ML Distributed training

Demo: https://youtu.be/OOPVA-eqBTY

Introduction

This project leverage the power of multiple GPUs with the target is to reduce the training time of complex models by data parallelism method with 2 approaches:

1. Multi-worker Training using 2 PCs with GeForce RTX GPU as Workers via:
   • Local area network (LAN).
   • VPN tunnel using OpenVPN (not included in the demo).
2. Parameter Server Training using 5 machines in LAN:
   • 2 Laptops as Parameter Server connected via 5GHz Wi-Fi.
   • 2 PCs with GeForce RTX GPU as Workers.
   • 1 PC just with CPU as a Coordinator.

Dataset

We used our self-built 30VNFoods dataset which includes collected and labeled images of 30 famous Vietnamese dishes. This dataset is divided into:

• 17,581 images for training.
• 2,515 images for validation.
• 5,040 images for testing.

In addition, we also used a small TensorFlow flowers dataset with about 3700 images of flowers, which includes 5 folders corresponding to 5 types of flowers (daisy, dandelion, roses, sunflowers, tulips).

Setup

Image size	(224, 224)
Batch size/worker	32
Optimizer	Adam
Learning rate	0.001

The iperf3 tool is used to measure the bandwidth of machines in network.

1. Multi-worker Training

2. Parameter Server Training

Result

Training method	Dataset	Connection	Avg. s/epoch
Single-worker	flowers	LAN	14
Multi-worker	flowers	LAN	18
Multi-worker	flowers	VPN Tunnel	635
Multi-worker	30VNFoods	LAN	184
Parameter Server	30VNFoods	LAN	115

⇒ For more information, see Report.pdf.

References

• Distributed training with Keras
• A friendly introduction to distributed training (ML Tech Talks)
• Distributed TensorFlow training (Google I/O '18)
• Inside TensorFlow: Parameter server training
• Performance issue for Distributed TF
• When is TensorFlow's ParameterServerStrategy preferable to its MultiWorkerMirroredStrategy?