Skip to content

Training Models

Pydantic Configuration

BioNeMo 2 provides two entrypoints for models with both argparse and pydantic. Both documented in the Models section below. Pydantic based configuration is designed to accept a configuration yaml file as input, along with context specific arguments (e.g., should we resume from existing checkpoints?). These YAML configs go through a Pydantic Validator, in this case referred to as MainConfig. This Config is composed of several other Pydantic models, see the class definition for details. To pre-populate a config with reasonable defaults for various standard models, we provide 'recipes.' These are simple methods that instantiate the config object and then serialize it to a YAML configuration file. From this file, you may either submit it directly, or modify the various parameters to meet your usecase. For example, Weights and biases, devices, precision, and dataset options are all extremely useful to modify. Then, you would submit this config for training.

These two workflows are packaged as executables when esm2 or geneformer are installed with pip. These commands will appear as:

bionemo-geneformer-recipe
bionemo-esm2-recipe
bionemo-geneformer-train
bionemo-esm2-train

ESM-2

Running

First off, we have a utility function for downloading full/test data and model checkpoints called download_bionemo_data that our following examples currently use. This will download the object if it is not already on your local system, and then return the path either way. For example if you run this twice in a row, you should expect the second time you run it to return the path almost instantly.

NOTE: NVIDIA employees should use pbss rather than ngc for the data source.

export MY_DATA_SOURCE="ngc"

or for NVIDIA internal employees with new data etc:

export MY_DATA_SOURCE="pbss"

# The fastest transformer engine environment variables in testing were the following two
TEST_DATA_DIR=$(download_bionemo_data esm2/testdata_esm2_pretrain:2.0 --source $MY_DATA_SOURCE); \
ESM2_650M_CKPT=$(download_bionemo_data esm2/650m:2.0 --source $MY_DATA_SOURCE); \

train_esm2     \
    --train-cluster-path ${TEST_DATA_DIR}/2024_03_sanity/train_clusters_sanity.parquet     \
    --train-database-path ${TEST_DATA_DIR}/2024_03_sanity/train_sanity.db     \
    --valid-cluster-path ${TEST_DATA_DIR}/2024_03_sanity/valid_clusters.parquet     \
    --valid-database-path ${TEST_DATA_DIR}/2024_03_sanity/validation.db     \
    --result-dir ./results     \
    --experiment-name test_experiment     \
    --num-gpus 1  \
    --num-nodes 1 \
    --val-check-interval 10 \
    --num-dataset-workers 1 \
    --num-steps 10 \
    --max-seq-length 1024 \
    --limit-val-batches 2 \
    --micro-batch-size 2 \
    --restore-from-checkpoint-path ${ESM2_650M_CKPT}

Running with Pydantic configs

Alternatively, we provide a validated and serialized configuration file entrypoint for executing the same workflow. These can be generated using the bionemo-esm2-recipe entrypoints. Recipes are available for 8m, 650m, and 3b ESM2 models. You may select which preset config to use by setting the --recipe parameter. The output is then a serialized configuration file that may be used in the associated bionemo-esm2-train commands.

# The fastest transformer engine environment variables in testing were the following two
TEST_DATA_DIR=$(download_bionemo_data esm2/testdata_esm2_pretrain:2.0 --source $MY_DATA_SOURCE); \
bionemo-esm2-recipe \
--train-cluster-path ${TEST_DATA_DIR}/2024_03_sanity/train_clusters_sanity.parquet     \
--train-database-path ${TEST_DATA_DIR}/2024_03_sanity/train_sanity.db     \
--valid-cluster-path ${TEST_DATA_DIR}/2024_03_sanity/valid_clusters.parquet     \
--valid-database-path ${TEST_DATA_DIR}/2024_03_sanity/validation.db     \
--result-dir ./results     \
--dest my_config.yaml\
--recipe esm2_8m_recipe

⚠️ IMPORTANT: Inspect and edit the contents of the outputted my_config.yaml as you see fit

NOTE: To continue training from an existing checkpoint, simply pass in the path --initial-ckpt-path to the recipe command. This will populate the YAML with the correct field to ensure pretraining is initialized from an existing checkpoint.

To submit a training job with the passed config, first update the yaml file with any additional execution parameters of your choosing: number of devices, workers, steps, etc. Second, invoke our training entrypoint. To do this, we need three things:

  • Configuration file, the YAML produced by the previous step
  • Model config type, in this case the pretraining config. This will validate the arguments in the config YAML against those required for pretraining. Alternatively, things like fine-tuning with custom task heads may be specified here. This allows for mixing/matching Data Modules with various tasks.
  • Data Config type, this specifies how to parse, validate, and prepare the DataModule. This may change depending on task, for example, pretraining ESM2 uses a protein cluster oriented sampling method. In the case of inference or fine-tuning a pretrained model, a simple fasta file may be sufficient. There is a one-to-one relationship between DataConfig types and DataModule types.

⚠️ Warning: This setup does NO configuration of Weights and Biases. Edit your config YAML and populate it with your WandB details.

bionemo-esm2-train \
--data-config-cls bionemo.esm2.run.config_models.ESM2DataConfig \
--model-config-cls bionemo.esm2.run.config_models.ExposedESM2PretrainConfig \
--config my_config.yaml

NOTE: both data-config-cls and model-config-cls have default values corresponding to ESM2DataConfig and ExposedESM2PretrainingConfig

DataConfigCls and ModelConfigCls can also refer to locally defined types by the user. As long as python knows how to import the specified path, they may be configured. For example, you may have a custom Dataset/DataModule that you would like to mix with an existing recipe. In this case, you define a DataConfig object with the generic specified as your DataModule type, and then pass in the config type to the training recipe.

