ppocr_ref.py

#!/usr/bin/env python3
"""Oracle onnxruntime pour la parité PP-OCRv6 tiny (burn vs ORT).

Deux modes par réseau :
  * « exact » : les tenseurs d'entrée pré-traités par burn_ppocr (fichiers .f32
    little-endian + formes dans le manifeste) sont donnés à ORT tels quels →
    ne compare que les réseaux ;
  * « own » : ce script refait le pré-traitement de PaddleOCR lui-même (PIL
    BILINEAR, BGR, mean/std, crops à partir des boîtes JSON de burn_ppocr) →
    compare réseaux + pré-traitement.

Usage :
  ppocr_ref.py --image hello.png --det det_pads.onnx --rec rec_pads.onnx \
      --work DIR   # DIR contient boxes.json (+ det_input.f32, rec_input.f32, manifest.json)
Écrit dans DIR : det_out_exact.f32, det_out_own.f32, rec_out_exact.f32,
rec_out_own.f32 et ref.json (formes).
"""
import argparse
import json
import math
import os

import numpy as np
import onnxruntime as ort
from PIL import Image

MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
STD = np.array([0.229, 0.224, 0.225], dtype=np.float32)


def det_resize_dims(w, h, limit, max_side):
    ratio = limit / min(w, h) if min(w, h) < limit else 1.0
    if max(w, h) * ratio > max_side:
        ratio = max_side / max(w, h)
    rw, rh = int(w * ratio), int(h * ratio)
    rw = max(int(round(rw / 32) * 32), 32)
    rh = max(int(round(rh / 32) * 32), 32)
    return rw, rh


def det_own_input(img, limit, max_side):
    rw, rh = det_resize_dims(img.width, img.height, limit, max_side)
    resized = np.asarray(img.resize((rw, rh), Image.BILINEAR), dtype=np.float32)  # HWC RGB
    bgr = resized[:, :, ::-1] / 255.0
    norm = (bgr - MEAN) / STD
    return norm.transpose(2, 0, 1)[None].astype(np.float32)


def crop_rotate(arr, x0, y0, x1, y1):
    c = arr[y0:y1, x0:x1]
    h, w = c.shape[:2]
    if w > 0 and h / w >= 1.5:
        c = np.rot90(c)
    return c


def rec_own_input(img, boxes, rec_h, img_w):
    arr = np.asarray(img)
    crops = [crop_rotate(arr, *b) for b in boxes]
    out = np.zeros((len(crops), 3, rec_h, img_w), dtype=np.float32)
    for i, c in enumerate(crops):
        h, w = c.shape[:2]
        if h == 0 or w == 0:
            continue
        rw = min(max(int(math.ceil(rec_h * w / h)), 1), img_w)
        r = np.asarray(Image.fromarray(c).resize((rw, rec_h), Image.BILINEAR), dtype=np.float32)
        bgr = r[:, :, ::-1] / 255.0
        norm = (bgr - 0.5) / 0.5
        out[i, :, :, :rw] = norm.transpose(2, 0, 1)
    return out


def read_f32(path, shape):
    return np.fromfile(path, dtype="<f4").reshape(shape)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--image", required=True)
    ap.add_argument("--det", required=True)
    ap.add_argument("--rec", required=True)
    ap.add_argument("--work", required=True)
    ap.add_argument("--limit", type=int, default=736)
    ap.add_argument("--max-side", type=int, default=4000)
    ap.add_argument("--rec-height", type=int, default=48)
    ap.add_argument("--rec-width", type=int, default=320, help="largeur du lot calculée par burn_ppocr")
    a = ap.parse_args()

    so = ort.SessionOptions()
    so.intra_op_num_threads = 1
    det = ort.InferenceSession(a.det, so, providers=["CPUExecutionProvider"])
    rec = ort.InferenceSession(a.rec, so, providers=["CPUExecutionProvider"])
    din, rin = det.get_inputs()[0].name, rec.get_inputs()[0].name

    img = Image.open(a.image).convert("RGB")
    work = a.work
    manifest = json.load(open(os.path.join(work, "manifest.json")))
    boxes = manifest["boxes"]
    out = {}

    # det, exact
    shape = manifest["det_input_shape"]
    x = read_f32(os.path.join(work, "det_input.f32"), shape)
    y = det.run(None, {din: x})[0]
    y.astype("<f4").tofile(os.path.join(work, "det_out_exact.f32"))
    out["det_exact_shape"] = list(y.shape)
    # det, own preprocessing
    x2 = det_own_input(img, a.limit, a.max_side)
    out["det_own_input_max_abs_diff"] = float(np.abs(x2 - x).max()) if x2.shape == x.shape else None
    y2 = det.run(None, {din: x2})[0]
    y2.astype("<f4").tofile(os.path.join(work, "det_out_own.f32"))
    out["det_own_shape"] = list(y2.shape)

    # rec, exact
    shape = manifest["rec_input_shape"]
    if shape[0] > 0:
        xr = read_f32(os.path.join(work, "rec_input.f32"), shape)
        yr = rec.run(None, {rin: xr})[0]
        yr.astype("<f4").tofile(os.path.join(work, "rec_out_exact.f32"))
        out["rec_exact_shape"] = list(yr.shape)
        xr2 = rec_own_input(img, boxes, a.rec_height, a.rec_width)
        out["rec_own_input_max_abs_diff"] = float(np.abs(xr2 - xr).max()) if xr2.shape == xr.shape else None
        yr2 = rec.run(None, {rin: xr2})[0]
        yr2.astype("<f4").tofile(os.path.join(work, "rec_out_own.f32"))
        out["rec_own_shape"] = list(yr2.shape)
        # texte décodé côté ORT (glouton, blank 0, espace dernier) pour lecture humaine
        idx = yr.argmax(axis=2)
        out["rec_exact_argmax"] = idx.tolist()
    else:
        out["rec_exact_shape"] = [0]

    json.dump(out, open(os.path.join(work, "ref.json"), "w"))
    print(json.dumps({k: v for k, v in out.items() if k != "rec_exact_argmax"}, indent=1))


if __name__ == "__main__":
    main()