model-loads 0.2.4

Loads GPU or CPU pytorch models

model_loads is an open-source Python package for pytorch load models easy.

PyTorch is a Python package that provides two high-level features:

• Tensor computation (like NumPy) with strong GPU acceleration
• Deep neural networks built on a tape-based autograd system

It's annoying to load cpu model to gpu devices or load multi-gpus trained model to single gpu devices sometimes, And this package try to simplify it.

Table of Contents

• Table of Contents
• Installation
• Getting Started
• TODO list

Installation

To install load_models, you can do as follow:

    pip install model-loads

Or from source

    git clone https://github.com/cwh94/model_loads.git
    cd load_models
    python setup.py bdist_egg
    python setup.py install

Getting Started

1. load pth model to GPU device

import model_loads as lo
import torchvision.models as models

model = models.MobileNetV2()
model_path = "../examples/models/pth/mobilenet_v2-b0353104.pth"
model, _ = lo.load_models(model_path, model, use_gpu=True)
print(model)
print(type(model))

2. load tar model(which contains state_dict and optimization info or accuracy) to CPU device

from models.tar.mobilenet_v2 import MobileNetV2

model = MobileNetV2()
model_path = "models/tar/checkpoint.pth.tar"

model, other_param = lo.load_models(model_path, model)
print(model)
print(other_param)

4. load model to CPU device

import os

os.environ["CUDA_VISIBLE_DEVICES"] = ""

model = models.MobileNetV2()
model_path = "models/pth/mobilenet_v2-b0353104.pth"
model, _ = lo.load_models(model_path, model, use_gpu=True)
print(model)
print(type(model))

Done:

load model which save model by

torch.save(model, "path/to/model")

Todo list

DATA PARALLELISM