Configs and overrides

An experiment config is a Python file that can inherit from others; a launch config is YAML plus flags. Both have a fixed precedence.

Experiment configs

# configs/my_study/semseg.py
_base_ = ["../panda/semseg/semseg-pt-v3m2-pilarnet-ft-5cls-fft.py"]

seed = 7
batch_size = 16
optimizer = dict(lr=3e-5)

_base_ paths are relative to the file. Dicts merge recursively with the base, while scalars and lists replace it; redefining hooks, transform, criteria or param_dicts drops the inherited list.

Overrides on the command line

Launcher flags come before a bare --; experiment overrides come after it as key=value tokens:

uv run pimm launch --train.config my_study/semseg --resources.nproc-per-node 4 --run.name semseg-seed7 \
  -- batch_size=16 optimizer.lr=3e-5 data.train.max_len=100000

Dotted keys reach into nested dicts, and values are parsed like YAML. Tokens after -- that start with -- are rejected.

Precedence

experiment:  base config(s)  →  child config  →  overrides after --
launch:      launch/defaults.yaml  →  launch/sites/<site>.yaml  →  --recipe YAML  →  flags

The run directory records the result in config.py and resolved_config.json.

Common experiment fields

Defaults come from configs/_base_/default_runtime.py; recipes override many of them.

Field	Default	Meaning
weight	None	weights to load at start: a path or hf://org/repo
resume	False	restore trainer state as well as weights
evaluate	True	run validation during training
seed	None	random seed; drawn and recorded when None
batch_size	16	events per step, across all GPUs
batch_size_val, batch_size_test	None	across all GPUs; None means one event per GPU
num_worker	16	data-loader workers, across all GPUs
epoch	800	total passes over the training data
eval_epoch	800	rounds the training is split into; validation and checkpoints follow each round
clip_grad	None	gradient-norm clipping threshold
enable_amp, amp_dtype	False, "bfloat16"	mixed precision
param_dicts	None	learning-rate scales for parameter groups, for example [dict(keyword="block", lr_scale=0.1)]
hooks	see How pimm works	ordered lifecycle hooks
train, test	DefaultTrainer, SemSegTester	trainer and tester
structured_logging	disabled	per-rank traces; see Monitor and debug
use_wandb	set by recipes	log to W&B when True, TensorBoard otherwise
checkpoint_format	set by recipes	"standard" (split) or "legacy" (single file)
parallel	DDP	dict(strategy="ddp"), or the experimental "fsdp2"

Launcher flags are listed in Command line.

Check before running

--dry-run resolves both configs and prints what would run, without building the model or opening data:

uv run pimm launch --train.config my_study/semseg --dry-run