| """Synthetic captions (27_azure_caption.py, GPT, 3-shot PathGen style) -> per-set meta for 08_extract_text.py. |
| VALIDATION ONLY: these texts are never used to train any head. |
| Output cache/orion_gentext_meta.csv (row i = ORION tile_id i of load_all_tiles(); `caption` column) |
| """ |
| import json |
| import os |
|
|
| import pandas as pd |
|
|
| from common import CACHE, load_all_tiles |
|
|
| C = pd.DataFrame([json.loads(l) for l in open(os.path.join(CACHE, "gen_text", "captions.jsonl"))]) |
| C = C[(C.set == "orion") & C.caption.notna() & (C.caption.str.len() > 0)].drop_duplicates("path", keep="last") |
| cap = dict(zip(C.path.map(os.path.basename), C.caption)) |
| df = load_all_tiles() |
| df["caption"] = df.image_path.map(lambda p: cap.get(os.path.basename(p))) |
| print(f"ORION tiles {len(df)}, with synthetic caption {df.caption.notna().sum()}") |
| assert df.caption.notna().all(), "missing captions -> rerun 27_azure_caption.py (resumable)" |
| df[["tile_id", "image_path", "caption"]].to_csv(os.path.join(CACHE, "orion_gentext_meta.csv"), index=False) |
|
|
| |
| A = pd.DataFrame([json.loads(l) for l in open(os.path.join(CACHE, "gen_text", "captions.jsonl"))]) |
| A = A[(A.set == "tcga_calib") & A.caption.notna()].drop_duplicates("path", keep="last") |
| if len(A): |
| I = pd.read_csv(os.path.join(CACHE, "allimg_meta.csv")) |
| shots = [s["key"][:12] for s in json.load(open(os.path.join(CACHE, "gen_text", "prompt.json")))["shots"]] |
| A["img_idx"] = A.path.map(dict(zip(I.image_path, I.index))) |
| A = A[A.img_idx.notna() & ~A.img_idx.map(lambda i: I.patient[int(i)] in shots)] |
| A["img_idx"] = A.img_idx.astype(int) |
| A[["img_idx", "path", "caption"]].reset_index(drop=True).to_csv(os.path.join(CACHE, "tcga_calib_meta.csv"), index_label="idx") |
| print(f"TCGA calibration captions: {len(A)} (few-shot patients excluded: {shots})") |
|
|
| |
| for st in ("dlpfc", "idc"): |
| X = pd.DataFrame([json.loads(l) for l in open(os.path.join(CACHE, "gen_text", "captions.jsonl"))]) |
| X = X[(X.set == st) & X.caption.notna() & (X.caption.str.len() > 0)].drop_duplicates("key", keep="last") |
| if not len(X): |
| continue |
| n = len(pd.read_csv(os.path.join(CACHE, "external", f"{st}_meta.csv"))) |
| cap = dict(zip(X.path.str.split(":").str[2].astype(int), X.caption)) |
| out = pd.DataFrame({"row": range(n), "caption": [cap.get(i) for i in range(n)]}) |
| print(f"{st}: {n} spots, with synthetic caption {out.caption.notna().sum()}") |
| assert out.caption.notna().all(), "missing captions -> rerun 27_azure_caption.py --sets dlpfc idc (resumable)" |
| out.to_csv(os.path.join(CACHE, f"{st}_gentext_meta.csv"), index=False) |
|
|
| |
| |
| N = pd.read_csv(os.path.join(CACHE, "notext_meta.csv")) |
| X = pd.DataFrame([json.loads(l) for l in open(os.path.join(CACHE, "gen_text", "captions.jsonl"))]) |
| X = X[X.caption.notna() & (X.caption.str.len() > 0)].drop_duplicates("path", keep="last") |
| cap = dict(zip(X.path.map(os.path.abspath), X.caption)) |
| N["caption"] = N.image_path.map(lambda p: cap.get(os.path.abspath(p))) |
| print(f"TCGA no-text images {len(N)}, with synthetic caption {N.caption.notna().sum()}") |
| assert N.caption.notna().all(), "missing captions -> rerun 27_azure_caption.py --sets tcga_notext" |
| N[["img_idx", "image_path", "patient", "caption"]].to_csv(os.path.join(CACHE, "tcga_notext_meta.csv"), index=False) |
|
|
