File size: 3,231 Bytes
6e15b0d | 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 | import argparse
import csv
import hashlib
import json
import shutil
from pathlib import Path
ROOT = Path(__file__).resolve().parents[3]
# registers new canonical patch cells: copies their packets into data/packets and adds the rows the tuning scripts read
# (hpc-run/jobs/cells.tsv, ex-042 best-top-p.csv, experiments/miscellaneous/shared/cells.json); cells already present are left alone
def main():
ap = argparse.ArgumentParser(description='Register new canonical patch cells from a staging folder.')
ap.add_argument('--cells', required=True, help='csv with dataset, encoder, scale, top1_bacc_rule, best_top_p')
ap.add_argument('--staging', required=True, help='folder holding top1-out/{native,multiscale}-build/packets, topp-out/pooled, text-out and best-top-p-hpc.csv')
ap.add_argument('--tag', required=True, help='suffix of the new top-1 folders, for example 261028')
a = ap.parse_args()
staging = Path(a.staging)
rows = list(csv.DictReader(open(a.cells)))
tsv, best_csv, cells_json = ROOT / 'embedding-acquire/hpc-run/jobs/cells.tsv', ROOT / 'experiments/ex-042-top-p/results/best-top-p.csv', ROOT / 'experiments/miscellaneous/shared/cells.json'
have = {tuple(l.split('\t')[:2]) for l in tsv.read_text().splitlines()[1:]}
hpc = {(r['dataset'], r['encoder'], r['scale']): r for r in csv.DictReader(open(staging / 'best-top-p-hpc.csv'))}
cfg = json.loads(cells_json.read_text())
with open(tsv, 'a') as t, open(best_csv, 'a', newline='') as b:
bw = csv.writer(b)
for r in rows:
d, e, s, p = r['dataset'], r['encoder'], r['scale'], int(r['best_top_p'])
if (d, e) in have:
print(f'skip {d} {e}: already a canonical cell')
continue
kind = 'native' if s == 'native' else 'multiscale'
name = f'{d}--{e}--{s}--short--filtered-class.pt'
top1 = ROOT / f'data/packets/patch/top1/{kind}-{a.tag}/packets/{name}'
top1.parent.mkdir(parents=True, exist_ok=True)
shutil.copy2(staging / f'top1-out/{kind}-build/packets/{name}', top1)
shutil.copy2(staging / f'text-out/{d}--{e}.pt', ROOT / f'data/packets/text/{d}--{e}.pt')
if p > 1:
shutil.copy2(staging / f'topp-out/pooled/{d}--{e}--{s}--P{p}.pt', ROOT / f'data/packets/patch/pooled/{d}--{e}--{s}--P{p}.pt')
rel = str(top1.relative_to(ROOT))
t.write(f'{d}\t{e}\t{s}\tpatch\t{float(r["top1_bacc_rule"]):.2f}\t{rel}\t{top1.stat().st_size}\t{hashlib.sha256(top1.read_bytes()).hexdigest()}\n')
h = hpc[(d, e, s)]
if int(h['P']) != p:
raise ValueError(f'{d} {e} {s}: best Top-P {h["P"]} in the job output differs from {p} in the cell list')
bw.writerow([h[k] for k in ('dataset', 'encoder', 'scale', 'P', 'f1_weighted', 'balanced_accuracy', 'auc', 'slides', 'slides_out_of_patches')])
for setting, top_p in (('top1', 1), ('best', p)):
cfg[setting][f'{d}--{e}'] = dict(dataset=d, encoder=e, scale=s, packet=rel, top_p=top_p)
print(f'added {d} {e} {s} P{p}')
cells_json.write_text(json.dumps(cfg, indent=1) + '\n')
if __name__ == '__main__':
main()
|