Skip to content

Pre-training CT-FM

Before you begin, ensure you have downloaded your data as explained in the Data Instructions. It is also a good idea to review the lighter documentation since our training configurations are based on its guidelines.

Pre-training Experiment Configurations

Pre-training configuration files are organized in the experiments directory. Explore the key folders below to find the setup that best meets your needs:

  • Ablations Setups for testing various experimental approaches.

  • FM
    Finalized configurations for the pre-training run.

Running the pretraining

After all adjustments have been made, navigate to the root directory of the CT-FM project and execute the following command to begin pre-training:

lighter fit --config=./experiments/fm/base.yaml,\ #(1)!
./experiments/fm/frameworks/intrasample_simclr.yaml,\ #(2)!
./experiments/fm/backbones/segresenc.yaml #(3)!

  1. The file establishes the core settings for pre-training the CT-FM model by defining:

    • Variables: Core parameters such as voxel spacing.
    • Trainer Settings: Parameters including 500 epochs, batch limits, GPU configuration, mixed precision, logging via WandB, and checkpoint callbacks.
    • System Settings: The model placeholder, optimizer (AdamW), learning rate scheduler (WarmupCosineSchedule), and dataloader setup for safely handling your dataset.
    • Adapters: Methods for batch processing and loss computation. In essence, base.yaml serves as the foundation upon which the entire pre-training process is built.

  2. The file configures the self-supervised SimCLR framework used during pre-training. It includes:

    • Model & Criterion: Defines the CT-FM model and applies a contrastive loss function with a specified temperature.
    • Data Augmentation Pipeline: Implements a series of transformations (such as random crops, flips, and intensity adjustments) to generate multiple augmented views from each input image. This configuration augments the base setup with specialized self-supervised learning components.

  3. The file sets up the backbone for the CT-FM model. It includes:

    • Backbone Identification: Sets the variable BACKBONE_NAME to "SegResNetDS".
    • Architectural Details: Configures the SegResNet encoder (via monai.networks.nets.segresnet_ds.SegResEncoder) by specifying parameters like spatial dimensions, input channels, initial filters, and block structures.
    • Integration with Base Config: Uses shared variable mappings (such as NUM_FTRS_BY_BACKBONE) and logger identifiers from base.yaml to ensure smooth integration. This configuration provides the essential backbone architecture for complete model training.

Click on the symbols to learn more about each yaml file

Customization Before Training

Before running the experiment, update your base.yaml configuration using the guidelines below:

Directory Paths Update

  • Set the paths for save_dir and dirpath to your preferred locations for saving logs and checkpoints.
  • Update the path for scan_list.pkl to reflect the file produced during the data preparation phase.

Training Parameter Adjustments

Modify the settings under the trainer: key (such as the number of GPUs, batch size, and training duration) to align with your system’s resources and experimental needs.

After applying these customizations, execute the pre-training command to initiate the process with your updated configurations.