File size: 14,926 Bytes
69770c7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 | """Process one IMPACT year: download, enrich, embed, push to HF Hub.
Runs on Modal GPU — no local disk used. Each year is ~4-5GB of images,
~2GB of text. Modal ephemeral storage handles it; everything is deleted
when the job exits.
Pipeline per year:
1. Download IMPACT year.zip + CSV from HF (images + metadata)
2. Download PatentsView text TSVs from S3 (draw_desc, detail_desc, patent meta)
3. Extract text TSVs with 7za (handles deflate64)
4. Join: IMPACT metadata × PatentsView text → enriched parquet
5. Run CLIP ViT-L/14 on GPU → embeddings parquet
6. Push both parquets to midah/patent-wireframes on HF Hub
7. Exit — ephemeral storage is cleared automatically
Safety:
- Idempotent: checks HF Hub before downloading anything
- One year at a time — run sequentially to verify, then parallelize
- Never touches local machine disk
Setup (one-time on any networked machine):
pip install modal
modal setup
modal secret create hf-secret HF_TOKEN=hf_...
Run one year (verify first):
modal run scripts/cloud/process_year_modal.py --year 2021
Run all years sequentially:
for year in 2022 2021 2020 2019 2018 2017 2016 2015 2014 2013 2012 2011 2010 2009 2008 2007; do
modal run scripts/cloud/process_year_modal.py --year $year
done
Run all years in parallel (after verifying one year works):
modal run scripts/cloud/process_year_modal.py --all-years
"""
import io
import os
import re
import subprocess
from pathlib import Path
import modal
# ── Modal image ───────────────────────────────────────────────────────────────
image = (
modal.Image.debian_slim(python_version="3.11")
.apt_install("p7zip-full") # for deflate64 extraction
.pip_install(
"open_clip_torch",
"huggingface_hub>=0.23",
"datasets",
"Pillow",
"pandas",
"numpy",
"tqdm",
"requests",
)
)
app = modal.App("patent-process-year", image=image)
hf_secret = modal.Secret.from_name("hf-secret")
PATENTSVIEW_URLS = {
"draw_desc": "https://s3.amazonaws.com/data.patentsview.org/draw-description-text/g_draw_desc_text_{year}.tsv.zip",
"detail_desc": "https://s3.amazonaws.com/data.patentsview.org/detail-description-text/g_detail_desc_text_{year}.tsv.zip",
"brf_sum": "https://s3.amazonaws.com/data.patentsview.org/brief-summary-text/g_brf_sum_text_{year}.tsv.zip",
"claims": "https://s3.amazonaws.com/data.patentsview.org/claims/g_claims_{year}.tsv.zip",
}
PATENT_META_URL = "https://s3.amazonaws.com/data.patentsview.org/download/g_patent.tsv.zip"
HF_REPO = "midah/patent-wireframes"
BATCH_SIZE = 64
# ── helpers ───────────────────────────────────────────────────────────────────
def already_on_hub(year: int, token: str) -> bool:
"""Check if this year's outputs are already on HF Hub."""
from huggingface_hub import list_repo_files
files = set(list_repo_files(HF_REPO, repo_type="dataset", token=token))
enriched = f"data/enriched_{year}.parquet"
embeddings = f"embeddings/embeddings_{year}_vitl14.parquet"
if enriched in files and embeddings in files:
print(f"Year {year}: already on Hub, skipping.")
return True
return False
def download_file(url: str, dest: Path, chunk_size: int = 8 * 1024 * 1024) -> Path:
import requests
print(f" Downloading {url.split('/')[-1]}...")
r = requests.get(url, stream=True, timeout=120)
r.raise_for_status()
with open(dest, "wb") as f:
for chunk in r.iter_content(chunk_size=chunk_size):
f.write(chunk)
size_mb = dest.stat().st_size / 1e6
print(f" → {dest.name} ({size_mb:.0f}MB)")
return dest
def extract_zip(zip_path: Path, out_dir: Path) -> Path | None:
"""Extract using 7za (handles deflate64 that Python zipfile can't)."""
