no attribute 'test_pipeline'

[scott-vsi]

I'm trying to run mmseg.apis.inference_model with the "Potsdam(i)2RescueNet.pth" model and am getting AttributeError: 'ConfigDict' object has no attribute 'test_pipeline'. (It looks like the CrossEarth_dinov2_mask2former_512x512_bs1x4.py config file is setup to run this model since it refers to the dg_potsdamRGB2rescue_512x512.py config file, so I didn't modify it any.

from mmseg.apis import init_model, inference_model

config_path = '/CrossEarth/configs/CrossEarth_dinov2/CrossEarth_dinov2_mask2former_512x512_bs1x4.py'
checkpoint_path = '/CrossEarth/checkpoints/Potsdam(i)2RescueNet.pth' # EDITED

model = init_model(config_path, checkpoint_path) # this does give size mismatch warnings, but I saw you said those were ok
result = inference_model(model, img_path)
# AttributeError: 'ConfigDict' object has no attribute 'test_pipeline'

Details

File /venv/lib/python3.10/site-packages/mmseg/apis/inference.py:143, in inference_model(model, img)
    130 """Inference image(s) with the segmentor.
    131 
    132 Args:
   (...)
    140     will be returned, otherwise return the segmentation results directly.
    141 """
    142 # prepare data
--> 143 data, is_batch = _preprare_data(img, model)
    145 # forward the model
    146 with torch.no_grad():

File /venv/lib/python3.10/site-packages/mmseg/apis/inference.py:99, in _preprare_data(imgs, model)
     96 def _preprare_data(imgs: ImageType, model: BaseSegmentor):
     98     cfg = model.cfg
---> 99     for t in cfg.test_pipeline:
    100         if t.get('type') == 'LoadAnnotations':
    101             cfg.test_pipeline.remove(t)

File /venv/lib/python3.10/site-packages/mmengine/config/config.py:1500, in Config.__getattr__(self, name)
   1499 def __getattr__(self, name: str) -> Any:
-> 1500     return getattr(self._cfg_dict, name)

File /venv/lib/python3.10/site-packages/mmengine/config/config.py:113, in ConfigDict.__getattr__(self, name)
    111         value = value.build()
    112 except KeyError:
--> 113     raise AttributeError(f"'{self.__class__.__name__}' object has no "
    114                          f"attribute '{name}'")
    115 except Exception as e:
    116     raise e

AttributeError: 'ConfigDict' object has no attribute 'test_pipeline'

python 3.10
torch==2.0.1
torchvision==0.15.2
mmengine= 0.10.7
mmcv==2.0.1
mmdet==3.0.0
mmsegmentation==1.0.0
xformers==0.0.20

[scott-vsi]

I tried to use the the Potsdam_Rescue_test_pipeline defined in potsdamRGB_rescue_512x512.py, which is reference by dg_potsdamRGB2rescue_512x512.py (which is reference by CrossEarth_dinov2_mask2former_512x512_bs1x4.py) by setting test_pipeline=Potsdam_Rescue_test_pipeline in that file, but that didn't work either.

[scott-vsi]

If I use

test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='PackSegInputs')
]

then I can get it to run, but the result seems to change significantly in segmentation quality between runs, and the labels seem to change their order so sometimes the labeling looks ok and sometimes it looks very bad.

[Cuzyoung]

Thanks for your mention! I was very busy a few days ago. The checkpoints have been updated several days ago. Could you provide more detailed information, such as which weight file you used and on what kind of images you applied it to?

[scott-vsi]

It seems the default model (Potsdam(i)2RescueNet.pth, dg_potsdamRGB2rescue_512x512.py) does work as expected. However, neither (WHU-Building-S2A.pth, dg_building_satellite2aerial.py) or (R2U.pth, dg_loveda_rural2urban_1024x1024.py) do.

For example,

config_path = '/CrossEarth/configs/CrossEarth_dinov2/CrossEarth_dinov2_mask2former_1024x1024_bs4x2.py'
checkpoint_path = '/CrossEarth/checkpoints/R2U.pth'

and add this to loveda_rural_1024x1024.py, which is reference by dg_loveda_rural2urban_1024x1024.py, which is reference in CrossEarth_dinov2_mask2former_1024x1024_bs4x2.py

diff --git a/configs/_base_/datasets/loveda_rural_1024x1024.py b/configs/_base_/datasets/loveda_rural_1024x1024.py
index b2b0e9b..7c06177 100644
--- a/configs/_base_/datasets/loveda_rural_1024x1024.py
+++ b/configs/_base_/datasets/loveda_rural_1024x1024.py
@@ -53,3 +53,9 @@ test_loveda_rural = dict(
     ),
     pipeline=loveda_rural_test_pipeline,
 )
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='PackSegInputs')
+]

then each time I reload the model, I get a different segmentation labeling. And a different ordering to the labels.