,13 @@ def __init__(self, opt):
         self.optim_dict_list = [{'params':self.proxies, 'lr':opt.lr * opt.loss_proxynca_lrmulti}]
 
 
+        ####
+        self.ALLOWED_MINING_OPS  = ALLOWED_MINING_OPS
+        self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
+        self.REQUIRES_OPTIM      = REQUIRES_OPTIM
+
+        
+
     def forward(self, batch, labels, **kwargs):
         #Empirically, multiplying the embeddings during the computation of the loss seem to allow for more stable training;
         #Acts as a temperature in the NCA objective.

criteria/quadruplet.py

@@ -22,6 +22,8 @@ def __init__(self, opt, batchminer):
         self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
         self.REQUIRES_OPTIM      = REQUIRES_OPTIM
 
+
+
     def triplet_distance(self, anchor, positive, negative):
         return torch.nn.functional.relu(torch.norm(anchor-positive, p=2, dim=-1)-torch.norm(anchor-negative, p=2, dim=-1)+self.margin_alpha_1)
 

criteria/snr.py

@@ -24,6 +24,8 @@ def __init__(self, opt, batchminer):
         self.REQUIRES_BATCHMINER = REQUIRES_BATCHMINER
         self.REQUIRES_OPTIM      = REQUIRES_OPTIM
 
+
+
     def forward(self, batch, labels, **kwargs):
         sampled_triplets = self.batchminer(batch, labels)
         anchors   = [triplet[0] for triplet in sampled_triplets]