PIMmodels

This repository includes PIM acceleratable models.
You can inference models on PIM SDK environment which is provided as docker.

PimAiCompiler supporting models

Below models are currently supported through PimAiCompiler.
Prerequisites : PimAiCompiler installation provided by PIM SDK is essential.

No	Model	Reference
1	RNNT (RNN Transducer)	link
2	GNMT (Google Neural Machine Translation)	link
3	HWR (Hand-Written text Recognition)	link

Each model folder has run.sh bash script for execute model.

Basic options are identical, however model specific options might be existed.

$ ./run.sh --help

usage: ./run.sh [options] [argument]
 Option               Argument
  --accuracy           measure accuracy (default: performance)
  --clean              clean submodule repository
  --use_pim            enable PIM (default: false)

Example for executing model

When you want to measure accuracy with enabled PIM, you can execute script as following:

$ cd (gnmt|rnnt|hwr ...)
$ ./run.sh --accuracy --use_pim

When you want to measure performance (end-to-end latency) with enabled PIM, you can execute script as following:

$ cd (gnmt|rnnt|hwr ...)
$ ./run.sh --use_pim

Known issues

Model	Issues	Status
GNMT	Occasionally the output of GPU-PIM is inaccuracte in comparison with the output of GPU.	In progress