//clika-runtime/io.clika.runtime/Ops/deformConv

deformConv

[common]
fun deformConv(input: Tensor, weight: Tensor, offset: Tensor, mask: Tensor? = null, bias: Tensor? = null, stride: LongArray = longArrayOf(), padding: LongArray = longArrayOf(), dilation: LongArray = longArrayOf(), groups: Long = 1, offsetGroups: Long = 1, activation: Activation = Activation.IDENTITY): Tensor

deformConv(input: Tensor, weight: Tensor, offset: Tensor, mask: Tensor? = null, bias: Tensor? = null, stride: LongArray = longArrayOf(), padding: LongArray = longArrayOf(), dilation: LongArray = longArrayOf(), groups: Long = 1L, offsetGroups: Long = 1L, activation: Activation = Activation.IDENTITY): the deform_conv operator. Deformable convolution, 1-D/2-D/3-D (rank derives from input), channels-last: x [N, D1..Dr, C], weight [O, K1..Kr, C/groups] (OHWI), offset [N, out-spatial..., offset_groups·∏K·r], mask [N, out-spatial..., offset_groups·∏K] (absent ⇒ unmodulated), bias [O]; returns [N, out-spatial..., O] at input's dtype. Each kernel tap samples at out·stride − pad_lo + tap·dilation + Δ (pixel units, bilinear; out-of-bounds reads 0); the offset channel for (group g, tap t, axis d) is (g·∏K + t)·r + d with taps row-major over the kernel and axes in layout order (2-D: Δh then Δw). padding is interleaved (lo, hi) pairs over the spatial axes; stride/dilation broadcast per spatial axis (empty ⇒ 1). All floating inputs must share input's dtype (f32/f64/f16/bf16; no silent promotion); activation is a fused elementwise epilogue applied after bias.