[opencode-agent]

Autograd robustness improvements completed across core engine:

neutro/autograd/tensor.py:18,33,37,42 - as_tensor handles numpy scalars robustly; Tensor.__init__ validates None, casts to float64 with minimal copy and contiguous layout; __hash__ uses identity to work with overridden __eq__ (array return); __bool__ raises ValueError for multi-element tensors (previously ambiguous bool(np.array) error), allows scalar tensors; __setitem__ clears stale grad; added dtype, __len__, __iter__ guards.

neutro/autograd/tape.py:6,10,15,22,34 - Stack safety: __exit__ validates top-of-stack and removes mismatched tape; watch accepts iterables and type-checks; _record_op defensive copy; gradient validates target/sources, normalizes single tensor, handles missing ops, wraps backward exceptions, pads/truncates mismatched grad lengths (preventing silent zip truncation), uses out-of-place grad = grad+ig to avoid aliasing, handles batched accumulation for shared inputs and broadcasted grads.

neutro/autograd/function.py:7,15 - _Ctx.save_for_backward isolated; Function.apply separates tensor positional and Tensor kwargs, tracks all_tensor_inputs, converts output via np.asarray, checks tape._watched via any, copies saved_data without mutating original ctx (converts remaining Tensor values to data), closure captures lists to avoid late binding, bw handles None returns, length mismatches (pads kwargs grads), calls broadcast_backward per input, returns list aligned with recorded inputs.

neutro/autograd/utils.py:4 - broadcast_backward handles scalar original shape, leading broadcast dims, keepdims expansion via building expanded shape and broadcast_to, raises informative error on mismatch; numeric_grad ensures scalar sum and correct dtype; gradclose coerces to array.

neutro/autograd/ops.py:6,40,81,95,112,134,147,158,171,182,207,220,236,247,258,269,290,303,318,330,340,351,356 - Extracted _normalize_axis/_expand_grad_to_shape helpers; _Add/_Sub removed unnecessary save_for_backward; _Pow handles k==0; _Matmul handles vector-vector g*b, vector-matrix batches via outer/matmul and generic _swap_last_two for batched ND; _Sum/_Mean/_Max use _expand_grad_to_shape supporting tuple axes, negative axes, keepdims True/False (fixes np.expand_dims tuple bug and np.ones memory waste); _Transpose validates permutation and normalizes negative axes; _Slice saves only shape/dtype/idx, uses np.add.at for duplicate fancy indices; _Relu saves bool mask; _Sigmoid saves view; _Tanh saves output t; _Sqrt saves out and clamps; _Exp saves out; _Log clamps; _Max divides ties equally and expands grad via helper; _Concatenate normalizes axis; _Tile robust to reps length mismatch (pads eff_x/reps_padded, splits iteratively, reshapes to original); _Clip handles scalar bounds; _Abs uses np.sign (0 at 0, fixes +1e-15 sign flip); _Repeat normalizes axis; _Maximum splits ties equally; added validation for reshape, tile reps, clip bounds, repeat repeats.

Verified with tests/autograd (46 tests) and full suite (358+53 tests) plus manual gradchecks for tuple-axis sum/mean, negative axes, batched matmul, tile with mismatched reps, duplicate slice indices, abs at 0, clip, max, maximum ties, var, nested tapes.

Closes #15

opencode session  |  github run

[sourcepirate]

/oc make it more readable. currently it looks code is very unreadable and convoluted.

[opencode-agent]

Readability refactor completed while preserving hardened semantics:

neutro/autograd/utils.py:1 – Added module docstring, simplified broadcast_backward (leading-dim + keepdims loop with clear error), concise numeric_grad/gradclose.

neutro/autograd/tensor.py:1 –Grouped into sections (conversion / creation / magic / properties / ops), identity __hash__, __bool__ guard, sorted comparison helpers, centralized float64 handling.

neutro/autograd/tape.py:1 – Extracted _normalize_sources, _align_grad_lengths, _accumulate_grads, _collect_source_grads; cleaned __exit__ stack handling and watch iterable support; removed inline defensive-copy noise.

neutro/autograd/function.py:1 – Split into _split_tensor_args, _split_tensor_kwargs, _build_saved_copy, _make_backward_fn, _align_backward_grads; removed late-binding closure clutter and dense comments.

neutro/autograd/ops.py:1 – Sectioned helpers/arithmetic/matmul/reductions/shape/unary; documented _normalize_axis/_expand_grad_to_shape/_swap_last_two; rewrote _Matmul via _matmul_vec_mat_left/_matmul_mat_vec_right (loop-based batched handling, removed if False dead code); simplified _Tile iterative summing, _Max/_Maximum tie handling, and removed redundant copies/verbose inline comments.

Verified: tests/autograd 46 passed and full suite 411 passed; manual gradchecks for tuple-axis/negative-axis sum, batched matmul vector cases, mismatched reps tile, duplicate slice indices, abs/clip/max/maximum ties, var, nested tapes all pass.

opencode session  |  github run