🧠 D: DOCUMENT — Problem & Opportunity

TransposedConv2d (skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/nn/TransposedConv2d.kt) has a full constructor, parameter validation, outputSize() and factory helpers, but forward cannot run: there is no 2D transposed-convolution op. TensorOps.convTranspose2d now exists as a placeholder whose default throws, marked @NotImplemented, so the gap is visible as ❌ in the operator coverage matrix (both in the TensorOps table and the NN-layers table).

Transposed convolution is the learned-upsampling block of GAN generators, U-Net / segmentation decoders and super-resolution nets, so this gap blocks a whole family of models. The 1D variant is implemented end to end (TransposedConv1d, DefaultCpuOps.convTranspose1d, convTranspose1dBackward), so this is mostly a careful extension to a second spatial axis.

Summary: implement convTranspose2d on the CPU backend with PyTorch-identical semantics, wire autodiff and StableHLO, and turn the ❌ rows green.

🔍 A: ASSESS — Feasibility & Impact

• Feasibility: high. Templates exist for every piece: DefaultCpuOps.convTranspose1d (scatter-add over out = in*stride - padding + k*dilation), DefaultCpuOps.conv2d (2D validation, bias, groups), and the input-gradient loop of conv2dGrads in DefaultExecutionTape.kt, which is mathematically a transposed conv2d already.
• Semantics to match: PyTorch torch.nn.functional.conv_transpose2d: weight [C_in, C_out/groups, kH, kW], output size (in-1)*stride - 2*padding + dilation*(k-1) + outputPadding + 1, outputPadding < stride.
• Risks: dtype — convTranspose1d casts to Float everywhere (FP32 only); the 2D kernel should at least not silently corrupt other dtypes. Promoting the op to an abstract @Diff interface method touches every TensorOps implementer and test fake, so it is its own lane.

📚 R: RESEARCH — What Must Be Understood First?

• PyTorch reference cases for the parameter grid (Lane 1)
• StableHLO lowering: stablehlo.convolution with lhs_dilation = stride, flipped kernel, swapped feature dims (Lane 2c)

🛠️ C: CODE — Implementation Plan

• CPU kernel (Lane 2a) — makes TransposedConv2d.forward work on CPU
• Promote to @Diff op + backward (Lane 2b)
• StableHLO converter (Lane 2c)
• Ground-truth CI (Lane 4)
• Docs partial (Lane 5)
• Review (Lane 6)

Acceptance Criteria

• TransposedConv2d.forward matches PyTorch on the Lane-1 cases
• @NotImplemented removed from TensorOps.convTranspose2d and TransposedConv2d; the regenerated matrix shows ✅ for cpu, StableHLO and Backward

Lane breakdown

• [Lane 1 · numerics] TransposedConv2d — PyTorch reference cases for conv_transpose2d #1372 — [Lane 1 · numerics] PyTorch reference cases for conv_transpose2d (size:s, good first issue)
• [Lane 2 · kotlin-core] TransposedConv2d — CPU kernel for convTranspose2d #1373 — [Lane 2 · kotlin-core] CPU kernel for convTranspose2d (size:m)
• [Lane 2 · kotlin-core] TransposedConv2d — promote convTranspose2d to a @Diff op with backward rule #1374 — [Lane 2 · kotlin-core] promote convTranspose2d to a @Diff op with backward rule (size:s)
• [Lane 2 · kotlin-core] TransposedConv2d — StableHLO lowering for convTranspose2d #1375 — [Lane 2 · kotlin-core] StableHLO lowering for convTranspose2d (size:s)
• [Lane 4 · kotlin-core] TransposedConv2d — ground-truth test for convTranspose2d #1376 — [Lane 4 · kotlin-core] ground-truth test for convTranspose2d (size:s)
• [Lane 5 · docs] TransposedConv2d — operator doc partial for convTranspose2d #1377 — [Lane 5 · docs] operator doc partial for convTranspose2d (size:s, good first issue)
• [Lane 6 · review] TransposedConv2d — DARC review of convTranspose2d #1378 — [Lane 6 · review] DARC review of convTranspose2d (size:s)

Lanes 0 (no SKEEP trigger — additive op behind an existing interface) and 3 (no platform-specific code) are skipped.

New here? Read Getting started as a contributor and the Issue taxonomy. Lanes marked good first issue assume no prior SKaiNET codebase knowledge. The gap markers and the coverage matrix were introduced in #1370.