PART 01
具体对象Concrete object
这里是一张具体图片，不是“猫图”这个抽象类别。This is one specific image, not the abstract category “cat images.”
01 · 先看桌上的四张图01 · Start with four images
三张猫图和一张狗图，是一条数据，还是四条？Are three cat photos and one dog photo one data item—or four?
02样本装篮台Sample Basket
拖拽归集型Drag and collect篮中样本Samples in basket
0
条items
猫Cats
0
张图images
狗Dogs
0
张图images
03一条样本的 4 个零件Four Parts of One Sample
对象到记录Object to recordPART 02
可观察内容Observed content
像素、来源与时间等内容可以被读取并记录。Pixels, source, and time can be observed and recorded.
PART 03
字段Fields
一条记录可以含 ID、类别和尺寸，但仍只对应一个对象。A record may hold an ID, class, and size while still referring to one object.
PART 04
批次Batch
多条样本合起来才是一批；批次大小不是样本字段。Several samples form a batch; batch size is not a field inside one sample.
04最小机制：数对象，不数篮子Minimal Mechanism: Count Objects, Not Baskets
N = Σ 1N = Σi=1batch 1
上方每放入一个可独立记录的对象，N 增加 1。拖出后 N 减少 1。Each independently recorded object adds one to N. Drag it out and N falls by one.
观察一个对象Observe one object确定记录边界Set the record boundary
生成一条记录Create one record字段属于同一对象Fields belong together
汇成一批Assemble a batch样本数等于记录数Samples equal records
05边界实验：把整篮误算成一条Boundary Test: Miscount the Whole Basket as One
主动制造失败Create a failure故意混淆一条与一批Confuse one item with one batch
点击后把四张图装成一个“批次文件”，却把记录数写成 1。观察错误口径会漏掉多少样本。Bundle four images into one batch file but record the sample count as one. See how many items disappear from the count.
尚未执行。先想想文件数和样本数是否总相等。Not run yet. Decide whether file count always equals sample count.
✗ 一个文件固定就是一条样本。One file always equals one sample.
→ ✓ 样本边界由可独立观察和记录的对象决定。The observation and record boundary defines a sample.
✗ 一批数据可以当作一个训练例子。A whole batch can be counted as one training example.
→ ✓ 批次是多条样本的集合。A batch is a collection of samples.
✗ 多个字段就是多条样本。Several fields mean several samples.
→ ✓ 同一对象的多个字段仍属于同一条记录。Several fields for one object still form one record.
06 · 一句话带走06 · One line to keep