tinymind-native-8b-remote-handoff/bundle/model/tinymind-apex/neural_core.py

from __future__ import annotations

import argparse
import json
from pathlib import Path

import torch
import torch.nn as nn


ROOT = Path(r"D:\ad\tinymind\model\tinymind-apex")


class NeuralCore(nn.Module):
    def __init__(self, vocab_size: int, domain_count: int, tool_count: int, emb: int = 96, hidden: int = 192):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, emb, padding_idx=0)
        self.conv3 = nn.Conv1d(emb, hidden, kernel_size=3, padding=1)
        self.conv5 = nn.Conv1d(emb, hidden, kernel_size=5, padding=2)
        self.norm = nn.LayerNorm(hidden * 2)
        self.dropout = nn.Dropout(0.0)
        self.domain_head = nn.Linear(hidden * 2, domain_count)
        self.tool_head = nn.Linear(hidden * 2, tool_count)
        self.risk_head = nn.Linear(hidden * 2, 1)

    def encode(self, x):
        emb = self.embedding(x).transpose(1, 2)
        h3 = torch.relu(self.conv3(emb)).amax(dim=2)
        h5 = torch.relu(self.conv5(emb)).amax(dim=2)
        h = torch.cat([h3, h5], dim=1)
        return self.dropout(self.norm(h))

    def forward(self, x):
        h = self.encode(x)
        return {
            "domain": self.domain_head(h),
            "tool": self.tool_head(h),
            "risk": self.risk_head(h).squeeze(-1),
        }


class TinyMindNeuralCore:
    def __init__(self, root: Path = ROOT):
        ckpt = torch.load(root / "artifacts" / "neural_core" / "tinymind_neural_core.pt", map_location="cpu", weights_only=False)
        self.vocab = ckpt["vocab"]
        self.domains = ckpt["domains"]
        self.tools = ckpt["tools"]
        self.id_to_domain = {v: k for k, v in self.domains.items()}
        self.id_to_tool = {v: k for k, v in self.tools.items()}
        self.max_len = ckpt["max_len"]
        cfg = ckpt["config"]
        self.model = NeuralCore(len(self.vocab), len(self.domains), len(self.tools), cfg["emb"], cfg["hidden"])
        self.model.load_state_dict(ckpt["state_dict"])
        self.model.eval()

    def encode(self, text: str):
        ids = [self.vocab.get(ch, 1) for ch in text[: self.max_len]]
        ids += [0] * (self.max_len - len(ids))
        return torch.tensor([ids], dtype=torch.long)

    def predict(self, text: str):
        with torch.no_grad():
            out = self.model(self.encode(text))
            domain_prob = torch.softmax(out["domain"], dim=1)[0]
            tool_prob = torch.softmax(out["tool"], dim=1)[0]
            risk_prob = torch.sigmoid(out["risk"])[0].item()
        domain_i = int(domain_prob.argmax().item())
        tool_i = int(tool_prob.argmax().item())
        return {
            "domain": {"label": self.id_to_domain[domain_i], "confidence": round(float(domain_prob[domain_i]), 6)},
            "tool_policy": {"label": self.id_to_tool[tool_i], "confidence": round(float(tool_prob[tool_i]), 6)},
            "high_risk_probability": round(float(risk_prob), 6),
            "model": "tinymind-neural-core",
        }


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("text")
    args = parser.parse_args()
    core = TinyMindNeuralCore()
    print(json.dumps(core.predict(args.text), ensure_ascii=False, indent=2))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())