[Fix] Fixed device_id in tools/test.py for the CPU. (open-mmlab#58)

* [Fix] fix bugs for mmcls performance test (open-mmlab#269)

* fix bugs for mmcls performance test

* fix yapf

* add comments of CLASSES attribute

* Rewrote the dictionary traversal for new versions of Python.

* Fix device_id for cpu

* Rewrite parse_device_id and tests

* Added None for cpu

Co-authored-by: hanrui1sensetime <[email protected]>

mmdeploy/backend/openvino/wrapper.py

@@ -107,11 +107,14 @@ def __process_outputs(
             name: torch.from_numpy(tensor)
             for name, tensor in outputs.items()
         }
-        for output_name in outputs.keys():
-            if '.' in output_name:
-                new_output_name = output_name.split('.')[0]
-                outputs[new_output_name] = outputs.pop(output_name)
-        return outputs
+        cleaned_outputs = {}
+        for name, value in outputs.items():
+            if '.' in name:
+                new_output_name = name.split('.')[0]
+                cleaned_outputs[new_output_name] = value
+            else:
+                cleaned_outputs[name] = value
+        return cleaned_outputs
 
     def forward(self, inputs: Dict[str,
                                    torch.Tensor]) -> Dict[str, torch.Tensor]:

mmdeploy/utils/device.py

@@ -1,23 +1,27 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+from typing import Optional
+
 import torch
 
 
-def parse_device_id(device: str) -> int:
-    """Parse cuda device index from a string.
+def parse_device_id(device: str) -> Optional[int]:
+    """Parse device index from a string.
 
     Args:
-        device (str): The typical style of string specifying cuda device,
-            e.g.: 'cuda:0'.
+        device (str): The typical style of string specifying device,
+            e.g.: 'cuda:0', 'cpu'.
 
     Returns:
-        int: The parsed device id, defaults to `0`.
+        Optional[int]: The return value depends on the type of device.
+            If device is 'cuda': cuda device index, defaults to `0`.
+            If device is 'cpu': `-1`.
+            Otherwise, `None` will be returned.
     """
     if device == 'cpu':
         return -1
-    device_id = 0
-    if len(device) >= 6:
-        device_id = torch.device(device).index
-    return device_id
+    if 'cuda' in device:
+        return parse_cuda_device_id(device)
+    return None
 
 
 def parse_cuda_device_id(device: str) -> int:

tests/test_utils/test_util.py

@@ -369,6 +369,30 @@ def test_can_get_axes_from_list_with_cfg(self):
         assert expected_dynamic_axes == dynamic_axes
 
 
+class TestParseDeviceID:
+
+    def test_cpu(self):
+        device = 'cpu'
+        assert util.parse_device_id(device) == -1
+
+    def test_cuda(self):
+        device = 'cuda'
+        assert util.parse_device_id(device) == 0
+
+    def test_cuda10(self):
+        device = 'cuda:10'
+        assert util.parse_device_id(device) == 10
+
+    def test_incorrect_cuda_device(self):
+        device = 'cuda_5'
+        with pytest.raises(RuntimeError):
+            util.parse_device_id(device)
+
+    def test_incorrect_device(self):
+        device = 'abcd:1'
+        assert util.parse_device_id(device) is None
+
+
 def test_AdvancedEnum():
     keys = [
         Task.TEXT_DETECTION, Task.TEXT_RECOGNITION, Task.SEGMENTATION,

tools/test.py

@@ -105,7 +105,8 @@ def main():
     # load the model of the backend
     model = task_processor.init_backend_model(args.model)
 
-    device_id = parse_device_id(args.device)
+    is_device_cpu = (args.device == 'cpu')
+    device_id = None if is_device_cpu else parse_device_id(args.device)
 
     model = MMDataParallel(model, device_ids=[device_id])
     # The whole dataset test wrapped a MMDataParallel class outside the module.
@@ -115,7 +116,7 @@ def main():
     if hasattr(model.module, 'CLASSES'):
         model.CLASSES = model.module.CLASSES
     if args.speed_test:
-        with_sync = device_id >= 0
+        with_sync = not is_device_cpu
         output_file = sys.stdout
         if args.log2file:
             output_file = args.log2file