Geneformer

Running

Similar to ESM-2, you can download the dataset and checkpoint through our utility function.

TEST_DATA_DIR=$(download_bionemo_data single_cell/testdata-20241203 --source $MY_DATA_SOURCE); \
GENEFORMER_10M_CKPT=$(download_bionemo_data geneformer/10M_240530:2.0 --source $MY_DATA_SOURCE); \
train_geneformer     \
    --data-dir ${TEST_DATA_DIR}/cellxgene_2023-12-15_small_processed_scdl    \
    --result-dir ./results     \
    --restore-from-checkpoint-path ${GENEFORMER_10M_CKPT} \
    --experiment-name test_experiment     \
    --num-gpus 1  \
    --num-nodes 1 \
    --val-check-interval 10 \
    --num-dataset-workers 0 \
    --num-steps 55 \
    --seq-length 128 \
    --limit-val-batches 2 \
    --micro-batch-size 2

To fine-tune, you to specify a different combination of model and loss. Pass the path to the outputted config file from the previous step as the --restore-from-checkpoint-path, and also change --training-model-config-class to the newly created model-config-class.

While no CLI option currently exists to hot swap in different data modules and processing functions now, you could copy the sub-projects/bionemo-geneformer/geneformer/scripts/train_geneformer.py and modify the DataModule class that gets initialized.

Simple fine-tuning example (NOTE: please change --restore-from-checkpoint-path to be the checkpoint directory path that was output last by the previous train run)

TEST_DATA_DIR=$(download_bionemo_data single_cell/testdata-20241203 --source $MY_DATA_SOURCE); \
train_geneformer     \
    --data-dir ${TEST_DATA_DIR}/cellxgene_2023-12-15_small_processed_scdl    \
    --result-dir ./results     \
    --experiment-name test_finettune_experiment     \
    --num-gpus 1  \
    --num-nodes 1 \
    --val-check-interval 10 \
    --num-dataset-workers 0 \
    --num-steps 55 \
    --seq-length 128 \
    --limit-val-batches 2 \
    --micro-batch-size 2 \
    --training-model-config-class FineTuneSeqLenBioBertConfig \
    --restore-from-checkpoint-path results/test_experiment/dev/checkpoints/test_experiment--val_loss=4.3506-epoch=1-last

Running with Pydantic configs

Alternatively, we provide a validated and serialized configuration file entrypoint for executing the same workflow. Recipes are available for 10m, and 106m geneformer models. Additionally we provide an example recipe of finetuning, where the objective is to 'regress' on token IDs rather than the traditional masked language model approach. In practice, you will likely need to implement your own DataModule, DataConfig, and Finetuning model. You can use the same overall approach, but with customizations for your task.

TEST_DATA_DIR=$(download_bionemo_data single_cell/testdata-20241203 --source $MY_DATA_SOURCE); \
bionemo-geneformer-recipe \
    --recipe 10m-pretrain \
    --dest my_config.json \
    --data-path ${TEST_DATA_DIR}/cellxgene_2023-12-15_small_processed_scdl \
    --result-dir ./results

⚠️ IMPORTANT: Inspect and edit the contents of the outputted my_config.yaml as you see fit

NOTE: To pretrain from an existing checkpoint, simply pass in the path --initial-ckpt-path to the recipe command. This will populate the YAML with the correct field to ensure pretraining is initialized from an existing checkpoint.

To submit a training job with the passed config, first update the yaml file with any additional execution parameters of your choosing: number of devices, workers, steps, etc. Second, invoke our training entrypoint. To do this, we need three things:

  • Configuration file, the YAML produced by the previous step
  • Model config type, in this case the pretraining config. This will validate the arguments in the config YAML against those required for pretraining. Alternatively, things like fine-tuning with custom task heads may be specified here. This allows for mixing/matching Data Modules with various tasks.
  • Data Config type, this specifies how to parse, validate, and prepare the DataModule. This may change depending on task, for example, while fine-tuning you may want to use a custom Dataset/DataModule that includes PERTURB-seq. In this case, the default pretraining DataConfig and DataModule will be insufficient. See ESM2 for additional example usecases.

⚠️ Warning: This setup does NO configuration of Weights and Biases. Edit your config YAML and populate it with your WandB details.

bionemo-geneformer-train \
--data-config-cls bionemo.geneformer.run.config_models.GeneformerPretrainingDataConfig \
--model-config-cls bionemo.geneformer.run.config_models.ExposedGeneformerPretrainConfig \
--config my_config.yaml

NOTE: both data-config-cls and model-config-cls have default values corresponding to GeneformerPretrainingDataConfig and ExposedGeneformerPretrainConfig

DataConfigCls and ModelConfigCls can also refer to locally defined types by the user. As long as python knows how to import the specified path, they may be configured. For example, you may have a custom Dataset/DataModule that you would like to mix with an existing recipe. In this case, you define a DataConfig object with the generic specified as your DataModule type, and then pass in the config type to the training recipe.