fix mmcls get classes (open-mmlab#215)

* fix mmcls get classes

* resolve comment

* resolve comment

Author: RunningLeon

mmdeploy/codebase/mmcls/deploy/classification_model.py

@@ -4,13 +4,12 @@
 import mmcv
 import numpy as np
 import torch
-from mmcls.datasets import DATASETS
 from mmcls.models.classifiers.base import BaseClassifier
 from mmcv.utils import Registry
 
 from mmdeploy.codebase.base import BaseBackendModel
 from mmdeploy.utils import (Backend, get_backend, get_codebase_config,
-                            load_config)
+                            get_root_logger, load_config)
 
 
 def __build_backend_model(cls_name: str, registry: Registry, *args, **kwargs):
@@ -150,20 +149,35 @@ def get_classes_from_config(model_cfg: Union[str, mmcv.Config]):
     Returns:
         list[str]: A list of string specifying names of different class.
     """
-    model_cfg = load_config(model_cfg)[0]
+    from mmcls.datasets import DATASETS
+
     module_dict = DATASETS.module_dict
+    model_cfg = load_config(model_cfg)[0]
     data_cfg = model_cfg.data
 
-    if 'train' in data_cfg:
-        module = module_dict[data_cfg.train.type]
-    elif 'val' in data_cfg:
-        module = module_dict[data_cfg.val.type]
-    elif 'test' in data_cfg:
-        module = module_dict[data_cfg.test.type]
-    else:
-        raise RuntimeError(f'No dataset config found in: {model_cfg}')
-
-    return module.CLASSES
+    def _get_class_names(dataset_type: str):
+        dataset = data_cfg.get(dataset_type, None)
+        if (not dataset) or (dataset.type not in module_dict):
+            return None
+
+        module = module_dict[dataset.type]
+        if module.CLASSES is not None:
+            return module.CLASSES
+        return module.get_classes(dataset.get('classes', None))
+
+    class_names = None
+    for dataset_type in ['val', 'test', 'train']:
+        class_names = _get_class_names(dataset_type)
+        if class_names is not None:
+            break
+
+    if class_names is None:
+        logger = get_root_logger()
+        logger.warning(f'Use generated class names, because \
+            it failed to parse CLASSES from config: {data_cfg}')
+        num_classes = model_cfg.model.head.num_classes
+        class_names = [str(i) for i in range(num_classes)]
+    return class_names
 
 
 def build_classification_model(model_files: Sequence[str],