ecfc31e21d0c30b1f543aacb8619582b60a1bef7,gluoncv/model_zoo/mask_rcnn/rcnn_target.py,MaskTargetGenerator,hybrid_forward,#MaskTargetGenerator#Any#Any#Any#Any#Any#,30

Before Change

                                             self._mask_size, 1.0, sample_ratio=2)
            pooled_mask = pooled_mask.reshape((-3, 0, 0))
            // duplicate to C * (N, MS, MS)
            mask_target = []
            mask_mask = []
            for cid in range(1, self._num_classes + 1):
                // boolean array (N,) -> (N, 1, 1)
                same_cid = (cls_target == cid).reshape((-1, 1, 1))
                // keep orig targets
                mask_target.append(pooled_mask)
                // but mask out the one not belong to this class [N, MS, MS]
                mask_mask.append(F.broadcast_mul(F.ones_like(pooled_mask), same_cid))
            // (C, N, MS, MS) -> (N, C, MS, MS)
            mask_targets.append(F.stack(*mask_target, axis=0).transpose((1, 0, 2, 3)))
            mask_masks.append(F.stack(*mask_mask, axis=0).transpose((1, 0, 2, 3)))

        // B * (N, C, MS, MS) -> (B, N, C, MS, MS)

After Change

            // (N,) -> (C, 1) -> (C, N, 1, 1)
            cls_target = F.expand_dims(cls_target, 0)
            same_cids = F.broadcast_equal(cls_target, cids)
            same_cids = same_cids.reshape((-2, 1, 1))

            // (N, MS, MS) -> (C, N, 1, 1) -> (C, N, MS, MS)
            mask_mask = F.broadcast_like(same_cids, pooled_mask, lhs_axes=(2, 3), rhs_axes=(1, 2))

            // (N, MS, MS) -> (C, N, MS, MS)
            mask_target = F.expand_dims(pooled_mask, 0)
            mask_target = F.broadcast_axis(mask_target, size=self._num_classes, axis=0)

Instances