feat(detect): 固定 160x160 输入，只用裁剪与镜像填充，禁止缩放
产线要求送进模型的图固定为 N×N 且不许 resize——缺陷的绝对尺度是判别依据， 拉伸会把"5 像素的黑点"变成别的东西。 - engine/tiling.py：切片几何。边长 < N 镜像填充（reflect，真镜像，原图居中）； 边长 >= N 按步长裁剪、末片贴边保证全覆盖。框跟着平移裁剪，露出面积低于 min_visibility 的丢弃；to_source 反向映射供推理时拼回原图。 - scripts/tile_dataset.py：整套数据集转换。有框图切出的空片会爆量 （524x959 能切 24 片），用 --max-empty-ratio 限量；原本无框的真负样本不受限。 - 训练配置 imgsz 改 160，并显式关掉 scale 与 translate——ultralytics 的默认值是 0.5 / 0.1 而不是 0，不写出来等于一直在做仿射缩放，违反约束。 尺寸正好等于 imgsz 时 letterbox 退化为空操作（r=1、dw=dh=0），核对过源码。 合成数据上验证：所有片正好 160x160、框区域与原图逐像素一致、填充是真镜像。 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Shaun Zhou
co-authored by
Claude Opus 5
Claude Opus 5
parent
1684fddd7f
commit
2a92275317
@@ -0,0 +1,183 @@
|
||||
"""把数据集切成固定边长的方片。"""
|
||||
|
||||
import csv
|
||||
import json
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
|
||||
pytest.importorskip("PIL")
|
||||
|
||||
import numpy as np # noqa: E402
|
||||
import yaml # noqa: E402
|
||||
from PIL import Image # noqa: E402
|
||||
|
||||
sys.path.insert(0, "scripts")
|
||||
import tile_dataset as td # noqa: E402
|
||||
|
||||
from engine.detection import parse_label # noqa: E402
|
||||
from engine.tiling import Tile, to_source # noqa: E402
|
||||
|
||||
SIZE = 160
|
||||
CLASSES = ["blackdot", "grease"]
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def dataset(tmp_path):
|
||||
"""含小图（要填充）、大图（要切片）与真负样本的最小数据集。"""
|
||||
root = tmp_path / "src"
|
||||
rng = np.random.default_rng(0)
|
||||
for split in ("train", "val"):
|
||||
(root / "images" / split).mkdir(parents=True)
|
||||
(root / "labels" / split).mkdir(parents=True)
|
||||
|
||||
def put(split, name, width, height, lines):
|
||||
Image.fromarray(rng.integers(0, 255, (height, width, 3), dtype=np.uint8)).save(
|
||||
root / "images" / split / f"{name}.png"
|
||||
)
|
||||
(root / "labels" / split / f"{name}.txt").write_text("".join(lines), encoding="utf-8")
|
||||
|
||||
put("train", "blackdot_0001", 128, 128, ["0 0.5 0.5 0.1 0.1\n"]) # 小图
|
||||
put("train", "grease_0001", 400, 300, ["1 0.25 0.3 0.1 0.1\n"]) # 大图
|
||||
put("train", "normal_0001", 116, 83, []) # 真负样本
|
||||
put("val", "blackdot_0002", 320, 160, ["0 0.25 0.5 0.1 0.2\n"])
|
||||
|
||||
(root / "data.yaml").write_text(
|
||||
yaml.safe_dump({"path": str(root), "train": "images/train", "val": "images/val",
|
||||
"nc": 2, "names": dict(enumerate(CLASSES))}),
|
||||
encoding="utf-8",
|
||||
)
|
||||
return root
|
||||
|
||||
|
||||
def run(dataset, out, *extra):
|
||||
argv = ["tile_dataset.py", "--data", str(dataset / "data.yaml"), "--out", str(out),
|
||||
"--size", str(SIZE), *extra]
|
||||
original, sys.argv = sys.argv, argv
|
||||
try:
|
||||
return td.main()
|
||||
finally:
|
||||
sys.argv = original
|
||||
|
||||
|
||||
def tiles_of(out):
|
||||
with (out / "tiles.csv").open(encoding="utf-8-sig", newline="") as handle:
|
||||
return list(csv.DictReader(handle))
|
||||
|
||||
|
||||
def test_every_tile_is_exactly_the_target_size(dataset, tmp_path):
|
||||
"""这是整件事的目的：模型输入固定 160x160。"""
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out) == 0
|
||||
images = list((out / "images").rglob("*.png"))
|
||||
assert images
|
||||
for path in images:
|
||||
with Image.open(path) as image:
|
||||
assert image.size == (SIZE, SIZE)
|
||||
|
||||
|
||||
def test_pixels_are_never_resampled(dataset, tmp_path):
|
||||
"""片上框住的像素必须与原图逐像素相同——有任何缩放都会露馅。"""
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out) == 0
|
||||
|
||||
checked = 0
|
||||
for row in tiles_of(out):
|
||||
if int(row["boxes"]) == 0:
|
||||
continue
|
||||
tile = Tile(int(row["x"]), int(row["y"]), int(row["crop_width"]),
|
||||
int(row["crop_height"]), int(row["pad_left"]), int(row["pad_top"]))
|
||||
patch = np.asarray(Image.open(out / "images" / row["split"] / row["tile"]).convert("RGB"))
|
||||
source = np.asarray(
|
||||
Image.open(dataset / "images" / row["split"] / row["source"]).convert("RGB")
|
||||
)
|
||||
label = out / "labels" / row["split"] / f"{Path(row['tile']).stem}.txt"
|
||||
