Skip to content

Tour the config (optional)

Ten minutes with the example project, no dataset needed: how a config is checked, what a run uses, and how a mistake is caught before anything is built. The example is a tiny untrained detector, wired like a real method.

1. Validate a config

grasppanda print-config projects/example_project/configs/toy_detector.yaml \
    --check

✓ projects/example_project/configs/toy_detector.yaml is valid
  inherits    default_runtime.yaml, adam_10e.yaml
  components  optimizer: Adam · scheduler: ExponentialLR · model: ToyDetector · model.backbone: ToyBackbone · model.grouper: ToyGrouper · model.head: ToyHead
  hash        53bed832c67f

The config inherits two shipped building blocks: the runtime defaults and a plain training recipe (_base_/schedules/adam_10e.yaml: Adam, a learning rate of 1e-3, 10 epochs).

2. See what a run would use

grasppanda print-config projects/example_project/configs/toy_detector.yaml \
    --defaults

It prints the config after inheritance, overrides and interpolation, with every default written out. This is exactly what a run stores as config.resolved.yaml.

The output 
train:
  amp: false
  accumulate: 1
  epochs: 10
  grad_clip: null
  grad_clip_mode: norm
  ema: null
  sync_bn: false
  scheduler_unit: epoch
evaluation:
  interval: 1
  variants:
    raw: []
  evaluator: null
  save_best: null
  split: test
  weights: model
  num_workers: 0
runtime:
  work_dir: work_dirs/toy_detector
  seed: 0
  deterministic: false
  checkpoint:
    interval: 1
    keep_last: 3
  log:
    interval: 50
    backends:
    - text
  run_id: null
  device: auto
  resume: false
  allow_config_mismatch: false
  load_from: null
  allow_weight_mismatch: false
  ops: {}
optimizer:
  type: Adam
  lr: 0.001
  weight_decay: 0.0
  betas:
  - 0.9
  - 0.999
  eps: 1.0e-08
  amsgrad: false
  foreach: null
  maximize: false
  capturable: false
  differentiable: false
  fused: null
  decoupled_weight_decay: false
scheduler:
  type: ExponentialLR
  gamma: 0.95
  last_epoch: -1
custom_imports:
- ../toy_components.py
model:
  type: ToyDetector
  num_seeds: 64
  backbone:
    type: ToyBackbone
    out_channels: 64
    in_channels: 3
  grouper:
    type: ToyGrouper
    radius: 0.05
    nsample: 16
  head:
    type: ToyHead
    num_angle: 12
    num_depth: 4
    depth_step: 0.01
    max_width: 0.1
    width_weight: 10.0
data: null
hooks: []

3. Make a mistake on purpose

A misspelt key is an error that names the closest parameter:

grasppanda print-config projects/example_project/configs/toy_detector.yaml \
    --opts model.head.num_angles=6

✗ 1 config problem in projects/example_project/configs/toy_detector.yaml

  model.head.num_angles
    ToyHead has no parameter 'num_angles'. Did you mean 'num_angle'? Parameters: num_angle, num_depth, depth_step, max_width, width_weight, in_channels.
    set in: <command line>

A swap that cannot work is refused the same way, before anything is built:

grasppanda print-config projects/example_project/configs/toy_detector.yaml \
    --opts model.grouper.type=ToyDepthGrouper

✗ 1 config problem in projects/example_project/configs/toy_detector.yaml

  model
    ToyDepthGrouper emits features with axes ('seed', 'depth', 'channel'), but ToyHead expects ('seed', 'channel'). Use a grouper and a head that agree; a depth axis is never added or squeezed implicitly.
    set in: projects/example_project/configs/toy_detector.yaml

Both exit with status 2, so a script can stop on them.

4. Build and run from Python

import torch
from grasppanda.config import Config
from grasppanda.registry import DETECTORS
from grasppanda.structures import Observation, PointSet, SampleKey

config = Config.load("projects/example_project/configs/toy_detector.yaml")
model = DETECTORS.build(config["model"]).eval()
points = PointSet.from_dense(torch.rand(1, 2048, 3))
observation = Observation(keys=[SampleKey("demo", 0, "none", 0)], points=points)
grasps = model.predict(observation)[0]
print(len(grasps), grasps.to_graspnet_array().shape)   # 64 (64, 17)

The toy detector is untrained, so its grasps are arbitrary.

Next