s3dis_with_stuff_11g.yaml

# @package _global_

# to execute this experiment run:
# python train.py experiment=panoptic/s3dis_with_stuff_11g

# This configuration allows training SPT on a single 11G GPU, with a
# training procedure comparable with the default
# experiment/semantic/s3dis configuration.
# Among the multiple ways of reducing memory impact, we choose here to
#   - divide the dataset into smaller tiles (facilitates preprocessing
#     and inference on smaller GPUs)
#   - reduce the number of samples in each batch (facilitates training
#     on smaller GPUs)
# To keep the total number of training steps consistent with the default
# configuration, while keeping informative gradient despite the smaller
# batches, we use gradient accumulation and reduce the number of epochs.
# DISCLAIMER: the tiling procedure may increase the preprocessing time
# (more raw data reading steps), and slightly reduce mode performance
# (less diversity in the spherical samples)

defaults:
  - override /datamodule: panoptic/s3dis_with_stuff.yaml
  - override /model: panoptic/spt-2.yaml
  - override /trainer: gpu.yaml

# all parameters below will be merged with parameters from default configurations set above
# this allows you to overwrite only specified parameters

datamodule:
  xy_tiling: 3  # split each cloud into xy_tiling^2=9 tiles, based on a regular XY grid. Reduces preprocessing- and inference-time GPU memory
  sample_graph_k: 2  # 2 spherical samples in each batch instead of 4. Reduces train-time GPU memory

trainer:
  max_epochs: 500  # to keep same nb of steps: 8x more tiles, 2-step gradient accumulation -> epochs/4

model:
  optimizer:
    lr: 0.1
    weight_decay: 1e-2

  partitioner:
    regularization: 10
    x_weight: 5e-2
    cutoff: 300

  partition_every_n_epoch: 5

logger:
  wandb:
    project: "spt_s3dis"
    name: "SPT-64"

# metric based on which models will be selected
optimized_metric: "val/pq"

# modify checkpointing callbacks to adapt to partition_every_n_epoch
# being potentially different
callbacks:
  model_checkpoint:
    every_n_epochs: ${eval:'max(${trainer.check_val_every_n_epoch}, ${model.partition_every_n_epoch})'}

  early_stopping:
    strict: False

  gradient_accumulator:
    scheduling:
      0:
        2  # accumulate gradient every 2 batches, to make up for reduced batch size