bpr loss + 5X NS

models.py

@@ -52,6 +52,20 @@ class EmbeddingLearner(nn.Module):
         n_score = score[:, pos_num:]
         return p_score, n_score
 
+def bpr_loss(p_scores, n_values):
+    p1 = p_scores[:,0,None]
+    p2 = p_scores[:,1,None]
+
+    num_neg = n_values.shape[1]
+    half_index = int(num_neg/2)
+
+    d1 = torch.sub(p1,n_values[:,0:half_index])
+    d2 = torch.sub(p2,n_values[:,half_index:])
+
+    t = F.logsigmoid(torch.add(d1,d2))
+
+    loss = (-1) * t.sum() / n_values.shape[1]
+    return loss
 
 class MetaTL(nn.Module):
     def __init__(self, itemnum, parameter):
@@ -68,6 +82,7 @@ class MetaTL(nn.Module):
 
         self.embedding_learner = EmbeddingLearner()
         self.loss_func = nn.MarginRankingLoss(self.margin)
+
         self.rel_q_sharing = dict()
 
     def split_concat(self, positive, negative):
@@ -101,15 +116,15 @@ class MetaTL(nn.Module):
             y = torch.Tensor([1]).to(self.device)
             self.zero_grad()
 
-            sorted,indecies = torch.sort(n_score, descending=True,dim=1)
-            n_values = sorted[:,0:p_score.shape[1]]
-            loss = self.loss_func(p_score, n_values, y)
+            # sorted,indecies = torch.sort(n_score, descending=True,dim=1)
+            # n_values = sorted[:,0:p_score.shape[1]]
+            n_values = n_score
+            loss = bpr_loss(p_score,n_values)
+            # loss = self.loss_func(p_score, n_values, y)
             loss.backward(retain_graph=True)
 
             grad_meta = rel.grad
             rel_q = rel - self.beta*grad_meta
-
-
             self.rel_q_sharing[curr_rel] = rel_q
 
         rel_q = rel_q.expand(-1, num_q + num_n, -1, -1)

sampler.py

@@ -31,7 +31,7 @@ def sample_function_mixed(user_train, usernum, itemnum, batch_size, maxlen, resu
             
             seq = np.zeros([maxlen], dtype=np.int32)
             pos = np.zeros([maxlen], dtype=np.int32)
-            neg = np.zeros([maxlen*30], dtype=np.int32)
+            neg = np.zeros([maxlen*5], dtype=np.int32)
 
             if len(user_train[user]) < maxlen:
                 nxt_idx = len(user_train[user]) - 1
@@ -62,7 +62,7 @@ def sample_function_mixed(user_train, usernum, itemnum, batch_size, maxlen, resu
 
             # for idx in range(maxlen*30 - 1):
             #     support_negative_triples.append([seq[-1], curr_rel, neg[idx]])
-            for j in range(30):
+            for j in range(5):
                 for idx in range(maxlen-1):
                     support_negative_triples.append([seq[idx], curr_rel, neg[j*maxlen + idx]])
 

utils.py

@@ -119,7 +119,7 @@ class DataLoader(object):
         seq = np.zeros([self.maxlen], dtype=np.int32)
         pos = np.zeros([self.maxlen - 1], dtype=np.int32)
         # neg = np.zeros([self.maxlen*30 - 1], dtype=np.int32)
-        neg = np.zeros([self.maxlen * 30], dtype=np.int32)
+        neg = np.zeros([self.maxlen * 5], dtype=np.int32)
         
         idx = self.maxlen - 1
 
@@ -144,7 +144,7 @@ class DataLoader(object):
 
         # for idx in range(len(neg)):
         #     support_negative_triples.append([seq[-1],curr_rel,neg[idx]])
-        for j in range(30):
+        for j in range(5):
             for idx in range(self.maxlen-1):
                 support_negative_triples.append([seq[idx], curr_rel, neg[j * self.maxlen + idx]])