Chapter 23 — Read the defaults, then start training

A config is a receipt, not decoration

A banquet recipe says “make 3,072 servings.” That number assumes two kitchens, eight ovens each, 96 servings per oven, and two rounds of accumulation. One home oven can test whether the batter bakes. It cannot honestly claim to have reproduced the banquet.

Read the public default end to end

The pinned conf/config.yaml trains ViT-B/16 on ten Walking Tours videos: 16 frames, stride 2, tubelet 1, 95% random dropping, and block_causal attention. Each clip produces one 224 global crop and ten 96 local crops. The projector is 768 → 2048 → 256. SIGReg uses 17 knots and 1,024 random projections; its weight remains 0.02.

AdamW uses learning rate 4e-4, weight decay 0.04, and betas (0.9, 0.95). The first 1,200 optimizer steps warm from 1e-4 to 4e-4; because end_lr == lr, the schedule stays flat afterward. The default runs 26 epochs, noted as roughly ten thousand optimizer steps.

Check the graph beside the recipe. Global and ten locals use one encoder. Locals reshape from [B,V,T,C,H,W] to [B×V,C,T,H,W]; eleven [CLS] summaries pass through one projector to [B,11,256]. The code evaluates (global_emb - embeddings).pow(2).mean()—including one harmless global-self zero—and sends [11,B,256] to SIGReg. Both losses update the shared encoder. There is no teacher forward.

Batch size is the easiest number to misquote. The comment assumes 2 nodes × 8 GPUs × 96 clips × accumulation 2 = 3,072. Move to different hardware without changing accumulation and the experiment has changed. EMA tracks only the encoder, updates every 32 optimizer steps at decay 0.9999, and saves as state_dict_ema. Online weights live in state_dict; the average neither enters the loss nor acts as a target.

Once data exists, a single-device, two-batch smoke run can look like this:

uv run python main.py \
  trainer.devices=1 trainer.num_nodes=1 trainer.strategy=auto \
  trainer.max_epochs=1 trainer.limit_train_batches=2 \
  loader.batch_size=2 loader.num_workers=0 accumulate_grad_batches=1 \
  model.name=vit_tiny augmentation.local_crops_number=2

This is a tiny effective-batch-2 pipeline check, not the default or a paper reproduction. Hydra writes run metadata under outputs/<date>/<time>/. But the pinned checkpoint.dirpath=checkpoints resolves from the launch directory because hydra.job.chdir=false, so checkpoints normally land in repository-root checkpoints/. A config comment suggesting the Hydra run directory conflicts with that behavior. Only explicit hydra.job.chdir=true makes a relative path follow the run directory.

Before a real launch, run uv run python main.py --cfg job and save the resolved YAML. Confirm data.train. Walking Tours has no held-out validation set: val points to train while trainer.limit_val_batches=0 disables validation. Training loss is therefore no generalization metric. Checkpoints save every 2,000 train steps and at the end; inspect both state_dict and state_dict_ema, and record actual world size.

Keep four recipes separate: the paper’s 20% K710, 240-epoch control; public ViT-B Walking Tours, 26-epoch default; 12-hour ViT-Tiny consumer demo on eight videos; released ViT-L trained on 1.8M VideoMix clips. Method identity does not make training conditions identical.

Finally, resume.weights_only=true loads weights but restarts optimizer and schedule. Set it to false to resume optimizer state and step count, and retain the original resolved config.

Configuration sources

• Pinned default config
• Pinned training entry and EMA callback
• Hydra working-directory behavior
• Controlled paper experiments and release model card

Keep the receipt

1. Effective batch 3,072 assumes 2×8 GPUs, 96 clips per GPU, and accumulation 2.
2. A two-batch local command checks plumbing; it inherits no paper result.
3. Shared method does not mean shared data, model, schedule, or budget.

Recipe questions

1. With one eight-GPU node and 96 clips per GPU, what accumulation keeps batch 3,072?
2. Does state_dict_ema provide an online encoder target?
3. Can the Walking Tours default reproduce the 20% K710 table?