InsightFace in OneFlow

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.

Implementation in OneFlow

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.

Preparations

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

Install OneFlow
Prepare training and validation datasets in form of OFRecord.

Install OneFlow

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

Data preparations

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, training parts and validation parts.

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

1. Download datasets

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.

2. Transformation from MS1M recordio to OFRecord

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_shuffled_npart.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.

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

3. Transformation from validation datasets to OFRecord

Run

python bin_2_ofrecord.py --data_dir=datasets/faces_emore --output_filepath=faces_emore/ofrecord/lfw/ --dataset_name="lfw"

python bin_2_ofrecord.py --data_dir=faces_emore --output_filepath=faces_emore/ofrecord/cfp_fp/ --dataset_name="cfp_fp"

python bin_2_ofrecord.py --data_dir=datasets/faces_emore --output_filepath=faces_emore/ofrecord/agedb_30/ --dataset_name="agedb_30"

Pretrained model

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:

Framework	African	Caucasian	Indian	Asian	All
OneFlow	90.4076	94.583	93.702	68.754	89.684
MXNet	90.45	94.60	93.96	63.91	88.23

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

Training and verification

Training

To reduce the usage cost of user, OneFlow draws close the scripts to MXNet style, you can directly modify parameters via sample_config.py. Meanwhile, it could Validate while training when adding --do_validataion_while_train=True.

Just change the parameters in the sample_config.py straightforward. Modify and copy config.py

cp sample_config.py config.py

vim config.py # edit dataset path etc.

run

python insightface_train.py --dataset emore --network r100 --loss arcface

In this way, you will do training and validation with the backbone of ResNet100 by face_emore dataset.

To achieve ambitions for a larger quantity of data, run

python insightface_train.py --dataset glint360k_8GPU --network r100_glint360k --loss cosface

In this way, you will do training and validation with the backbone of ResNet100 by glint360k dataset.

Varification

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

run

python insightface_val.py \
--device_num_per_node=1 \
--network="r100" \
--model_load_dir=path/to/model_load_dir

Benchmark

Training Speed Benchmark

Face_emore Dataset & FP32

Backbone	GPU	model_parallel	partial_fc	BatchSize / it	Throughput img / sec
R100	8 * Tesla V100-SXM2-16GB	False	False	64	1836.8
R100	8 * Tesla V100-SXM2-16GB	True	False	64	1854.15
R100	8 * Tesla V100-SXM2-16GB	True	True	64	1872.81
R100	8 * Tesla V100-SXM2-16GB	False	False	96(Max)	1931.76
R100	8 * Tesla V100-SXM2-16GB	True	False	115(Max)	1921.87
R100	8 * Tesla V100-SXM2-16GB	True	True	120(Max)	1962.76
Y1	8 * Tesla V100-SXM2-16GB	False	False	256	14298.02
Y1	8 * Tesla V100-SXM2-16GB	True	False	256	14049.75
Y1	8 * Tesla V100-SXM2-16GB	False	False	350(Max)	14756.03
Y1	8 * Tesla V100-SXM2-16GB	True	True	400(Max)	14436.38

Glint360k Dataset & FP32

Backbone	GPU	partial_fc sample_ratio	BatchSize / it	Throughput img / sec
R100	8 * Tesla V100-SXM2-16GB	0.1	64	1858.57
R100	8 * Tesla V100-SXM2-16GB	0.1	115	1933.88

Evaluation on Lfw, Cfp_fp, Agedb_30

Data Parallelism

Backbone	Dataset	Lfw	Cfp_fp	Agedb_30
R100	MS1M	99.717	98.643	98.150
MobileFaceNet	MS1M	99.5	92.657	95.6

Model Parallelism

Backbone	Dataset	Lfw	Cfp_fp	Agedb_30
R100	MS1M	99.733	98.329	98.033
MobileFaceNet	MS1M	99.483	93.457	95.7

Partial FC

Backbone	Dataset	Lfw	Cfp_fp	Agedb_30
R100	MS1M	99.817	98.443	98.217

Evaluation on IFRT

r denotes the sampling rate of negative class centers.

Backbone	Dataset	African	Caucasian	Indian	Asian	ALL
R100	Glint360k(r=0.1)	90.4076	94.583	93.702	68.754	89.684

Max num_classses

node_num	gpu_num_per_node	batch_size_per_device	fp16	Model Parallel	Partial FC	num_classes
1	1	64	True	True	True	2000000
1	8	64	True	True	True	13500000

More test details could refer to OneFlow DLPerf.

xrosliang / oneflow_face Goto Github PK

oneflow_face's Introduction