Skip to content

Latest commit

 

History

History
170 lines (142 loc) · 6.76 KB

README.md

File metadata and controls

170 lines (142 loc) · 6.76 KB
Raw

AIDO.ModelGenerator

AIDO.ModelGenerator is a software stack powering the development of an AI-driven Digital Organism by enabling researchers to adapt pretrained models and generate finetuned models for downstream tasks. To read more about AIDO.ModelGenerator's integral role in building the world's first AI-driven Digital Organism, see AIDO.

AIDO.ModelGenerator is open-sourced as an opinionated plug-and-play research framework for cross-disciplinary teams in ML & Bio. It is designed to enable rapid and reproducible prototyping with four kinds of experiments in mind:

  1. Applying pre-trained foundation models to new data
  2. Developing new finetuning and inference tasks for foundation models
  3. Benchmarking foundation models and creating leaderboards
  4. Testing new architectures for finetuning performance

while also scaling with hardware and integrating with larger data pipelines or research workflows.

AIDO.ModelGenerator is built on PyTorch, HuggingFace, and Lightning, and works seamlessly with these ecosystems.

See the AIDO.ModelGenerator documentation for installation, usage, tutorials, and API reference.

Who uses ModelGenerator?

🧬 Biologists

  • Intuitive one-command CLIs for in silico experiments
  • Pre-trained model zoo
  • Broad data compatibility
  • Pipeline-oriented workflows

🤖 ML Researchers

  • Reproducible-by-design experiments
  • Architecture A/B testing
  • Automatic hardware scaling
  • Integration with PyTorch, Lightning, HuggingFace, and WandB

☕ Software Engineers

  • Extensible and modular models, tasks, and data
  • Strict typing and documentation
  • Fail-fast interface design
  • Continuous integration and testing

🤝 Everyone benefits from

  • A collaborative hub and focal point for multidisciplinary work on experiments, models, software, and data
  • Community-driven development
  • Permissive license for academic and non-commercial use

Projects using AIDO.ModelGenerator

Installation

git clone https://github.com/genbio-ai/ModelGenerator.git
cd ModelGenerator
pip install -e .

Source installation is necessary to add new backbones, finetuning tasks, and data transformations, as well as use convenience configs and scripts. If you only need to run inference, reproduce published experiments, or finetune on new data, you can use

pip install modelgenerator
pip install git+https://github.com/genbio-ai/openfold.git@c4aa2fd0d920c06d3fd80b177284a22573528442
pip install git+https://github.com/NVIDIA/dllogger.git@0540a43971f4a8a16693a9de9de73c1072020769

Quick Start

Get embeddings from a pre-trained model

mgen predict --model Embed --model.backbone aido_dna_dummy \
  --data SequencesDataModule --data.path genbio-ai/100m-random-promoters \
  --data.x_col sequence --data.id_col sequence --data.test_split_size 0.0001 \
  --config configs/examples/save_predictions.yaml

Get token probabilities from a pre-trained model

mgen predict --model Inference --model.backbone aido_dna_dummy \
  --data SequencesDataModule --data.path genbio-ai/100m-random-promoters \
  --data.x_col sequence --data.id_col sequence --data.test_split_size 0.0001 \
  --config configs/examples/save_predictions.yaml

Finetune a model

mgen fit --model ConditionalDiffusion --model.backbone aido_dna_dummy \
  --data ConditionalDiffusionDataModule --data.path "genbio-ai/100m-random-promoters"

Evaluate a model checkpoint

mgen test --model ConditionalDiffusion --model.backbone aido_dna_dummy \
  --data ConditionalDiffusionDataModule --data.path "genbio-ai/100m-random-promoters" \
  --ckpt_path logs/lightning_logs/version_X/checkpoints/<your_model>.ckpt

Save predictions

mgen predict --model ConditionalDiffusion --model.backbone aido_dna_dummy \
  --data ConditionalDiffusionDataModule --data.path "genbio-ai/100m-random-promoters" \
  --ckpt_path logs/lightning_logs/version_X/checkpoints/<your_model>.ckpt \
  --config configs/examples/save_predictions.yaml

Configify your experiment

This command

mgen fit --model ConditionalDiffusion --model.backbone aido_dna_dummy \
  --data ConditionalDiffusionDataModule --data.path "genbio-ai/100m-random-promoters"

is equivalent to mgen fit --config my_config.yaml with

# my_config.yaml
model:
  class_path: ConditionalDiffusion
  init_args:
    backbone: aido_dna_dummy
data:
  class_path: ConditionalDiffusionDataModule
  init_args:
    path: "genbio-ai/100m-random-promoters"

Use composable configs to customize workflows

mgen fit --model SequenceRegression --data PromoterExpressionRegression \
  --config configs/defaults.yaml \
  --config configs/examples/lora_backbone.yaml \
  --config configs/examples/wandb.yaml

We provide some useful examples in configs/examples. Configs use the LAST value for each attribute. Check the full configuration logged with each experiment in logs/lightning_logs/your-experiment/config.yaml, or if using wandb logs/config.yaml.

Use LoRA for parameter-efficient finetuning

This also avoids saving the full model, only the LoRA weights are saved.

mgen fit --data PromoterExpressionRegression \
  --model SequenceRegression --model.backbone.use_peft true \
  --model.backbone.lora_r 16 \
  --model.backbone.lora_alpha 32 \
  --model.backbone.lora_dropout 0.1

Use continued pretraining for finetuning domain adaptation

First run pretraining objective on finetuning data

# https://arxiv.org/pdf/2310.02980
mgen fit --model MLM --model.backbone aido_dna_dummy \
  --data MLMDataModule --data.path leannmlindsey/GUE \
  --data.config_name prom_core_notata

Then finetune using the adapted model

mgen fit --model SequenceClassification --model.strict_loading false \
  --data SequenceClassificationDataModule --data.path leannmlindsey/GUE \
  --data.config_name prom_core_notata \
  --ckpt_path logs/lightning_logs/version_X/checkpoints/<your_adapted_model>.ckpt

Make sure to turn off strict_loading to replace the adapter!

Use the head/adapter/decoder that comes with the backbone

mgen fit --model SequenceClassification --data GUEClassification \
  --model.use_legacy_adapter true