[BugFix & Feature]

1.Fix set_max_subnet or set_min_subnet not found exception
2 Add the fleibility to define custom subnet kinds
3.Support to specifiy None with ResorceEstimator

Signed-off-by: Ming-Hsuan-Tu <qrnnis2623891@gmail.com>

Author: twmht

mmrazor/engine/runner/subnet_val_loop.py

@@ -21,8 +21,8 @@ class SubnetValLoop(ValLoop, CalibrateBNMixin):
         evaluator (Evaluator or dict or list): Used for computing metrics.
         fp16 (bool): Whether to enable fp16 validation. Defaults to
             False.
-        evaluate_fixed_subnet (bool): Whether to evaluate a fixed subnet only
-            or not. Defaults to False.
+        fix_subnet_kind (str): fix subnet kinds when evaluate, this would be
+            `sample_kinds` if not specified
         calibrate_sample_num (int): The number of images to compute the true
             average of per-batch mean/variance instead of the running average.
             Defaults to 4096.
@@ -36,7 +36,7 @@ def __init__(
         dataloader: Union[DataLoader, Dict],
         evaluator: Union[Evaluator, Dict, List],
         fp16: bool = False,
-        evaluate_fixed_subnet: bool = False,
+        fix_subnet_kinds: List[str] = [],
         calibrate_sample_num: int = 4096,
         estimator_cfg: Optional[Dict] = dict(type='mmrazor.ResourceEstimator')
     ) -> None:
@@ -48,9 +48,18 @@ def __init__(
             model = self.runner.model
 
         self.model = model
-        self.evaluate_fixed_subnet = evaluate_fixed_subnet
+        if fix_subnet_kinds is None and not hasattr(self.model,
+                                                    'sample_kinds'):
+            raise ValueError(
+                'neither fix_subnet_kinds nor self.model.sample_kinds exists')
+
+        self.evaluate_kinds = fix_subnet_kinds if len(
+            fix_subnet_kinds) > 0 else getattr(self.model, 'sample_kinds')
+
         self.calibrate_sample_num = calibrate_sample_num
-        self.estimator = TASK_UTILS.build(estimator_cfg)
+        self.estimator = None
+        if estimator_cfg:
+            self.estimator = TASK_UTILS.build(estimator_cfg)
 
     def run(self):
         """Launch validation."""
@@ -59,24 +68,19 @@ def run(self):
 
         all_metrics = dict()
 
-        if self.evaluate_fixed_subnet:
+        for kind in self.evaluate_kinds:
+            if kind == 'max':
+                self.model.mutator.set_max_choices()
+            elif kind == 'min':
+                self.model.mutator.set_min_choices()
+            elif 'random' in kind:
+                self.model.mutator.set_choices(
+                    self.model.mutator.sample_choices())
+            else:
+                raise NotImplementedError(f'Unsupported Subnet {kind}')
+
             metrics = self._evaluate_once()
-            all_metrics.update(add_prefix(metrics, 'fix_subnet'))
-        elif hasattr(self.model, 'sample_kinds'):
-            for kind in self.model.sample_kinds:
-                if kind == 'max':
-                    self.model.mutator.set_max_choices()
-                    metrics = self._evaluate_once()
-                    all_metrics.update(add_prefix(metrics, 'max_subnet'))
-                elif kind == 'min':
-                    self.model.mutator.set_min_choices()
-                    metrics = self._evaluate_once()
-                    all_metrics.update(add_prefix(metrics, 'min_subnet'))
-                elif 'random' in kind:
-                    self.model.mutator.set_choices(
-                        self.model.mutator.sample_choices())
-                    metrics = self._evaluate_once()
-                    all_metrics.update(add_prefix(metrics, f'{kind}_subnet'))
+            all_metrics.update(add_prefix(metrics, f'{kind}_subnet'))
 
         self.runner.call_hook('after_val_epoch', metrics=all_metrics)
         self.runner.call_hook('after_val')
@@ -90,7 +94,8 @@ def _evaluate_once(self) -> Dict:
             self.run_iter(idx, data_batch)
 
         metrics = self.evaluator.evaluate(len(self.dataloader.dataset))
-        resource_metrics = self.estimator.estimate(self.model)
-        metrics.update(resource_metrics)
+        if self.estimator:
+            resource_metrics = self.estimator.estimate(self.model)
+            metrics.update(resource_metrics)
 
         return metrics

tests/test_engine/test_runner/test_subnet_val_loop.py

@@ -0,0 +1,50 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from unittest import TestCase
+from unittest.mock import MagicMock, call, patch
+
+from mmrazor.engine.runner import SubnetValLoop
+
+class TestSubnetValLoop(TestCase):
+    def test_subnet_val_loop():
+        runner = MagicMock()
+        runner.distributed = False
+        runner.model = MagicMock()
+        dataloader = MagicMock()
+        evaluator = [MagicMock()]
+        fix_subnet_kinds = ['max', 'min']
+        loop = SubnetValLoop(
+            runner,
+            dataloader,
+            evaluator,
+            fix_subnet_kinds=fix_subnet_kinds,
+            estimator_cfg=None)
+        runner.train_dataloader = MagicMock()
+        with patch.object(loop, '_evaluate_once') as evaluate_mock:
+            evaluate_mock.return_value = dict(acc=10)
+            all_metrics = dict()
+            all_metrics['max_subnet.acc'] = 10
+            all_metrics['min_subnet.acc'] = 10
+            loop.run()
+            loop.runner.call_hook.assert_has_calls([
+                call('before_val'),
+                call('before_val_epoch'),
+                call('after_val_epoch', metrics=all_metrics),
+                call('after_val')
+            ])
+            evaluate_mock.assert_has_calls([call(), call()])
+
+        runner.dataloader = MagicMock()
+        runner.dataloader.dataset = MagicMock()
+        loop.dataloader.__iter__.return_value = ['data_batch1']
+        with patch.object(loop, 'calibrate_bn_statistics') as calibration_bn_mock:
+            with patch.object(loop, 'run_iter') as run_iter_mock:
+                eval_result = dict(acc=10)
+                loop.evaluator.evaluate.return_value = eval_result
+                result = loop._evaluate_once()
+                calibration_bn_mock.assert_called_with(runner.train_dataloader,
+                                                       loop.calibrate_sample_num)
+                runner.model.eval.assert_called()
+                run_iter_mock.assert_called_with(0, 'data_batch1')
+                loop.evaluator.evaluate.assert_called_with(
+                    len(runner.dataloader.dataset))
+                assert result == eval_result