CHANGES

Author: wbw

draw_tools.py

@@ -8,7 +8,7 @@
 
 
 def draw_bar(data):
-    plt.figure(figsize=(8, 6), dpi=80)
+    plt.figure(figsize=(20, 6), dpi=80)
     font = 22
     x_bar = np.arange(0, len(data), 1)
     plt.bar(x_bar, data)
@@ -22,7 +22,7 @@ def draw_bar(data):
 
 def draw_plot(data):
     font = 22
-    plt.figure(figsize=(8, 6), dpi=80)
+    plt.figure(figsize=(20, 6), dpi=80)
     b, c = data.shape
     for i in range(c):
         plt.boxplot(data[:, i], positions=[i*10], widths=5, showmeans=True)

engine.py

@@ -48,17 +48,17 @@ def train(args, model, device, loader, optimizer, epoch):
             pred_acces.update(acc)
 
             if epoch >= args.lr_drop:
-                s = 0
-                k = 2
+                s = 5
+                k = 5
                 q = 5
                 t = 2
             else:
-                s = 0
+                s = 1
                 k = 1
                 q = 1
-                t = 0.5
+                t = 1
 
-            loss_total = retri_loss + s * attn_loss + t * quantity_loss + 0.5 * loss_pred + q * consistence_loss - k * batch_dis_loss + 0 * att_dis_loss
+            loss_total = retri_loss + s * attn_loss + t * quantity_loss + 0.5 * loss_pred - q * consistence_loss + k * batch_dis_loss
         else:
             cpt = model(data)
             retri_loss, quantity_loss = get_retrieval_loss(cpt, label, args.num_classes, device)

loss.py

@@ -1,5 +1,6 @@
 import torch
 from torch.autograd import Variable
+import torch.nn.functional as F
 
 
 def pairwise_loss(outputs1, outputs2, label1, label2, sigmoid_param=1.):
@@ -44,10 +45,10 @@ def batch_cpt_discriminate(data, att):
         current_att = att.sum(-1)[:, i]
         indices = current_att > current_att.mean()
         b, d = current_f[indices].shape
-        current_f = current_f[indices] # / current_att[indices].unsqueeze(-1).expand(b, d)
+        current_f = current_f[indices] / current_att[indices].unsqueeze(-1).expand(b, d)
         record.append(torch.mean(current_f, dim=0, keepdim=True))
     record = torch.cat(record, dim=0)
-    sim = ((record[None, :, :] - record[:, None, :]) ** 2).sum(-1)
+    sim = F.cosine_similarity(record[None, :, :], record[:, None, :], dim=-1)
     return sim.mean()
 
 
@@ -76,8 +77,8 @@ def att_consistence(update, att):
         current_att = att[:, i, :].sum(-1)
         indices = current_att > current_att.mean()
         b, d = current_up[indices].shape
-        need = current_up[indices] # / current_att[indices].unsqueeze(-1).expand(b, d)
-        consistence_loss += ((need[None, :, :] - need[:, None, :]) ** 2).sum(-1).mean()
+        need = current_up[indices] / current_att[indices].unsqueeze(-1).expand(b, d)
+        consistence_loss += F.cosine_similarity(need[None, :, :], need[:, None, :], dim=-1).mean()
     return consistence_loss/cpt
 
 

model/slots.py

@@ -55,8 +55,9 @@ def forward(self, inputs_pe, inputs, weight=None, things=None):
             else:
                 raise RuntimeError(f"unsupported input to tensor dot, got slot mode={self.slot_mode}")
 
+            # attn2 = attn / (attn.sum(dim=-1, keepdim=True) + self.eps)
             updates = torch.einsum('bjd,bij->bid', inputs, attn)
-            updates = updates / inputs.size(2)
+            # updates = updates / (attn.sum(-1).unsqueeze(-1) + 1)
 
         if self.vis:
             slots_vis_raw = attn.clone()

vis_retri.py

@@ -90,7 +90,7 @@ def main():
     test_labels = f1["test_labels"]
 
     query = np.zeros((1, args.num_cpt)) - 1
-    location = 2
+    location = 14
     query[0][location] = 1
     ids = for_retrival(args, np.array(database_hash), query, location=location)
     print("-------------------------")