Add softmax to cls model (#1573)

* Add softmax to cls model

* fix cls ci

* multihead

* update classification_model.py

mmdeploy/codebase/mmcls/deploy/classification.py

@@ -333,8 +333,8 @@ def get_postprocess(self, *args, **kwargs) -> Dict:
         if 'topk' not in postprocess:
             topk = (1, )
             logger = get_root_logger()
-            logger.warning('no topk in postprocess config, using default \
-                 topk value.')
+            logger.warning('no topk in postprocess config, using default '
+                           'topk value.')
         else:
             topk = postprocess.topk
         postprocess.topk = max(topk)

mmdeploy/codebase/mmcls/deploy/classification_model.py

@@ -35,6 +35,7 @@ def __init__(self,
                  backend: Backend,
                  backend_files: Sequence[str],
                  device: str,
+                 model_cfg: Union[str, Config] = None,
                  deploy_cfg: Union[str, Config] = None,
                  data_preprocessor: Optional[Union[dict, nn.Module]] = None,
                  **kwargs):
@@ -46,8 +47,18 @@ def __init__(self,
             backend_files=backend_files,
             device=device,
             **kwargs)
+        self.model_cfg = model_cfg
+        self.head = None
+        if model_cfg is not None:
+            self.head = self._get_head()
         self.device = device
 
+    def _get_head(self):
+        from mmcls.models import build_head
+        head_config = self.model_cfg['model']['head']
+        head = build_head(head_config)
+        return head
+
     def _init_wrapper(self, backend: Backend, backend_files: Sequence[str],
                       device: str, **kwargs):
         output_names = self.output_names
@@ -84,11 +95,38 @@ def forward(self,
         cls_score = self.wrapper({self.input_name:
                                   inputs})[self.output_names[0]]
 
-        from mmcls.models.heads.cls_head import ClsHead
-        predict = ClsHead._get_predictions(
-            None, cls_score, data_samples=data_samples)
-
-        return predict
+        from mmcls.models.heads import MultiLabelClsHead
+        from mmcls.structures import ClsDataSample
+        pred_scores = cls_score
+
+        if self.head is None or not isinstance(self.head, MultiLabelClsHead):
+            pred_labels = pred_scores.argmax(dim=1, keepdim=True).detach()
+
+            if data_samples is not None:
+                for data_sample, score, label in zip(data_samples, pred_scores,
+                                                     pred_labels):
+                    data_sample.set_pred_score(score).set_pred_label(label)
+            else:
+                data_samples = []
+                for score, label in zip(pred_scores, pred_labels):
+                    data_samples.append(ClsDataSample().set_pred_score(
+                        score).set_pred_label(label))
+        else:
+            if data_samples is None:
+                data_samples = [
+                    ClsDataSample() for _ in range(cls_score.size(0))
+                ]
+
+            for data_sample, score in zip(data_samples, pred_scores):
+                if self.head.thr is not None:
+                    # a label is predicted positive if larger than thr
+                    label = torch.where(score >= self.head.thr)[0]
+                else:
+                    # top-k labels will be predicted positive for any example
+                    _, label = score.topk(self.head.topk)
+                data_sample.set_pred_score(score).set_pred_label(label)
+
+        return data_samples
 
 
 @__BACKEND_MODEL.register_module('sdk')
@@ -204,6 +242,7 @@ def build_classification_model(
             backend=backend,
             backend_files=model_files,
             device=device,
+            model_cfg=model_cfg,
             deploy_cfg=deploy_cfg,
             data_preprocessor=data_preprocessor,
             **kwargs))

mmdeploy/codebase/mmcls/models/classifiers/base.py

@@ -1,8 +1,10 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from typing import List, Optional
 
+import torch
 from mmengine.structures import BaseDataElement
 from torch import Tensor
+from torch.nn import functional as F
 
 from mmdeploy.core import FUNCTION_REWRITER
 
@@ -32,4 +34,10 @@ def base_classifier__forward(
     output = self.extract_feat(batch_inputs)
     if self.head is not None:
         output = self.head(output)
+
+    from mmcls.models.heads import MultiLabelClsHead
+    if isinstance(self.head, MultiLabelClsHead):
+        output = torch.sigmoid(output)
+    else:
+        output = F.softmax(output, dim=1)
     return output

tests/test_codebase/test_mmcls/test_mmcls_models.py

@@ -9,6 +9,10 @@
 from mmdeploy.utils import Backend, Codebase
 from mmdeploy.utils.test import WrapModel, check_backend, get_rewrite_outputs
 
+try:
+    from torch.testing import assert_close as torch_assert_close
+except Exception:
+    from torch.testing import assert_allclose as torch_assert_close
 try:
     import_codebase(Codebase.MMCLS)
 except ImportError:
@@ -77,6 +81,7 @@ def __init__(self, backbone):
         def extract_feat(self, batch_inputs: torch.Tensor):
             return batch_inputs
 
+    input = torch.rand(1, 1000)
     backbone_cfg = dict(
         type='ResNet',
         depth=18,
@@ -90,8 +95,8 @@ def extract_feat(self, batch_inputs: torch.Tensor):
     with RewriterContext({}):
         backend_output = model(input)
 
-    assert model_output == input
-    assert backend_output == input
+    torch_assert_close(model_output, input)
+    torch_assert_close(backend_output, torch.nn.functional.softmax(input, -1))
 
 
 @pytest.mark.parametrize(