build_model_pool.py

"""Build the candidate model pool consumed by the recommendation web app.

The output is a single .npz that bundles, for every candidate model:
  - model_name (str)
  - size_id   (int, bucket id matching the trained MLPMetric)
  - family_id (int)
  - popularity (int, HF downloads in the last 30d; 0 if unknown)
  - hf_url    (str, https://huggingface.co/<name> if name looks like a repo id)

Run from the project root:
    python web/build_model_pool.py \
        --data-dir data/unified_augmented \
        --args     checkpoint/mlp/unified_augmented/ablation_no_model_id_no_dataset_id/args.json \
        --out      web/assets/model_pool.npz
"""
from __future__ import annotations

import argparse
import json
import os
import numpy as np


SIZE_EDGES_DEFAULT = [
    0.001, 0.003, 0.01, 0.03, 0.06, 0.1, 0.15, 0.2, 0.3, 0.4,
    0.5, 0.6, 0.8, 1, 3, 7, 14, 35, 70, 100, 1000,
]


def assign_size_bucket(size_b: float, size_edges: np.ndarray, unknown_id: int) -> int:
    try:
        x = float(size_b)
    except (TypeError, ValueError):
        return unknown_id
    if not np.isfinite(x) or x == 0.0:
        return unknown_id
    return int(np.searchsorted(size_edges, x, side="right"))


def get_size_b(profile_entry) -> float:
    if not isinstance(profile_entry, dict):
        return float("nan")
    size = profile_entry.get("size")
    try:
        if isinstance(size, str) and size.strip().lower() == "unknown":
            return float("nan")
        x = float(size)
        return x if x != 0.0 else float("nan")
    except Exception:
        return float("nan")


def hf_url_for(name: str) -> str:
    return f"https://huggingface.co/{name}" if "/" in name else ""


def main(argv=None):
    p = argparse.ArgumentParser()
    p.add_argument("--data-dir", default="data/unified_augmented")
    p.add_argument(
        "--args",
        default="checkpoint/mlp/unified_augmented/ablation_no_model_id_no_dataset_id/args.json",
        help="Path to the training args.json — used to read size_bucket so bucket ids align with the checkpoint.",
    )
    p.add_argument("--out", default="web/assets/model_pool.npz")
    p.add_argument(
        "--min-popularity",
        type=int,
        default=0,
        help="Drop candidate models with HF download count below this. 0 keeps all.",
    )
    args = p.parse_args(argv)

    os.makedirs(os.path.dirname(args.out), exist_ok=True)

    with open(os.path.join(args.data_dir, "model2id.json")) as f:
        model2id = json.load(f)
    with open(os.path.join(args.data_dir, "model2family.json")) as f:
        model2family = json.load(f)
    with open(os.path.join(args.data_dir, "family2id.json")) as f:
        family2id = json.load(f)
    with open(os.path.join(args.data_dir, "model_profile.json")) as f:
        model_profile = json.load(f)
    pop_path = os.path.join(args.data_dir, "model_popularity.json")
    pop_map = {}
    if os.path.exists(pop_path):
        pop_doc = json.load(open(pop_path))
        # Doc shape: {fetched_at, source, num_models, status_counts, models: {name: {downloads, status}}}
        models_field = pop_doc.get("models", pop_doc)
        for name, entry in models_field.items():
            if isinstance(entry, dict):
                pop_map[name] = int(entry.get("downloads", 0) or 0)
            else:
                try:
                    pop_map[name] = int(entry)
                except Exception:
                    pop_map[name] = 0

    if os.path.exists(args.args):
        train_args = json.load(open(args.args))
        size_edges = np.array(train_args.get("size_bucket", SIZE_EDGES_DEFAULT), dtype=float)
    else:
        size_edges = np.array(SIZE_EDGES_DEFAULT, dtype=float)
    unknown_size_id = len(size_edges) + 1

    unknown_family_id = family2id.get("unknown", len(family2id) - 1)

    names = []
    size_ids = []
    sizes_b = []
    family_ids = []
    popularities = []
    urls = []
    dropped_pop = 0
    for name in model2id.keys():
        pop = pop_map.get(name, 0)
        if pop < args.min_popularity:
            dropped_pop += 1
            continue
        size_b = get_size_b(model_profile.get(name))
        sid = assign_size_bucket(size_b, size_edges, unknown_size_id)
        fam = model2family.get(name, "unknown")
        fid = family2id.get(fam, unknown_family_id)
        names.append(name)
        size_ids.append(sid)
        sizes_b.append(size_b)  # NaN means unknown
        family_ids.append(fid)
        popularities.append(pop)
        urls.append(hf_url_for(name))

    names_arr = np.array(names, dtype=object)
    size_arr = np.array(size_ids, dtype=np.int64)
    sizes_b_arr = np.array(sizes_b, dtype=np.float32)
    fam_arr = np.array(family_ids, dtype=np.int64)
    pop_arr = np.array(popularities, dtype=np.int64)
    url_arr = np.array(urls, dtype=object)

    np.savez(
        args.out,
        names=names_arr,
        size_ids=size_arr,
        sizes_b=sizes_b_arr,
        family_ids=fam_arr,
        popularities=pop_arr,
        urls=url_arr,
    )
    print(f"Wrote {len(names)} models to {args.out} (dropped {dropped_pop} below min-popularity={args.min_popularity})")
    print(f"  unique families: {len(set(family_ids))}, unique size buckets: {len(set(size_ids))}")
    print(f"  models with HF URL: {sum(1 for u in urls if u)} / {len(urls)}")


if __name__ == "__main__":
    main()