out_dir.mkdir(parents=True, exist_ok=True)
result = subprocess.run(
["7za", "x", str(zip_path), f"-o{out_dir}", "-y"],
capture_output=True, text=True
)
if result.returncode != 0:
print(f" 7za error: {result.stderr[:200]}")
return None
tsv_files = list(out_dir.glob("*.tsv"))
if tsv_files:
print(f" Extracted: {tsv_files[0].name} ({tsv_files[0].stat().st_size/1e6:.0f}MB)")
return tsv_files[0]
return None
# ── core processing ───────────────────────────────────────────────────────────
@app.function(
gpu="A10G",
timeout=7200, # 2 hours max per year
memory=32768, # 32GB RAM
ephemeral_disk=51200, # 50GB ephemeral disk
secrets=[hf_secret],
)
def process_year(year: int) -> dict:
import ast
import csv
import zipfile
import numpy as np
import open_clip
import pandas as pd
import torch
from huggingface_hub import HfApi, hf_hub_download
from PIL import Image
from tqdm import tqdm
token = os.environ["HF_TOKEN"]
api = HfApi(token=token)
work = Path(f"/tmp/patent_{year}")
work.mkdir(exist_ok=True)
print(f"\n{'='*50}")
print(f"Processing year {year}")
print(f"{'='*50}")
# ── 0. Idempotency check ─────────────────────────────────────────────────
if already_on_hub(year, token):
return {"year": year, "status": "skipped"}
# ── 1. Download IMPACT metadata CSV ─────────────────────────────────────
print("\n[1/5] IMPACT metadata...")
csv_path = hf_hub_download(
repo_id="AI4Patents/IMPACT", filename=f"{year}.csv",
repo_type="dataset", token=token, local_dir=str(work)
)
impact_df = pd.read_csv(csv_path)
print(f" {len(impact_df):,} patents")
# Explode to figure level
rows = []
for _, row in impact_df.iterrows():
try:
fnames = ast.literal_eval(str(row["file_names"]))
fig_descs = ast.literal_eval(str(row["fig_desc"])) if pd.notna(row.get("fig_desc")) else []
except Exception:
continue
for i, fname in enumerate(fnames):
rows.append({
"patent_id": "D" + str(row["id"]).replace("D","").lstrip("0").zfill(7),
"figure_number": i,
"image_filename": fname,
"patent_title": row.get("title",""),
"caption": row.get("caption",""),
"class": row.get("class",""),
"year": year,
})
df = pd.DataFrame(rows)
print(f" Exploded to {len(df):,} figures")
# ── 2. Download PatentsView text TSVs ────────────────────────────────────
print("\n[2/5] PatentsView text tables...")
pv_dir = work / "patentsview"
pv_dir.mkdir(exist_ok=True)
text_dfs = {}
for table, url_template in PATENTSVIEW_URLS.items():
url = url_template.format(year=year)
zip_path = pv_dir / url.split("/")[-1]
try:
download_file(url, zip_path)
tsv_path = extract_zip(zip_path, pv_dir / table)
if tsv_path:
tdf = pd.read_csv(tsv_path, sep="\t", dtype=str, low_memory=False)
tdf["patent_id"] = tdf["patent_id"].apply(
lambda x: "D" + str(x).replace("D","").lstrip("0").zfill(7)
)
text_dfs[table] = tdf
zip_path.unlink() # free space immediately
except Exception as e:
print(f" {table}: skipped ({e})")
# Patent metadata (year-independent, check if already fetched)
meta_path = pv_dir / "g_patent.tsv"
if not meta_path.exists():
try:
zip_path = pv_dir / "g_patent.tsv.zip"
download_file(PATENT_META_URL, zip_path)
extract_zip(zip_path, pv_dir)
zip_path.unlink()
except Exception as e:
print(f" patent meta: skipped ({e})")
# ── 3. Join ──────────────────────────────────────────────────────────────
print("\n[3/5] Joining text tables...")
col_map = {
"draw_desc": ("draw_desc_text", "drawing_description"),
"detail_desc": ("detail_desc_text", "detailed_description"),
"brf_sum": ("brf_sum_text", "brief_summary"),
"claims": ("claims_text", "claims"),
}
for table, (src_col, dst_col) in col_map.items():
tdf = text_dfs.get(table, pd.DataFrame())
if not tdf.empty and src_col in tdf.columns:
agg = (tdf.groupby("patent_id")[src_col]
.apply(lambda x: "\n".join(x.dropna().astype(str)))
.reset_index().rename(columns={src_col: dst_col}))
df = df.merge(agg, on="patent_id", how="left")
df[dst_col] = df[dst_col].fillna("")
else:
df[dst_col] = ""
filled = (df[dst_col] != "").sum()
print(f" {dst_col}: {filled:,}/{len(df):,} filled")
# Merge patent metadata
meta_path = pv_dir / "g_patent.tsv"
if meta_path.exists():
meta = pd.read_csv(meta_path, sep="\t", dtype=str, low_memory=False)
meta["patent_id"] = meta["patent_id"].apply(
lambda x: "D" + str(x).replace("D","").lstrip("0").zfill(7)
)
for col in ["patent_date", "patent_type"]:
if col in meta.columns:
df = df.merge(meta[["patent_id", col]].drop_duplicates("patent_id"),
on="patent_id", how="left")
# Figure ID + siblings
df["figure_id"] = df["patent_id"] + "_" + df["figure_number"].astype(str)
patent_groups = df.groupby("patent_id")["figure_id"].apply(list).to_dict()
df["n_figures_in_patent"] = df["patent_id"].map(df.groupby("patent_id").size())
df["sibling_figure_ids"] = df.apply(
lambda r: [f for f in patent_groups[r["patent_id"]] if f != r["figure_id"]],
axis=1
)
# ── 4. CLIP embeddings ───────────────────────────────────────────────────
print("\n[4/5] CLIP embeddings...")
