Skip to content

Latest commit

 

History

History
79 lines (70 loc) · 3.25 KB

README_serving.md

File metadata and controls

79 lines (70 loc) · 3.25 KB
Raw

OpenRetro-serving

Serving modules for models in OpenRetro

Environment Setup (CPU only)

First follow the instruction on https://docs.docker.com/engine/install to install the Docker engine. Then either

  • pull pre-built image from ASKCOS docker registry and tag (MLPDS permission required)
docker pull registry.gitlab.com/mlpds_mit/askcos/askcos-data/openretro:serving-cpu-public
docker tag registry.gitlab.com/mlpds_mit/askcos/askcos-data/openretro:serving-cpu-public openretro:serving-cpu

or

  • build from local
docker build -f Dockerfile_serving -t openretro:serving-cpu .

Download Pre-compiled Model Archives

sh scripts_serving/download_prebuilt_archives.sh

Run Docker for Serving Trained Models (CPU only)

sh scripts_serving/gln_serve_in_docker.sh
sh scripts_serving/neuralsym_serve_in_docker.sh
sh scripts_serving/transformer_serve_in_docker.sh
sh scripts_serving/retroxpert_serve_in_docker.sh
sh scripts_serving/localretro_serve_in_docker.sh
sh scripts_serving/retrocomposer_serve_in_docker.sh

By default, any one of the serving scripts would serve all available model archives. Change the arguments after the --models flag (in the script) before running to specify only the desired archives otherwise.

Sample query

curl http://0.0.0.0:port/predictions/MODEL_NAME \
    --header "Content-Type: application/json" \
    --request POST \
    --data '{"smiles": ["[Br:1][CH2:2]/[CH:3]=[CH:4]/[C:5](=[O:6])[O:7][Si:8]([CH3:9])([CH3:10])[CH3:11]", "CC(C)(C)OC(=O)N1CCC(OCCO)CC1"]}'

This query format should work for any served model. Specifically,
  0.0.0.0: the server ip or server name
  port: the port of served model. By default, it is 9917, 9927, 9937, 9947, 9957, 9967 for GLN, NeuralSym, Transformer, RetroXpert, LocalRetro, and RetroComposer respectively.
  MODEL_NAME: any of USPTO_50k_gln, {USPTO_50k,USPTO_full,pistachio_21Q1}_neuralsym, {USPTO_50k,USPTO_full,pistachio_21Q1}_transformer, {USPTO_50k,USPTO_full}_retroxpert, USPTO_50k_localretro, USPTO_50k_retrocomposer.   --data： a single json dict with "smiles" as the key, and list of SMILES as the value. It is okay to pass in SMILES with atom mapping numbers. All models would remove atom mapping numbers and canonicalize the target SMILES before inference.

Sample return

  • For NeuralSym, GLN and RetroComposer
List[{
    "templates": List[str], list of top k templates,
    "reactants": List[str], list of top k proposed reactants based on the templates,
    "scores": List[float], list of top k corresponding scores
}]

Note that the reactants may be duplicated since different templates can give the same reactants.

  • For Transformer, RetroXpert and LocalRetro
List[{
    "reactants": List[str], list of top k proposed reactants,
    "scores": List[float], list of top k corresponding scores
}]

[Optional] Create own model archives

If you want to create servable model archives from own checkpoints (e.g. trained on different datasets), please refer to the archiving scripts (scripts_serving/{gln,neuralsym,transformer,retroxpert,localretro}_archive_model_in_docker.sh). Change the arguments accordingly in the script before running. It's mostly bookkeeping by replacing the data name and/or checkpoint paths; the script should be self-explanatory.