scripts/train.py

"""Train all learned components and persist them to models/:
  1. Siamese semantic encoder   (DL, contrastive loss, pure-numpy backprop)
  2. DQN investigation planner  (RL, replay buffer + target network)
  3. Plagiarism-type classifier (DL, softmax MLP over evidence features)
  4. Causal query bandit + Bayesian evidence weights (initial priors)
"""
import json
import os
import sys
import time

import numpy as np

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from plagdetect.causal import CausalQueryBandit  # noqa: E402
from plagdetect.evidence import EvidenceWeights, compare_section  # noqa: E402
from plagdetect.forensics import CLASSES, TypeClassifier  # noqa: E402
from plagdetect.ingestion import load_corpus  # noqa: E402
from plagdetect.rl_planner import train_dqn  # noqa: E402
from plagdetect.siamese import SiameseEncoder  # noqa: E402
from plagdetect.textutils import mosaic_mix, sentences, synonymize  # noqa: E402
from plagdetect.understanding import build_idf  # noqa: E402
from gen_corpus import TOPICS, render_sections  # noqa: E402

ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
CORPUS_DIR = os.path.join(ROOT, "data", "corpus")
MODELS_DIR = os.path.join(ROOT, "models")


def word_dropout(text, rng, p=0.15):
    words = text.split()
    kept = [w for w in words if rng.rand() > p]
    return " ".join(kept if len(kept) > 3 else words)


def build_siamese_pairs(corpus, rng, n_pos=1500, n_neg=1500):
    all_sents = []
    for d in corpus:
        for sec in ("introduction", "literature_review", "methodology", "results"):
            all_sents.extend(sentences(d.sections.get(sec, "")))
    pairs = []
    for _ in range(n_pos):
        s = all_sents[rng.randint(len(all_sents))]
        t = synonymize(s, rng, p=0.7) if rng.rand() < 0.5 else word_dropout(s, rng)
        pairs.append((s, t, 1))
    for _ in range(n_neg):
        a = all_sents[rng.randint(len(all_sents))]
        b = all_sents[rng.randint(len(all_sents))]
        if a[:40] != b[:40]:
            pairs.append((a, b, 0))
    return pairs


def build_classifier_data(corpus, encoder, idf, default_idf, rng, n_per=110):
    """Self-consistent training: transforms -> real evidence features -> label."""
    raws = {}
    for fn in os.listdir(CORPUS_DIR):
        if fn.endswith(".json"):
            with open(os.path.join(CORPUS_DIR, fn), "r", encoding="utf-8") as f:
                raws[fn[:-5]] = json.load(f)
    by_topic = {}
    for d in corpus:
        by_topic.setdefault(raws[d.doc_id]["topic"], []).append(d)

    X, y = [], []
    secs = ["literature_review", "methodology", "results"]
    docs = list(corpus)
    for it in range(n_per):
        src = docs[rng.randint(len(docs))]
        topic = raws[src.doc_id]["topic"]
        sec = secs[rng.randint(len(secs))]
        src_text = src.sections[sec]
        src_sents = sentences(src_text)
        other_topic = [k for k in by_topic if k != topic][rng.randint(len(by_topic) - 1)]
        other = by_topic[other_topic][rng.randint(len(by_topic[other_topic]))]

        variants = {
            "clean": (other.sections[sec], other.references),
            "clone": (src_text, list(src.references)),
            "find_replace": (synonymize(src_text, rng, p=0.9),
                             list(src.references[:6])),
            "mosaic": (mosaic_mix(src_sents, sentences(other.sections[sec]), rng),
                       list(src.references[:5]) + list(other.references[:3])),
            "idea": (render_sections(TOPICS[topic], raws[src.doc_id]["vals"],
                                     np.random.RandomState(rng.randint(10_000)),
                                     alt=True)[sec],
                     list(src.references[:6])),
        }
        for label, (text, refs) in variants.items():
            ev = compare_section(text, refs, src, encoder, idf, default_idf)
            X.append(ev["features"])
            y.append(CLASSES.index(label))
        if (it + 1) % 25 == 0:
            print(f"  classifier data {5 * (it + 1)}/{5 * n_per} samples")
    return np.array(X), np.array(y)


def main():
    os.makedirs(MODELS_DIR, exist_ok=True)
    rng = np.random.RandomState(5)
    corpus = load_corpus(CORPUS_DIR)
    idf, default_idf = build_idf(corpus)
    print(f"[1/4] Siamese encoder -- corpus={len(corpus)} docs")
    t0 = time.time()
    enc = SiameseEncoder()
    enc.train(build_siamese_pairs(corpus, rng), epochs=5)
    enc.save(os.path.join(MODELS_DIR, "siamese.npz"))
    print(f"      saved siamese.npz ({time.time() - t0:.1f}s)")

    print("[2/4] DQN investigation planner")
    t0 = time.time()
    qnet, returns = train_dqn(episodes=900)
    qnet.save(os.path.join(MODELS_DIR, "dqn.npz"))
    print(f"      saved dqn.npz, final avg return "
          f"{np.mean(returns[-100:]):+.2f} ({time.time() - t0:.1f}s)")

    print("[3/4] Plagiarism-type classifier (features via real evidence pipeline)")
    t0 = time.time()
    X, y = build_classifier_data(corpus, enc, idf, default_idf, rng)
    split = int(0.85 * len(X))
    perm = rng.permutation(len(X))
    Xtr, ytr = X[perm[:split]], y[perm[:split]]
    Xte, yte = X[perm[split:]], y[perm[split:]]
    clf = TypeClassifier()
    clf.train(Xtr, ytr, epochs=60)
    preds = [clf.predict(x)[0] for x in Xte]
    acc = float(np.mean([CLASSES.index(p) == t for p, t in zip(preds, yte)]))
    clf.save(os.path.join(MODELS_DIR, "classifier.npz"))
    print(f"      saved classifier.npz, held-out acc={acc:.3f} "
          f"({time.time() - t0:.1f}s)")

    print("[4/4] Causal bandit priors + Bayesian evidence weights")
    CausalQueryBandit().save(os.path.join(MODELS_DIR, "bandit.json"))
    EvidenceWeights().save(os.path.join(MODELS_DIR, "weights.json"))
    print("      saved bandit.json, weights.json")
    print("Training complete.")


if __name__ == "__main__":
    main()