Train and pretrain#

Pick a recipe, check it on a few events, then start the full run.

Pick a recipe#

uv run pimm ls prints every recipe; uv run pimm ls panda semseg lists one directory. The recipe catalog shows each recipe’s model, data, warm start, epochs and batch size.

Goal	Recipe
Pretrain Panda	`panda/pretrain/pretrain-sonata-v1m1-pilarnet-smallmask`
Panda semantic segmentation	`panda/semseg/semseg-pt-v3m2-pilarnet-ft-5cls-fft`
Panda particle detector	`panda/panseg/detector-v5-pt-v3m2-ft-pid-fft`
Panda interaction detector	`panda/panseg/detector-v5-pt-v3m2-ft-vtx-fft`
Pretrain PoLAr-MAE	`polarmae/pretrain-polarmae-pilarnet`
PoLAr-MAE semantic segmentation	`polarmae/semseg/semseg-polarmae-pilarnet-fft`
JAXTPC semantic segmentation	`detector/semseg/semseg-pt-v3m2-jaxtpc-5cls`

To fine-tune with a recipe, pass the starting weights with --train.weight; see Fine-tune.

Each recipe names a PILArNet-M revision. The reader accepts v3 and v3_extra; to run a recipe that names v1, point each split at v3 with an override, such as data.train.revision=v3 data.val.revision=v3.

Check it on a few events#

uv run pimm launch \
  --train.config panda/pretrain/pretrain-sonata-v1m1-pilarnet-smallmask \
  --resources.nproc-per-node 1 \
  --run.name sonata-smoke \
  -- epoch=1 data.train.max_len=32 data.val.max_len=16 batch_size=4 num_worker=0 use_wandb=False \
     data.train.revision=v3 data.val.revision=v3

This needs PILArNet-M v3 on disk (see Supported detectors). Run it once with --dry-run first. When it finishes, model/last/.complete exists in the run directory.

Keep a variant as a child config#

# configs/my_study/sonata_v1.py
_base_ = ["../panda/pretrain/pretrain-sonata-v1m1-pilarnet-smallmask.py"]

seed = 17
batch_size = 32
epoch = 100
optimizer = dict(lr=3e-5, weight_decay=0.2)
data = dict(train=dict(revision="v3"), val=dict(revision="v3"))

Dicts merge with the base; lists replace it. Redefining hooks or transform replaces the whole list.

Start the run#

uv run pimm launch --train.config my_study/sonata_v1 --resources.nproc-per-node 4 --run.name sonata-v1

The launcher appends a timestamp to the run name and copies pimm/ and configs/ into the run. Pass --run.no-timestamp for a fixed directory you can resume, and --train.no-code-copy to train from the checkout instead of the copy.

From a notebook#

Start training as a child process so it gets the launcher’s setup:

!uv run pimm launch --resources.nproc-per-node 1 --train.config tests/tiny_semseg --run.name notebook-smoke

A local launch stays attached to the notebook kernel; use pimm submit for a run that should outlive it. Loading data, transforms and models for inference works in ordinary notebook cells.