segformer preprocessing problems

[concrete13377]

python easyportrait\pipelines\demo\image_demo.py "images\2.jpg" "easyportrait\pipelines\local_configs\easyportrait_experiments_v2\segformer-fp\segformer-fp.py" "easyportrait\ckpt\faceparse_segformer-fp-512.pth" --palette=easy_portrait --out-file="result.png"

I got the following error:

mmsegmentation\mmseg\datasets\pipelines\formatting.py", line 281, in __call__
img_meta[key] = results[key]
KeyError: pad_shape

torch==2.1.1
mmcv-full==1.7.1
mmsegmentation==0.30.0
mmcls==0.25.0

I dont have your torch version, but the problem does not seems to be in torch, looks like the config is wrong

[concrete13377]

the only thing that helped is to replace this line with img_meta[key] = results.get(key, None)

[concrete13377]

very good results with segformer-1024 model, thanks for your work :)

[CkShibit]

python ./pipelines/demo/image_demo.py "/root/workspace/facefusion/input_source/mmexport1705161230807.jpg" "/root/workspace/easyportrait/pipelines/local_configs/easyportrait_experiments_v2/segformer-fp/segformer-fp.py" "/root/workspace/easyportrait/pipelines/models/segformer-fp-512.pth" --palette=easy_portrait --out-file="ttt.png"
/root/miniconda3/envs/easyportrait/lib/python3.8/site-packages/mmcv/init.py:20: UserWarning: On January 1, 2023, MMCV will release v2.0.0, in which it will remove components related to the training process and add a data transformation module. In addition, it will rename the package names mmcv to mmcv-lite and mmcv-full to mmcv. See https://github.com/open-mmlab/mmcv/blob/master/docs/en/compatibility.md for more details.
warnings.warn(
/root/miniconda3/envs/easyportrait/lib/python3.8/site-packages/mmseg/models/losses/cross_entropy_loss.py:235: UserWarning: Default avg_non_ignore is False, if you would like to ignore the certain label and average loss over non-ignore labels, which is the same with PyTorch official cross_entropy, set avg_non_ignore=True.
warnings.warn(
load checkpoint from local path: /root/workspace/easyportrait/pipelines/models/segformer-fp-512.pth
Traceback (most recent call last):
File "./pipelines/demo/image_demo.py", line 46, in
main()
File "./pipelines/demo/image_demo.py", line 40, in main
get_palette(args.palette),
File "/root/miniconda3/envs/easyportrait/lib/python3.8/site-packages/mmseg/core/evaluation/class_names.py", line 324, in get_palette
raise ValueError(f'Unrecognized dataset: {dataset}')
ValueError: Unrecognized dataset: easy_portrait
How to soulve the error??

[hradec]

try to add the pipelines folder to the top of your PYTHONPATH:

export PYTHONPATH=./pipelines/:$PYTHONPATH

and run again.
This depot has a copy of the mmseg module in the pipelines folder, where they edited the class_names.py to add the easy_portrait palette.
So you have to force python to load the copy version of mmseg module from the pipelines folder, by adding it to the top of PYTHONPATH.

[Refl-ex]

the only thing that helped is to replace this line with img_meta[key] = results.get(key, None)

You have to manually add padding to test pipelines inside a local config for your own images. This may be because original test/val datasets were preprocessed before to match input shape. It works with additional resize.

pipeline=[
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(512, 512),
        flip=False,
        transforms=[
            dict(type='Resize', img_scale=(512, 512), keep_ratio=True),
            dict(type='Pad', size=(512, 512), pad_val=0),
            dict(
                type='Normalize',
                mean=[143.55267075, 132.96705975, 126.94924335],
                std=[60.2625333, 60.32740275, 59.30988645],
                to_rgb=True),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img'])
        ])
]