test before refactoring

bosaris/key.py

@@ -235,8 +235,8 @@ class Key:
         with h5py.File(input_file_fame, "r") as f:
 
             key = Key()
-            key.modelset = f.get("modelset")[()]
-            key.segset = f.get("segset")[()]
+            key.modelset = f["modelset"][()]
+            key.segset = f["segset"][()]
 
             # if running python 3, need a conversion to unicode
             if sys.version_info[0] == 3:

nnet/loss.py

@@ -251,13 +251,10 @@ class ArcMarginProduct(torch.nn.Module):
         self.mm = math.sin(math.pi - self.m) * self.m
 
     def forward(self, input, target=None):
-        if target is not None:
-            assert input.size()[0] == target.size()[0]
-        assert input.size()[1] == self.in_features
-
         # cos(theta)
-        cosine = torch.nn.functional.linear(torch.nn.functional.normalize(input),
-                                            torch.nn.functional.normalize(self.weight))
+        cosine = torch.nn.functional.linear(torch.nn.functional.normalize(input), torch.nn.functional.normalize(self.weight))
+        if target == None:
+            return cosine * self.s
         # cos(theta + m)
         sine = torch.sqrt((1.0 - torch.mul(cosine, cosine)).clamp(0, 1))
         phi = cosine * self.cos_m - sine * self.sin_m
@@ -278,7 +275,7 @@ class ArcMarginProduct(torch.nn.Module):
 
 class SoftmaxAngularProto(torch.nn.Module):
     # from https://github.com/clovaai/voxceleb_trainer/blob/3bfd557fab5a3e6cd59d717f5029b3a20d22a281/loss/angleproto.py
-    def __init__(self, spk_count, init_w=10.0, init_b=-5.0, **kwargs):
+    def __init__(self, spk_count, emb_dim=256, init_w=10.0, init_b=-5.0, **kwargs):
         super(SoftmaxAngularProto, self).__init__()
 
         self.test_normalize = True
@@ -288,7 +285,7 @@ class SoftmaxAngularProto(torch.nn.Module):
         self.criterion  = torch.nn.CrossEntropyLoss()
 
         self.cce_backend = torch.nn.Sequential(OrderedDict([
-                    ("linear8", torch.nn.Linear(256, spk_count))
+                    ("linear8", torch.nn.Linear(emb_dim, spk_count))
                 ]))
 
     def forward(self, x, target=None):