OneFlow-Models

Models and examples implement with OneFlow(version >= 0.5.0).

Introduction

English | 简体中文

OneFlow-Models is an open source repo which contains official implementation of different models built on OneFlow. In each model, we provide at least two scripts train.sh and infer.sh for a quick start. For each model, we provide a detailed README to introduce the usage of this model.

Features

• various models and pretrained weight
• easy use for beginners

Quick Start

Please check our the following demos for a quick start

• image classification quick start lenet demo
• speaker recognition speaker identification demo

Model List

Image Classification

• Lenet
• Alexnet
• VGG16/19
• Resnet50
• InceptionV3
• Densenet
• Resnext50_32x4d
• Shufflenetv2
• MobilenetV2
• mobilenetv3
• Ghostnet
• RepVGG
• DLA
• PoseNet
• Scnet
• Mnasnet
• ViT

Video Classification

• TSN

Object Detection

• CSRNet

Semantic Segmentation

• FODDet
• FaceSeg
• U-Net

Generative Adversarial Networks

• DCGAN
• SRGAN
• Pix2Pix
• CycleGAN

Neural Style Transform

• FastNeuralStyle

Person Re-identification

• BoT

Natural Language Processing

• RNN
• Seq2Seq
• LSTMText
• TextCNN
• Transformer
• Bert
• CPT
• Roberta
• MoE

Audio

• SincNet
• Wav2Letter
• AM_MobileNet1D
• Speech-Emotion-Analyer
• Speech-Transformer
• CycleGAN-VC2
• MaskCycleGAN-VC
• StarGAN-VC
• Adaptive_Voice_Conversion

Deep Reinforcement Learning

• FlappyBird

Quantization Aware Training

• Quantization

Installation and Environment setup

Install Oneflow

https://github.com/Oneflow-Inc/oneflow#install-with-pip-package

Build custom ops from source

In the root directory, run:

mkdir build
cd build
cmake ..
make -j$(nrpoc)

Example of using ops:

from ops import RoIAlign
pooler = RoIAlign(output_size=(14, 14), spatial_scale=2.0, sampling_ratio=2)