soco-device 0.0.5.2

Log memory usage and auto choose device for machine learning model inference

soco-device

A package for logging, saving, and automatically choosing device during model inference stage.

Usage

1. install package

   pip install soco-device

2. example usage

from soco_device import DeviceCheck

dc = DeviceCheck()
# Return a device name ('cpu'/'cuda') and a list of gpu ids, if any
model_name_or_path = <model name> 
n_gpu_needed = 2
device_name, device_ids = dc.get_device_by_model(model_name_or_path, n_gpu=n_gpu_needed)

# If only single gpu, set device_name as 'cuda:gpu_id'
# For either cpu or multi gpu case, set device_name as 'cpu'/'cuda'
device_name = '{}:{}'.format(device_name, device_ids[0]) if len(device_ids) == 1 else device_name

device = torch.device(device_name)

# ......
# Set up multi gpu training if available
if len(device_ids) > 1:
    model = torch.nn.DataParallel(model, device_ids=device_ids)


# log gpu memory needed in inference stage
dc.log_start()
# ......
# inference model for one time here
dc.log_end()
dc.save(model_name_or_path)