checking n-pair

loss/embedding_aug_mx.py

@@ -113,24 +113,29 @@ def get_min_dis(F, dis0,label,a1l,a2l):
 def get_sum_exp_dis(F, dis0,label,a1l,a2l):
     k=0
     for l in range(label.shape[0]):
-      id1=[i for i in range(a1l.shape[0]) if a1l[i]==label[l] ]
-      id2=[i for i in range(a2l.shape[0]) if a2l[i]==label[l] ]
+      id1=(a1l==label[l]) #[i for i in range(a1l.shape[0]) if a1l[i]==label[l] ]
+      id2=(a2l==label[l]) #[i for i in range(a2l.shape[0]) if a2l[i]==label[l] ]
 
-      if len(id1)>0 or len(id2)>0:
-        if len(id1)<1:
-          dist=F.sum(F.exp(-dis0[id2]))
-        elif len(id2)<1:
-          dist=F.sum(F.exp(-dis0[id1]))
+      if F.sum(id1)>0 or F.sum(id2)>0:
+        if F.sum(id1)<1:
+          dist=F.sum(F.exp(-F.contrib.boolean_mask(dis0,id2)))
+          num=F.sum(id2)
+        elif F.sum(id2)<1:
+          dist=F.sum(F.exp(-F.contrib.boolean_mask(dis0,id1)))
+          num=F.sum(id1)
         else:
-          dist=F.sum(F.concat(F.sum(F.exp(-dis0[id1])), F.sum(F.exp(-dis0[id2])), dim=0))
+          dist=F.sum(F.concat(F.sum(F.exp(-F.contrib.boolean_mask(dis0,id1))), F.sum(F.exp(-F.contrib.boolean_mask(dis0,id2))), dim=0))
+          num=F.sum(id1)+F.sum(id2)
 
         if k==0:
           k=k+1
           dis=dist
+          numf=num
         else:
           dis=F.concat(dis,dist,dim=0)
+          numf=F.concat(numf,num,dim=0)
 
-    return dis
+    return dis,numf
 
 def get_pos_dis(F, dis_ap, labelsorg):
     N = dis_ap.shape[0]
@@ -252,9 +257,27 @@ def pair_mining(F, dis_ap, dis_an, ids, a1l, a2l, ind, labels, num_ins, th, alph
 
     #print('COUNT FOR emptiness per pair...', float(count)/float(len(ids)))
     return dis_neg, dis_pos
+
+
+def check_corners(F,X1,X2,X3,X4):
+    dis13=F.expand_dims(F.sum(X1*X3, axis=0),axis=1)
+    dis14=F.expand_dims(F.sum(X1*X4, axis=0),axis=1)
+    dis23=F.expand_dims(F.sum(X2*X3, axis=0),axis=1)
+    dis24=F.expand_dims(F.sum(X2*X4, axis=0),axis=1)
+
+    dis1=dis13+7*(F.sign(dis13-dis14)**2-1+F.sign(dis13-dis23)**2-1+F.sign(dis13-dis24)**2-1)
+    dis2=dis14+8*(F.sign(dis14-dis23)**2-1+F.sign(dis14-dis24)**2-1)
+    dis3=dis23+9*(F.sign(dis23-dis24)**2-1)
+    dis4=dis24
 
+    dis=F.concat(dis1,dis2,dis3,dis4,dim=1)
+    #print(dis)
+    dis=F.max(dis,axis=1)
+    #print(dis)
+    return dis
+
 
-def get_opt_emb_dis(F, embeddings, labels, num_instance, l2_norm=True, multisim=False):
+def get_opt_emb_dis(F, embeddings, labels, num_instance, l2_norm=True, multisim=False, npair=False):
     batch_size = embeddings.shape[0]
     dim=embeddings.shape[1]
 
@@ -282,8 +305,8 @@ def get_opt_emb_dis(F, embeddings, labels, num_instance, l2_norm=True, multisim=
           X2l=X2l[ind]
 
       if len(ind)<2 or len(indx)<2:
-        print('============================YESS=======================================')
-        print('Similarities...', sim) 
+        #print('============================YESS=======================================')
+        #print('Similarities...', sim) 
         ind=[i for i in range(sim.shape[0])]
 
       if num_instance==2:
@@ -297,7 +320,10 @@ def get_opt_emb_dis(F, embeddings, labels, num_instance, l2_norm=True, multisim=
           print('dis_ap', dis_ap1)
 
     else:
-      dis_ap1 = F.sum(X1*X2, axis=1) #num_ins=2 for n-pair
+      if npair:
+        dis_ap1 = F.sum(X1*X2, axis=1) #num_ins=2 for n-pair
+      else:
+        dis_ap1 = F.sqrt(F.sum((X1-X2)*(X1-X2), axis=1)+1e-20)
 
     X1, X2, X3, X4, a1l, a2l, ids = concat(F,X1,X2,X1l,X2l)
 
@@ -309,6 +335,9 @@ def get_opt_emb_dis(F, embeddings, labels, num_instance, l2_norm=True, multisim=
         dis = F.sqrt(F.sum((X1-X3)*(X1-X3), axis=1)+1e-20)
 
     else:
+      if npair:
+        dis = check_corners(F,F.transpose(X1), F.transpose(X2), F.transpose(X3), F.transpose(X4))
+      else:
         dis = opt_pts_lin(F.transpose(X1), F.transpose(X2), F.transpose(X3), F.transpose(X4))
 
     #dis_an = get_min_dis(F, dis, ids, a1l, a2l) #for hphn-triplet

loss/ml_losses.py

@@ -133,11 +133,18 @@ def hybrid_forward(self, F, embeddings, labels):
         self.batch_size = embeddings.shape[0]
 
         gen_start_time = time.time()
-        dist_ap, dist_an0, ids, a1l, a2l = get_opt_emb_dis(F, embeddings, labels, self.num_instance, self.l2_norm)
-        dist_an = get_sum_exp_dis(F, -dist_an0, ids, a1l, a2l)
+        dist_ap, dist_an0, ids, a1l, a2l = get_opt_emb_dis(F, embeddings, labels, self.num_instance, l2_norm=False,npair=True )
+        dist_an, numf = get_sum_exp_dis(F, -dist_an0, ids, a1l, a2l)
+        print(dist_an)
+        print(F.exp(-dist_ap))
         gen_time = time.time() - gen_start_time
+
+        X1=embeddings[0:self.batch_size:2]
+        X2=embeddings[1:self.batch_size:2]
+        l2_reg=F.sum(X1*X1, axis = 1)+F.sum(X2*X2, axis = 1)
+        print(F.sqrt(F.sum((X1-X2)**2, axis = 1)))
 
-        loss = F.log(1.0 + F.exp(-dist_ap)*dist_an)
+        loss = F.log(1.0 + F.exp(-dist_ap)*dist_an) #+ 0.000075*l2_reg
 
         total_time = time.time() - total_start_time
 

main.py

@@ -220,7 +220,7 @@ def main():
             # evaluate_and_log(summary_writer, evaluator, ranks, step, epoch, best_metrics)
             best_metrics = evaluate_and_log(summary_writer, evaluator, args.recallk,
                                         global_step, epoch + 1,
-                                        best_metrics=best_metrics, args.data_name)
+                                        best_metrics=best_metrics, data_name=args.data_name)
             if best_metrics[0] != old_best_metric:
                 save_path = os.path.join(args.save_dir, 'model_epoch_%05d.params' % (epoch + 1))
                 model.save_parameters(save_path)