Transforms complex documents like PDFs and Office docs into LLM-ready markdown/JSON for your Agentic workflows.
Python ai4sciencedocument-analysisdocxextract-datalayout-analysisocrparserpdfpdf-converterpdf-extractor-llmpdf-extractor-pretrainpdf-extractor-ragpdf-parserpptxpythonxlsx
"""仅供冻结语料实验使用的 N/H/W 归并;不改变生产 det 的预处理和后处理。"""
from __future__ import annotations
import math
import time
from typing import Any
SHAPE_STRATEGIES: dict[str, tuple[int | None, int | None]] = {
"F": (1, None),
"G": (4, None),
"H": (None, 64),
"I": (None, 128),
}
def project_shape(shape: list[int], policy: str) -> dict[str, Any]:
"""预先计算真实送入网络的形状、拆批数量和额外像素工作量。"""
n, c, height, width = shape
batch, alignment = SHAPE_STRATEGIES.get(policy, (None, None))
target_batch = batch or n
target_height = math.ceil(height / alignment) * alignment if alignment else height
target_width = math.ceil(width / alignment) * alignment if alignment else width
chunks = math.ceil(n / target_batch)
shapes = [[target_batch, c, target_height, target_width] for _ in range(chunks)]
return {
"original_shape": shape,
"network_shapes": shapes,
"real_images": n,
"processed_images": target_batch * chunks,
"original_pixels": n * height * width,
"processed_pixels": target_batch * chunks * target_height * target_width,
"spatial_alignment": alignment,
"batch_bucket": batch,
"spatial_pad_normalized_value": 0.0 if alignment else None,
"batch_pad": "repeat_last_real_image" if target_batch * chunks > n else None,
}
def process_bucketed_batch(detector: Any, batch_items: list[Any], policy: str) -> tuple[list[Any], float]:
"""保留原图 resize/坐标,只在网络入口拆批补齐,并在 CPU 恢复原概率图后执行一次 DB 后处理。"""
import numpy as np
import torch
if not batch_items:
return [], 0.0
started = time.perf_counter()
images = np.ascontiguousarray(np.stack([item[1] for item in batch_items]))
shape = list(images.shape)
projection = project_shape(shape, policy)
n, _, height, width = shape
batch, _, padded_height, padded_width = projection["network_shapes"][0]
maps = []
padding_seconds = 0.0
for offset in range(0, n, batch):
padding_started = time.perf_counter()
real = images[offset : offset + batch]
count = len(real)
padded = np.zeros((batch, shape[1], padded_height, padded_width), dtype=images.dtype)
padded[:count, :, :height, :width] = real
if count < batch:
# eval 网络不存在跨样本归一化,补齐样本的输出会在 CPU 丢弃。
padded[count:] = padded[count - 1]
padding_seconds += time.perf_counter() - padding_started
with torch.inference_mode():
inputs = detector._to_inference_dtype(torch.from_numpy(padded).to(detector.device))
outputs = detector.net(inputs)
predictions = detector._build_det_preds(outputs)["maps"]
if predictions.shape[:2] != (batch, 1) or predictions.shape[2:] != (padded_height, padded_width):
raise ValueError(f"归并后的概率图形状不符合 DBHead 约定:{predictions.shape}")
# 先回传 CPU,再裁切,避免额外 MPS slice 图缓存和错误的后处理缩放。
maps.append(np.ascontiguousarray(predictions[:count, :, :height, :width]))
del predictions, inputs, outputs, padded
concatenation_started = time.perf_counter()
restored = np.concatenate(maps, axis=0)
padding_seconds += time.perf_counter() - concatenation_started
detector.bucket_padding_seconds += padding_seconds
if restored.shape != (n, 1, height, width):
raise ValueError("归并后的概率图未恢复原始 N/H/W")
if detector.capture_maps:
detector.last_maps = restored
boxes = detector._postprocess_det_batch(
{"maps": restored}, np.stack([item[2] for item in batch_items]), [item[3] for item in batch_items]
)
elapsed = time.perf_counter() - started
return [(box, elapsed / n) for box in boxes], elapsed