Created by: LordAaron

paddle官方api文档里面写了deformable conv的两种用法，一种可以用mask，一种可以不用mask；但是我将mask设置为None会报错： 代码： offset = fluid.layers.conv2d(input=wh_pred, num_filters=29n_anchors, groups=n_anchors, filter_size=1, stride=1, padding=0) offset.stop_gradient = True out = fluid.layers.deformable_conv(input=input, offset=offset, mask=None, num_filters=out_ch, filter_size=3, stride=1,
padding=1, deformable_groups=n_anchors)

如果使用mask 代码： offset_mask = fluid.layers.conv2d(input=wh_pred, num_filters=39n_anchors, groups=n_anchors, filter_size=1, stride=1, padding=0) offset_mask.stop_gradient = True

offset_channel = 9n_anchors * 2 mask_channel = 9n_anchors

offset, mask = fluid.layers.split( input=offset_mask, num_or_sections=[offset_channel, mask_channel], dim=1) mask = fluid.layers.sigmoid(mask)

out = fluid.layers.deformable_conv(input=input, offset=offset, mask=mask, num_filters=out_ch, filter_size=3, stride=1,
padding=1, deformable_groups=n_anchors) 在训练时，就会报下面的错误：