dcn layer error

[as754770178]

I run faster_rcnn_x101_32x4d_fpn_dconv_c3-c5_1x_coco by mmdet:2.4.0 and mmcv:1.1.3, report error:

  File "/usr/local/lib/python3.6/site-packages/torch/autograd/function.py", line 77, in apply
    return self._forward_cls.backward(self, *args)
  File "/usr/local/lib/python3.6/site-packages/torch/autograd/function.py", line 189, in wrapper
    outputs = fn(ctx, *args)
  File "/usr/local/lib/python3.6/site-packages/mmcv/ops/deform_conv.py", line 126, in backward
    im2col_step=cur_im2col_step)
RuntimeError: shape '[1024, 1024, 3, 3]' is invalid for input of size 294912

I change to mmdet:2.3.0 and mmcv:1.0.5, not report error.

[ZwwWayne]

This is a bug introduced in open-mmlab/mmcv#565. Thanks for your bug report, we will fix it ASAP.

[Charlyo]

@yhcao6 Reopening this since it still happens in mmcv-full 1.3.9 and mmdet 2.15.0

/venv/lib/python3.8/site-packages/mmcv/ops/deform_conv.py", line 127, in backward ext_module.deform_conv_backward_input( RuntimeError: shape '[1, 1, 2, 256, 3]' is invalid for input of size 4608