Configuration File Guide#

This guide explains how to construct a configuration file to launch a run with stable_ssl. The configuration file is written in YAML format, and various sections map to different configuration classes corresponding to optimization, hardware, log, data and model settings.

Optimization Configuration (optim)#

The optim keyword is used to define the optimization settings for your model. Here’s an example of how to define the optim section in your YAML file:

optim:
  lr: 0.001
  batch_size: 256
  epochs: 500

The complete list of parameters for the optim section, including their descriptions and default values, is provided below:

config.OptimConfig([optimizer, lr, ...])

Configuration for the 'optimizer' parameters.

Hardware Configuration (hardware)#

Use the hardware keyword to configure hardware-related settings such as the number of workers or GPU usage. The complete list of parameters for the hardware section is available below:

config.HardwareConfig([seed, float16, gpu, ...])

Configuration for the 'hardware' parameters.

Log Configuration (log)#

The log keyword configures the logging settings for your run. The complete list of parameters for the log section is provided here:

config.LogConfig([folder, run, load_from, ...])

Configuration for the 'log' parameters.

Data Configuration (data)#

The data keyword defines data loading, preprocessing and data augmentation settings. Here is an example of how to define the data section in your YAML file:

data:
   train_on: base
   base:
      name: CIFAR10
      batch_size: 32
      drop_last: True
      shuffle: True
      split: train
      num_workers: 10
      transforms:
         view1:
         - name: RandomResizedCrop
         kwargs:
            size: 32
            scale:
               - 0.1
               - 0.2
         - name: RandomHorizontalFlip
         - name: SpeckleNoise
         kwargs:
            severity: 2
         p: 0.5
         - name: GaussianBlur
         kwargs:
            kernel_size: 5
         p: 0.2
         view2:
         - name: RandomResizedCrop
         kwargs:
            size: 32
            scale:
               - 0.1
               - 0.2
         - name: RandomHorizontalFlip
   test_in:
      name: CIFAR10
      batch_size: 32
      drop_last: False
      split: train
      num_workers: 10
   test_out:
      name: CIFAR10
      batch_size: 32
      drop_last: False
      num_workers: 10
      split: test

The complete list of parameters for the data section can be found here:

config.DataConfig(train_on, *args, **datasets)

Configuration for multiple datasets used for training the model.

Model Configuration (model)#

The model keyword is used to define the model settings, including the architecture of the backbone, objectives, and more. Below is a list of parameters shared across all models:

config.BaseModelConfig([name, ...])

Base configuration for the 'model' parameters.

When defining a specific method, you can set method-specific parameters by creating a configuration class that inherits from BaseModelConfig. Examples of configurations for different methods in the library are provided below:

`joint_embedding.SimCLRConfig`(name, ...)	Configuration for the SimCLR model parameters.
`joint_embedding.BarlowTwinsConfig`(name, ...)	Configuration for the BarlowTwins model parameters.
`joint_embedding.VICRegConfig`(name, ...)	Configuration for the VICreg model parameters.
`joint_embedding.WMSEConfig`(name, ...)	Configuration for the WMSE model parameters.

Configuration File Guide

Contents