| """Numpy-only neural inference; no semantic phrase rules. Fixed feature hashing.""" |
|
|
| from importlib.resources import files |
|
|
| from core import Diagnostic, assemble, compile_js, frontend |
|
|
| |
| |
| import numpy as np |
|
|
|
|
| def default_checkpoint(): |
| return files("tiny_codemode_weights").joinpath("model.npz") |
|
|
|
|
| class Model: |
| def __init__(self, path=None): |
| with np.load(path if path is not None else default_checkpoint()) as z: |
| self.w = {k: z[k] for k in z.files} |
|
|
| def logits(self, item): |
| _, ids, f, _ = frontend(item) |
| return self.forward(ids, f) |
|
|
| def forward(self, ids, f): |
| w = self.w |
| dims = w["convs.0.weight"].shape[1] - w["embed.weight"].shape[1] |
| h = np.concatenate( |
| [w["embed.weight"][ids % len(w["embed.weight"])], f[:, :dims]], axis=1 |
| ) |
| for i in range(3): |
| weight = w[f"convs.{i}.weight"] |
| bias = w[f"convs.{i}.bias"] |
| pad = weight.shape[2] // 2 |
| p = np.pad(h, ((pad, pad), (0, 0))) |
| h = np.maximum( |
| sum( |
| p[k : k + len(h)] @ weight[:, :, k].T |
| for k in range(weight.shape[2]) |
| ) |
| + bias, |
| 0, |
| ) |
| pooled = np.concatenate([h.mean(0), h.max(0)]) |
| heads = ( |
| h @ w["roles.weight"].T + w["roles.bias"], |
| pooled @ w["mode.weight"].T + w["mode.bias"], |
| pooled @ w["output.weight"].T + w["output.bias"], |
| ) |
| return ( |
| heads + (pooled @ w["reject.weight"].T + w["reject.bias"],) |
| if "reject.weight" in w |
| else heads |
| ) |
|
|
| def predict(self, item): |
| heads = self.logits(item) |
| if len(heads) == 4 and int(heads[3].argmax()) == 1: |
| raise Diagnostic("learned rejection: unsupported/invalid request") |
| r, m, o = heads[:3] |
| return assemble(item, r.argmax(1), int(m.argmax()), int(o.argmax())) |
|
|
| def program(self, item): |
| return compile_js(self.predict(item), item["tools"]) |
|
|
