advanced_example_config.yaml

# Experiment configuration file.
#
# There are two special blocks. The 'seml' block is required for every experiment.
# It has to contain the following values:
# executable:        Name of the Python script containing the experiment. The path should be relative to the `project_root_dir`.
#                    For backward compatibility SEML also supports paths relative to the location of the config file.
#                    In case there are files present both relative to the project root and the config file,
#                    the former takes precedence.
# It can optionally also contain the following values:
# name:              Prefix for output file and Slurm job name. Default: Collection name
# output_dir:        Directory to store log files in. Default: Current directory
# conda_environment: Specifies which Anaconda virtual environment will be activated before the experiment is executed.
#                    Default: The environment used when queuing.
# project_root_dir:  (Relative or absolute) path to the root of the project. seml will then upload all the source
#                    files imported by the experiment to the MongoDB. Moreover, the uploaded source files will be
#                    downloaded before starting an experiment, so any changes to the source files in the project
#                    between queueing and starting the experiment will have no effect.
#
# The special 'slurm' block contains the slurm parameters. This block and all values are optional. Possible values are:
# experiments_per_job:   Number of parallel experiments to run in each Slurm job.
#                        Note that only experiments from the same batch share a job. Default: 1
# max_simultaneous_jobs: Maximum number of simultaneously running Slurm jobs per job array. Default: No restriction
# sbatch_options:        dictionary that contains custom values that will be passed to `sbatch`, specifying e.g.
#                        the memory and number of GPUs to be allocated (prepended dashes are not required). See
#                        https://slurm.schedmd.com/sbatch.html for all possible options.
#
# Parameters under 'fixed' will be used for all the experiments.
#
# Under 'grid' you can define parameters that should be sampled from a regular grid. Options are:
#   - choice:     List the different values you want to evaluate under 'choices' as in the example below.
#   - range:      Specify the min, max, and step. Parameter values will be generated using np.arange(min, max, step).
#   - uniform:    Specify the min, max, and num. Parameter values will be generated using
#                 np.linspace(min, max, num, endpoint=True)
#   - loguniform: Specify min, max, and num. Parameter values will be uniformly generated in log space (base 10).
# Additionally, one may supply the 'zip_id' argument to zip multiple parameters to a single dimension.
# This causes these parameters to only change jointly. All parameters within a group must have the same number of options.
#
#
# Under 'random' you can specify parameters for which you want to try several random values. Specify the number
# of samples per parameter with the 'samples' value as in the examples below.
# Specify the the seed under the 'random' dict or directly for the desired parameter(s).
# Supported parameter types are:
#   - choice:      Randomly samples <samples> entries (with replacement) from the list in parameter['options']
#   - uniform:     Uniformly samples between 'min' and 'max' as specified in the parameter dict.
#   - loguniform:  Uniformly samples in log space between 'min' and 'max' as specified in the parameter dict.
#   - randint:     Randomly samples integers between 'min' (included) and 'max' (excluded).
#
# The configuration file can be nested (as the example below) so that we can run different parameter sets
# e.g. for different datasets or models.
# We take the cartesian product of all `grid` parameters on a path and sample all random parameters on the path.
# The number of random parameters sampled will be max{n_samples} of all n_samples on the path. This is done because
# we need the same number of samples from all random parameters in a configuration.
#
# More specific settings (i.e., further down the hierarchy) always overwrite more general ones.

seml:
  executable: advanced_example_experiment.py
  name: advanced_example_experiment
  output_dir: logs
  project_root_dir: .
  description: "An advanced example configuration. We can also use variable interpolation here: ${config.model.model_type}"
  reschedule_timeout: 300 # The time (in seconds) that are left on the job before SEML will try to reschedule unfinished experiments.
  # Note that you have to implement a `reschedule_hook` to use this feature.

slurm:
  - experiments_per_job: 1
    max_simultaneous_jobs: 4 # Restrict number of simultaneously running jobs per job array
    sbatch_options:
      gres: gpu:1 # num GPUs
      mem: 16G # memory
      cpus-per-task: 2 # num cores
      time: 0-08:00 # max time, D-HH:MM
      partition: gpu_gtx1080 # use the 1080ti partition
  # We can also increase the number of jobs if we run on A100s
  - experiments_per_job: 4
    max_simultaneous_jobs: 4 # Restrict number of simultaneously running jobs per job array
    sbatch_options:
      gres: gpu:1 # num GPUs
      mem: 16G # memory
      cpus-per-task: 8 # num cores
      time: 0-08:00 # max time, D-HH:MM
      partition: gpu_a100 # Use A100

###### BEGIN PARAMETER CONFIGURATION ######

fixed:
  training.patience: 20
  training.num_epochs: 100
  optimization:
    optimizer_type: Adam
  +batchnorm.priority: 2

grid:
  # SEML supports dot-notation for nested parameter dictionaries. E.g., `model.model_type` resolves to
  # {'model': {'model_type': xxx}}
  model.model_type:
    type: choice
    options:
      - variant_1
      - variant_2
    zip_id: model_dataset # The model_type jointly changes with all other parameters of the mode_dataset group

  # The `+` prefix indicates that a named config should be used. The `name` attribute selects which named config is run
  # Here, setting `batchnorm.name` to `batchnorm`, will run the named config `batchnorm` (see `advanced_example_experiment.py`)
  # Setting `batchnorm.priority` to `2`, will affect the order in which the named configs are loaded
  +batchnorm.name:
    type: choice
    options:
      - batchnorm
      - no_batchnorm

  # Instead of setting the `name` attribute, you can equivalently set `preprocessing` to a the corresponding named config.
  # seml will infer to set the `name` attribute for you
  +preprocessing:
    type: choice
    options:
      - preprocessing_none
      - preprocessing_normalize

  # You can also specify named configs from file paths, e.g. yaml files. These paths are expected to be relative to `project_root_dir`.
  +augmentation:
    type: choice
    options:
      - config/flip_augmentation.yaml

  optimization.regularization.weight_decay:
    type: loguniform
    min: 1e-6
    max: 1e-3
    num: 4

random:
  samples: 5
  seed: 7059
  model.model_params.dropout:
    type: uniform
    min: 0.0
    max: 0.5

large_datasets:
  grid:
    data.dataset:
      type: choice
      options:
        - large_dataset_1
        - large_dataset_2
      zip_id: model_dataset # Use variant_1 for large_dataset_1 and variant_2 for large_dataset_2

    model.model_params.hidden_sizes:
      type: choice
      options:
        - [64]
        - [64, 32]

small_datasets:
  grid:
    data.dataset:
      type: choice
      options:
        - small_dataset_1
        - small_dataset_2
      zip_id: model_dataset # Use variant_1 for small_dataset_1 and variant_2 for small_dataset_2

    model.model_params.hidden_sizes:
      type: choice
      options:
        - [32]
        - [32, 16]