Don't apply activations on export in classification

mpa/modules/models/heads/custom_cls_head.py

@@ -2,6 +2,9 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
+import torch
+import torch.nn.functional as F
+
 from mmcls.models.builder import HEADS
 from mmcls.models.heads import LinearClsHead
 from .non_linear_cls_head import NonLinearClsHead
@@ -83,6 +86,17 @@ def loss(self, cls_score, gt_label, feature=None):
         losses['loss'] = loss
         return losses
 
+    def simple_test(self, img):
+        """Test without augmentation."""
+        cls_score = self.fc(img)
+        if isinstance(cls_score, list):
+            cls_score = sum(cls_score) / float(len(cls_score))
+        if torch.onnx.is_in_onnx_export():
+            return cls_score
+        pred = F.softmax(cls_score, dim=1) if cls_score is not None else None
+
+        return self.post_process(pred)
+
     def forward_train(self, x, gt_label):
         cls_score = self.fc(x)
         losses = self.loss(cls_score, gt_label, feature=x)

mpa/modules/models/heads/custom_hierarchical_linear_cls_head.py

@@ -115,13 +115,17 @@ def simple_test(self, img):
         for i in range(self.hierarchical_info['num_multiclass_heads']):
             multiclass_logit = cls_score[:, self.hierarchical_info['head_idx_to_logits_range'][i][0]:
                                             self.hierarchical_info['head_idx_to_logits_range'][i][1]]
-            multiclass_logit = torch.softmax(multiclass_logit, dim=1)
+            if not torch.onnx.is_in_onnx_export():
+                multiclass_logit = torch.softmax(multiclass_logit, dim=1)
             multiclass_logits.append(multiclass_logit)
         multiclass_pred = torch.cat(multiclass_logits, dim=1) if multiclass_logits else None
 
         if self.compute_multilabel_loss:
             multilabel_logits = cls_score[:, self.hierarchical_info['num_single_label_classes']:]
-            multilabel_pred = torch.sigmoid(multilabel_logits) if multilabel_logits is not None else None
+            if not torch.onnx.is_in_onnx_export():
+                multilabel_pred = torch.sigmoid(multilabel_logits) if multilabel_logits is not None else None
+            else:
+                multilabel_pred = multilabel_logits
             if multiclass_pred is not None:
                 pred = torch.cat([multiclass_pred, multilabel_pred], axis=1)
             else:

mpa/modules/models/heads/custom_multi_label_linear_cls_head.py

@@ -81,9 +81,9 @@ def simple_test(self, img):
         cls_score = self.fc(img) * self.scale
         if isinstance(cls_score, list):
             cls_score = sum(cls_score) / float(len(cls_score))
-        pred = torch.sigmoid(cls_score) if cls_score is not None else None
         if torch.onnx.is_in_onnx_export():
-            return pred
+            return cls_score
+        pred = torch.sigmoid(cls_score) if cls_score is not None else None
         pred = list(pred.detach().cpu().numpy())
         return pred
 

mpa/modules/models/heads/custom_multi_label_non_linear_cls_head.py

@@ -105,9 +105,9 @@ def simple_test(self, img):
         cls_score = self.classifier(img) * self.scale
         if isinstance(cls_score, list):
             cls_score = sum(cls_score) / float(len(cls_score))
-        pred = torch.sigmoid(cls_score) if cls_score is not None else None
         if torch.onnx.is_in_onnx_export():
-            return pred
+            return cls_score
+        pred = torch.sigmoid(cls_score) if cls_score is not None else None
         pred = list(pred.detach().cpu().numpy())
         return pred
 

mpa/modules/models/heads/non_linear_cls_head.py

@@ -76,9 +76,9 @@ def simple_test(self, img):
         cls_score = self.classifier(img)
         if isinstance(cls_score, list):
             cls_score = sum(cls_score) / float(len(cls_score))
-        pred = F.softmax(cls_score, dim=1) if cls_score is not None else None
         if torch.onnx.is_in_onnx_export():
-            return pred
+            return cls_score
+        pred = F.softmax(cls_score, dim=1) if cls_score is not None else None
         pred = list(pred.detach().cpu().numpy())
         return pred