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f" Device: {device}")
model, _, preprocess = open_clip.create_model_and_transforms(
"ViT-L-14", pretrained="openai"
)
model = model.to(device).eval()
# Download IMPACT zip
print(" Downloading IMPACT images (~4-5GB)...")
zip_path = hf_hub_download(
repo_id="AI4Patents/IMPACT", filename=f"{year}.zip",
repo_type="dataset", token=token, local_dir=str(work)
)
print(f" Zip: {Path(zip_path).stat().st_size/1e9:.1f}GB")
all_ids, all_vecs = [], []
def load_image(fname: str) -> Image.Image | None:
parts = fname.split("-D0")
if len(parts) < 2:
return None
inner = f"{year}/{parts[0]}/{fname}"
try:
with zipfile.ZipFile(zip_path) as z:
with z.open(inner) as f:
return Image.open(io.BytesIO(f.read())).convert("RGB")
except Exception:
return None
batch_imgs, batch_ids = [], []
def flush():
if not batch_imgs:
return
tensors = torch.stack([preprocess(im) for im in batch_imgs]).to(device)
with torch.no_grad():
feats = model.encode_image(tensors)
feats = feats / feats.norm(dim=-1, keepdim=True)
all_vecs.append(feats.cpu().numpy())
all_ids.extend(batch_ids)
batch_imgs.clear()
batch_ids.clear()
for _, row in tqdm(df.iterrows(), total=len(df), desc=" Embedding"):
img = load_image(row["image_filename"])
if img is None:
continue
batch_imgs.append(img)
batch_ids.append(row["figure_id"])
if len(batch_imgs) >= BATCH_SIZE:
flush()
flush()
vecs = np.vstack(all_vecs).astype(np.float32)
norms = np.linalg.norm(vecs, axis=1, keepdims=True)
vecs /= np.maximum(norms, 1e-8)
print(f" Embedded {len(all_ids):,} figures, shape {vecs.shape}")
emb_df = pd.DataFrame({"figure_id": all_ids, "embedding": list(vecs)})
# ── 5. Push to Hub ───────────────────────────────────────────────────────
print("\n[5/5] Pushing to HF Hub...")
enriched_path = work / f"enriched_{year}.parquet"
df.to_parquet(enriched_path, index=False)
api.upload_file(
path_or_fileobj=str(enriched_path),
path_in_repo=f"data/enriched_{year}.parquet",
repo_id=HF_REPO, repo_type="dataset",
commit_message=f"Add enriched_{year}.parquet ({len(df):,} figures)"
)
print(f" Pushed data/enriched_{year}.parquet")
emb_path = work / f"embeddings_{year}_vitl14.parquet"
emb_df.to_parquet(emb_path, index=False)
api.upload_file(
path_or_fileobj=str(emb_path),
path_in_repo=f"embeddings/embeddings_{year}_vitl14.parquet",
repo_id=HF_REPO, repo_type="dataset",
commit_message=f"Add embeddings_{year}_vitl14.parquet"
)
print(f" Pushed embeddings/embeddings_{year}_vitl14.parquet")
return {
"year": year,
"status": "done",
"n_figures": len(df),
"n_embedded": len(all_ids),
}
# ── entrypoints ───────────────────────────────────────────────────────────────
@app.local_entrypoint()
def main(year: int = 2021, all_years: bool = False):
"""
Test with one year first:
modal run scripts/cloud/process_year_modal.py --year 2021
Then run all years:
modal run scripts/cloud/process_year_modal.py --all-years
"""
if all_years:
years = list(range(2007, 2023)) # 2007–2022
print(f"Processing all {len(years)} years: {years}")
# Sequential for safety — can switch to .map() after verifying one year
for y in years:
result = process_year.remote(y)
print(f"Year {y}: {result}")
else:
print(f"Processing year {year} (test run)...")
result = process_year.remote(year)
print(f"Result: {result}")
|