Loading Models Without Disk Storage Using InMemoryModelStore¶

Introduction¶

This tutorial introduces the InMemoryModelStore class, an alternative to ModelStore designed for environments with limited disk space. This approach is particularly beneficial when memory is more readily available than disk space or disk access is limited or slow.

InMemoryModelStore enables model loading directly into memory, bypassing the need for local storage and disk I/O operations.

Using InMemoryModelStore¶

Here's how to use InMemoryModelStore:

In [ ]:

from molfeat.store import InMemoryModelStore

# Initialize the in-memory store
store = InMemoryModelStore(model_store_root="s3://my-modelstore-bucket")

# Load a model directly into memory
model, model_info = store.load("My-Model")

from molfeat.store import InMemoryModelStore # Initialize the in-memory store store = InMemoryModelStore(model_store_root="s3://my-modelstore-bucket") # Load a model directly into memory model, model_info = store.load("My-Model")

The model is now ready for use without any disk I/O overhead

In [ ]:

import datamol as dm
from molfeat.trans.pretrained.hf_transformers import PretrainedHFTransformer

smiles = dm.freesolv().iloc[:100].smiles

transformer = PretrainedHFTransformer(kind=model, notation=model_info.inputs, dtype=float)
features = transformer(smiles)

import datamol as dm from molfeat.trans.pretrained.hf_transformers import PretrainedHFTransformer smiles = dm.freesolv().iloc[:100].smiles transformer = PretrainedHFTransformer(kind=model, notation=model_info.inputs, dtype=float) features = transformer(smiles)

Benefits and Considerations¶

• Reduced Latency: By loading models directly into memory, you eliminate the time needed for disk reads.
• Efficient Resource Use: Ideal for serverless environments where disk access might be limited or costly.

However, keep in mind that this approach requires sufficient memory to hold the entire model. Ensure your deployment environment has adequate RAM for your model size.