fix(test-2label): --PCB and --use_dense flags select (+1 more)

[andrewwhitecdw]

Small fixes in reid_eval/test_2label.py:

fix: --PCB and --use_dense flags select unsupported model

Fix: Replace:

model_structure = ft_netAB(config['ID_class'], norm=config['norm_id'], stride=config['ID_stride'], pool=config['pool'])

if opt.PCB:
    model_structure = PCB(config['ID_class'])

with:

if opt.use_dense or opt.PCB:
    raise NotImplementedError("Only ft_netAB is supported by this test script")

model_structure = ft_netAB(config['ID_class'], norm=config['norm_id'], stride=config['ID_stride'], pool=config['pool'])

fix: PCB test buffer uses 6 parts but PCB model defines 4

Fix: Apply patch:

--- a/reid_eval/test_2label.py
+++ b/reid_eval/test_2label.py
@@ -95,7 +95,7 @@ def extract_feature(model,dataloaders):
         if opt.use_dense:
             ff = torch.FloatTensor(n,1024).zero_()
         else:
             ff = torch.FloatTensor(n,1024).zero_()
         if opt.PCB:
-            ff = torch.FloatTensor(n,2048,6).zero_() # we have six parts
+            ff = torch.FloatTensor(n,2048,4).zero_() # we have four parts
         for i in range(2):
             if(i==1):
@@ -116,8 +116,8 @@ def extract_feature(model,dataloaders):
 
         # norm feature
         if opt.PCB:
-            # feature size (n,2048,6)
+            # feature size (n,2048,4)
             # 1. To treat every part equally, I calculate the norm for every 2048-dim part feature.
-            # 2. To keep the cosine score==1, sqrt(6) is added to norm the whole feature (2048*6).
-            fnorm = torch.norm(ff, p=2, dim=1, keepdim=True) * np.sqrt(6) 
+            # 2. To keep the cosine score==1, sqrt(4) is added to norm the whole feature (2048*4).
+            fnorm = torch.norm(ff, p=2, dim=1, keepdim=True) * np.sqrt(4) 
             ff = ff.div(fnorm.expand_as(ff))
             ff = ff.view(ff.size(0), -1)

Files changed

• reid_eval/test_2label.py