English | 简体中文

It introduces how to train InsightFace in OneFlow, and do verification over the validation datasets via the well-toned networks.

- InsightFace in OneFlow

- Contents

- Background

- InsightFace opensource project

- Implementation in OneFlow

- Preparations

- Install OneFlow

- Data preparations

- 1. Download datasets

- 2. Transformation from MS1M recordio to OFRecord

- Pretrained model

- Training and verification

- Training

- Varification

- Benchmark

InsightFace is an open-source 2D&3D deep face analysis toolbox, mainly based on MXNet.

In InsightFace, it supports:

• Datasets typically used for face recognition, such as CASIA-Webface、MS1M、VGG2(Provided with the form of a binary file which could run in MXNet, here is more details about the datasets and how to download.

• Backbones of ResNet, MobilefaceNet, InceptionResNet_v2, and other deep-learning networks to apply in facial recognition.

• Implementation of different loss functions, including SphereFace Loss、Softmax Loss、SphereFace Loss, etc.

Based upon the currently existing work of Insightface, OneFlow ported basic models from it, and now OneFlow supports:

• Training datasets of MS1M、Glint360k, and validation datasets of Lfw、Cfp_fp and Agedb_30, scripts for training and validating.

• Backbones of ResNet100 and MobileFaceNet to recognize faces.

• Loss function, e.g. Softmax Loss and Margin Softmax Loss（including Arcface、Cosface and Combined Loss）.

• Model parallelism and Partial FC optimization.

• Model transformation via MXNet.

To be coming further:

• Additional datasets transformation.

• Plentiful backbones.

• Full-scale loss functions implementation.

• Incremental tutorial on the distributed configuration.

This project is open for every developer to PR, new implementation and animated discussion will be most welcome.

First of all, before execution, please make sure that:

1. Install OneFlow

2. Prepare training and validation datasets in form of OFRecord.

According to steps in Install OneFlow install the newest release master whl packages.

python3 -m pip install --find-links https://release.oneflow.info oneflow_cu102 --user

According to Load and Prepare OFRecord Datasets, datasets should be converted into the form of OFREcord, to test InsightFace.

It has provided a set of datasets related to face recognition tasks, which have been pre-processed via face alignment or other processions already in InsightFace. The corresponding datasets could be downloaded from here and should be converted into OFRecord, which performs better in OneFlow. Considering the cumbersome steps, it is suggested to download converted OFrecord datasets:

MS1M-ArcFace(face_emore)

MS1MV3

It illustrates how to convert downloaded datasets into OFRecords, and take MS1M-ArcFace as an example in the following.

The structure of the downloaded MS1M-ArcFace is shown as follown：

faces_emore/

​    train.idx

​    train.rec

​    property

​    lfw.bin

​    cfp_fp.bin

​    agedb_30.bin

The first three files are MXNet recordio format files of MS1M training dataset, the last three .bin files are different validation datasets.

Only need to execute 2.1 or 2.2 2.1 Use Python scripts directly

Run

python tools/mx_recordio_2_ofrecord_shuffled_npart.py  --data_dir datasets/faces_emore --output_filepath faces_emore/ofrecord/train --part_num 16

And you will get the number of part_num parts of OFRecord, it's 16 parts in this example, it showed like this

tree ofrecord/test/
ofrecord/test/
|-- _SUCCESS
|-- part-00000
|-- part-00001
|-- part-00002
|-- part-00003
|-- part-00004
|-- part-00005
|-- part-00006
|-- part-00007
|-- part-00008
|-- part-00009
|-- part-00010
|-- part-00011
|-- part-00012
|-- part-00013
|-- part-00014
`-- part-00015

0 directories, 17 files

2.2 Use Python scripts + Spark Shuffle + Spark partition

Run

python tools/dataset_convert/mx_recordio_2_ofrecord.py --data_dir datasets/faces_emore --output_filepath faces_emore/ofrecord/train

And you will get one part of OFRecord(part-0) with all data in this way. Then you should use Spark to shuffle and partition.

1. Get jar package available You can download Spark-oneflow-connector-assembly-0.1.0.jar via Github or OSS

2. Run in Spark Assign that you have already installed and configured Spark. Run

//Start Spark 
./Spark-2.4.3-bin-hadoop2.7/bin/Spark-shell --jars ~/Spark-oneflow-connector-assembly-0.1.0.jar --driver-memory=64G --conf Spark.local.dir=/tmp/
// shuffle and partition in 16 parts
import org.oneflow.Spark.functions._
Spark.read.chunk("data_path").shuffle().repartition(16).write.chunk("new_data_path")
sc.formatFilenameAsOneflowStyle("new_data_path")

Hence you will get 16 parts of OFRecords, it shown like this

tree ofrecord/test/
ofrecord/test/
|-- _SUCCESS
|-- part-00000
|-- part-00001
|-- part-00002
|-- part-00003
|-- part-00004
|-- part-00005
|-- part-00006
|-- part-00007
|-- part-00008
|-- part-00009
|-- part-00010
|-- part-00011
|-- part-00012
|-- part-00013
|-- part-00014
`-- part-00015

0 directories, 17 files

The accuracy comparison of OneFlow and MXNet pretrained models on the verification set of the 1:1 verification accuracy on insightface recognition test (IFRT) are as follows:

The download link of the OneFlow pretrain model:of_005_model.tar.gz

We also provide the MXNet model which converted from OneFlow:of_to_mxnet_model_005.tar.gz

pip install oneflow-onnx==0.3.4
./convert.sh

To reduce the usage cost of user, OneFlow draws close the scripts to Torch style, you can directly modify parameters via configs/*.py

./run.sh

Moreover, OneFlow offers a validation script to do verification separately, val.py, which facilitates you to check the precision of the pre-training model saved.

./val.sh

• Data Parallelism

• Model Parallelism

• Partial FC

r denotes the sampling rate of negative class centers.

More test details could refer to OneFlow DLPerf.