ppwwyyxx / groupnorm-reproduce Goto Github PK

Would it be possible for you to share pretrained weights for resnet GN? Transferring directly from the pytorch BN weights gives very bad accuracy.

Hello,

Could you provide ResNet-18 pretrained weights with GroupNorm?

Hi,

Is it possible or did you try training the resnet101 on ImageNet? Could you release the pretrained weights?

Thx

about the performance variation

using the pretrained models, how can I run detection on any image?

python imagenet-resnet-gn.py --data C:\yolo\resnet\GroupNorm-reproduce\ImageNet-ResNet-TensorFlow\image\ -d 101 --eval --load ImageNet-R101-GroupNorm32-AlignPadding.npz

In the image folder I have .png files on which I want to perform detection. But it keeps on giving the error of

File "imagenet-resnet-gn.py", line 128, in
ds = get_imagenet_dataflow(args.data, 'val', batch)
File "C:\yolo\resnet\GroupNorm-reproduce\ImageNet-ResNet-TensorFlow\imagenet_utils.py", line 95, in get_imagenet_dataflow
ds = dataset.ILSVRC12Files(datadir, name, shuffle=False)
File "C:\Users\paul90\AppData\Local\Continuum\anaconda3\envs\yolov3\lib\site-packages\tensorpack\dataflow\dataset\ilsvrc.py", line 164, in init
assert os.path.isfile(fname), fname
AssertionError: C:\yolo\resnet\GroupNorm-reproduce\ImageNet-ResNet-TensorFlow\image\val\ILSVRC2012_val_00000001.JPEG

can anyone please tell me how to run the detection on any image?

nn.moments returns variance instead of std.

Best,
Huiyu