# place the two tarballs (no extraction):
#   ~/tensorflow_datasets/downloads/manual/ILSVRC2012_img_train.tar
#   ~/tensorflow_datasets/downloads/manual/ILSVRC2012_img_val.tar
python -c "import tensorflow_datasets as tfds; \
           tfds.builder('imagenet2012').download_and_prepare()"

That validates, shuffles, and reshards into version 5.1.0 under ~/tensorflow_datasets/imagenet2012/5.1.0/: 1,281,167 training images across 1024 tfrecord shards (147.90 GB) and 50,000 validation images across 64 shards (6.74 GB), which is the exact layout every ImageNet trainer in this book consumes. Budget disk accordingly: the raw tarballs (\(\sim \)144 GB) plus the 5.1.0 output (\(\sim \)155 GB) plus TFDS’s intermediate extraction mean you want roughly 500 GB of free space to prepare it in one pass. From there the JAX bridge does no image-loading of its own: it calls tfds.load(’imagenet2012’, split=…, data_dir=…) and streams through tfds.as_numpy, so TFDS’s tf.data pipeline carries the loading (parallel shard reads, JPEG decode, prefetch, all engineered to scale), and we inherit that performance work for free.

How the ImageNet numbers are validated. Every top-1 and top-5 in this book follows the validation protocol of timm’s “ResNet Strikes Back” work, which is the reference the ImageNet recipes are ported from, so that a number here and a number there are answers to the same question. Concretely, four things are matched, and each was checked against timm rather than assumed:

• The denominator is all 50,000 validation images. This is worth stating because it moved: the loaders previously took drop_remainder=True on the validation split and scored \(49{,}920\), discarding \(80\) images. A top-1 over a different denominator is not an error bar, it is a different measurement.

• The centre-crop ratio is read from timm’s own model configuration, not from a global default. resnet50.a1_in1k, a2_in1k and a3_in1k all resolve crop_pct \(= 0.95\); the \(2018\) torchvision weights (tv_in1k) resolve \(0.875\). Both values appear in this book, attached to the recipe each belongs to.

• The field of view is identical. We decode a centred square of side \(\mathrm{crop\_ pct} \cdot \min (h,w)\) and resize it to the evaluation resolution; timm resizes the whole image so its shorter side is \(\mathrm{img\_ size}/\mathrm{crop\_ pct}\) and then centre-crops. The two orders end on the same pixels, and the ordering itself is measured at \(0.02\)–\(0.04\) points.

• Top-5 is the true label’s rank, matching the reference’s own \(\sum \mathbb {1}[\ell _i {\gt} \ell _y] {\lt} 5\), and the train/test resolution split is the recipe’s (RSB-A3 trains at \(160\) and evaluates at \(224\)).

• The resampler is timm’s. Validation resizes the whole image so its shorter side is \(\mathrm{img\_ size}/\mathrm{crop\_ pct}\) and only then centre-crops, which is timm’s order rather than the Google/TPU “Inception” crop-then-resize; and both that resize and the training RandomResizedCrop run antialiased, because PIL antialiases when downscaling and tf.image.resize does not by default. The two have to move together: a network partly fits the aliasing of the resampler it trained under, so matching only the evaluation side would be worse than matching neither. Antialiased tf.image.resize agrees with PIL to a mean \(|\Delta |\) of \(0.30\) of \(255\), flat across the whole range of downscale ratios ImageNet’s validation set spans.

Demo datasets. The demos/ trainers step past image classification into the three other task families in the book, which are detection, segmentation and language, and each pulls a dataset of its own:

• MSD Task01 (BraTS) is the segmentation dataset. It is four-modality brain-tumour MRI, and the thin-class collapse is the whole problem rather than a nuisance (enhancing tumour is \(\sim \)0.5 % of voxels). Two trainers read it: unet-brats-train (a from-scratch UNet at native \(240 \times 240\)) and unet-brats-r34 (the ResNet-34 encoder of §10.2.3, which needs the \(224\) centre-cropped build since its stride-32 backbone does not divide 240).

• tiny Shakespeare, a single 1 MB character stream, is what the TinyGPT char-level transformer (tinygpt-shakespeare) trains on.

• TinyStories, Eldan & Li’s synthetic short-story corpus at \(\sim \)2 GB, is its scaled-up sibling, byte-level-BPE tokenized to a deliberately small 4096-token vocabulary for the \(\sim \)8.5 M-parameter decoder-only GPT (tinystories).

• VisDrone-DET2019 is where detection leaves the household frame entirely. It is drone-altitude imagery where the objects are the problem: a median box is \(20 \times 25\) px in the source frame, some 70 of them per image, so a single \(7\times 7\) grid collapses outright and the multi-scale FPN head of Definition 95 becomes necessary rather than decorative.

• MS-COCO 2017 is the standard detection benchmark alongside it, at the opposite end of the same axis (photographic scale, 80 classes, 118k training images). The repository ships its loader and the two standard subsets described below, though no trainer in this book consumes it yet.

Same fetch-and-unpack story as the table above, with one extra step: each is tokenized or tiled into the trainer’s input format by the matching preprocess_*.py script before the first lake exe run.

The demo datasets, in the same columns as above. Each also has a preprocess_*.py step (e.g. preprocess_shakespeare.py builds train.bin/val.bin/vocab.txt) between download and training. For TinyStories, preprocess_tinystories.py trains a byte-level BPE tokenizer on a slice of the corpus and emits it in HuggingFace format (vocab.json/merges.txt) alongside the flat int32-LE token streams (train.bin/val.bin), and the small vocabulary keeps the in-graph one-hot embedding and head matmuls cheap and the tokenizer auditable. The two detection sets follow the same pattern: preprocess_visdrone.py and preprocess_coco.py write the same three record formats (single-grid, single-scale anchor, and multi-scale FPN), so one set of Lean loaders and one codegen serve both. VisDrone’s official Google-Drive/Baidu links are large-file-flaky, so download_visdrone.sh pulls the byte-identical Ultralytics release-asset mirror over plain HTTPS. COCO’s own hosting needs no such workaround.

Building the BraTS dataset. BraTS 2021 proper (1,251 cases) is gated behind a free Synapse account and a signed research-use agreement, so no script can fetch it. The Medical Segmentation Decathlon’s Task01_BrainTumour is the same task from open storage: 484 labelled 4D volumes built from the BraTS 2016/2017 cases, the same four co-registered MRI modalities (FLAIR / T1w / T1gd / T2w) and the same glioma sub-regions, as a single 7.6 GB tarball over S3. Thus we have fewer volumes and numbers that are not comparable to a BraTS 2021 leaderboard, and we also have a real segmentation result on real medical data that a reader can reproduce.