for box in parse_label(label.read_text(encoding="utf-8")):
|
||||
back = to_source(box, tile, SIZE)
|
||||
region = patch[
|
||||
round(box.y1 * SIZE):round(box.y2 * SIZE),
|
||||
round(box.x1 * SIZE):round(box.x2 * SIZE),
|
||||
]
|
||||
original = source[round(back.y1):round(back.y2), round(back.x1):round(back.x2)]
|
||||
assert np.array_equal(region, original)
|
||||
checked += 1
|
||||
assert checked >= 3
|
||||
|
||||
|
||||
def test_small_image_is_mirror_padded_not_stretched(dataset, tmp_path):
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out) == 0
|
||||
row = next(r for r in tiles_of(out) if r["source"] == "blackdot_0001.png")
|
||||
tile = Tile(int(row["x"]), int(row["y"]), int(row["crop_width"]), int(row["crop_height"]),
|
||||
int(row["pad_left"]), int(row["pad_top"]))
|
||||
|
||||
patch = np.asarray(Image.open(out / "images" / row["split"] / row["tile"]).convert("RGB"))
|
||||
source = np.asarray(Image.open(dataset / "images/train/blackdot_0001.png").convert("RGB"))
|
||||
inner = patch[tile.pad_top:tile.pad_top + 128, tile.pad_left:tile.pad_left + 128]
|
||||
assert np.array_equal(inner, source) # 原图原样居中
|
||||
# 填充是镜像而不是复制边缘：左侧第一个填充列等于原图第 1 列（不是第 0 列）
|
||||
assert np.array_equal(patch[tile.pad_top, tile.pad_left - 1], source[0, 1])
|
||||
|
||||
|
||||
def test_large_image_is_cut_into_covering_tiles(dataset, tmp_path):
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out, "--max-empty-ratio", "-1") == 0
|
||||
rows = [r for r in tiles_of(out) if r["source"] == "grease_0001.png"]
|
||||
# 400x300 -> x 起点 [0,160,240]，y 起点 [0,140]
|
||||
assert len(rows) == 6
|
||||
assert {int(r["x"]) for r in rows} == {0, 160, 240}
|
||||
assert {int(r["y"]) for r in rows} == {0, 140}
|
||||
assert all(int(r["pad_left"]) == 0 and int(r["pad_top"]) == 0 for r in rows)
|
||||
|
||||
|
||||
def test_true_negative_images_keep_all_their_tiles(dataset, tmp_path):
|
||||
"""原本就没框的图是真负样本，不受空片配额限制。"""
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out, "--max-empty-ratio", "0") == 0
|
||||
assert [r for r in tiles_of(out) if r["source"] == "normal_0001.png"]
|
||||
|
||||
|
||||
def test_empty_tile_quota_limits_background_from_positive_images(dataset, tmp_path):
|
||||
"""有框图切出的空片会爆量，要能限量。"""
|
||||
strict = tmp_path / "strict"
|
||||
assert run(dataset, strict, "--max-empty-ratio", "0") == 0
|
||||
loose = tmp_path / "loose"
|
||||
assert run(dataset, loose, "--max-empty-ratio", "-1") == 0
|
||||
|
||||
def empties(out):
|
||||
return [r for r in tiles_of(out) if r["source"] == "grease_0001.png" and r["boxes"] == "0"]
|
||||
|
||||
assert empties(strict) == []
|
||||
assert len(empties(loose)) == 5 # 6 片里 1 片有框
|
||||
|
||||
|
||||
def test_box_split_by_a_cut_line_is_dropped_when_barely_visible(dataset, tmp_path):
|
||||
"""min_visibility > 0.5 时，一个框最多只会出现在一片里。"""
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out, "--min-visibility", "0.6") == 0
|
||||
# val 里的 320x160：框中心在 x=80，完整落在第一片内
|
||||
rows = [r for r in tiles_of(out) if r["source"] == "blackdot_0002.png"]
|
||||
assert sum(int(r["boxes"]) for r in rows) == 1
|
||||
|
||||
|
||||
def test_outputs_are_a_valid_yolo_dataset(dataset, tmp_path):
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out) == 0
|
||||
|
||||
meta = yaml.safe_load((out / "data.yaml").read_text(encoding="utf-8"))
|
||||
assert [meta["names"][i] for i in range(meta["nc"])] == CLASSES
|
||||
assert (out / "classes.txt").read_text(encoding="utf-8").split() == CLASSES
|
||||
for row in tiles_of(out):
|
||||
assert (out / "images" / row["split"] / row["tile"]).exists()
|
||||
assert (out / "labels" / row["split"] / f"{Path(row['tile']).stem}.txt").exists()
|
||||
|
||||
summary = json.loads((out / "summary.json").read_text(encoding="utf-8"))
|
||||
assert summary["size"] == SIZE
|
||||
assert summary["counts"]["源图"] == 4
|
||||
|
||||
|
||||
def test_dry_run_writes_nothing(dataset, tmp_path):
|
||||
out = tmp_path / "out"
|
||||
assert run(dataset, out, "--dry-run") == 0
|
||||
assert not out.exists()
|
||||
Reference in New Issue
Block a user