CERN Summer Student Project
by Alex Burlacu
[Full topic of the project here]
CERN Summer Student Project
by Alex Burlacu
Benchmarking TMVA package against TensorFlow on event-by-event inference performance on multi-layered perceptrons for HEP
CERN Summer Student Project
by Alex Burlacu
TL;DR: TMVA DeepNets vs Tensorflow DeepNets
What is TMVA?
What are the use cases?
-
Provides Machine Learning capabilities to ROOT users
-
... also Feature Selection and data manipulation
-
Is part of the ROOT package
-
Low-latency inference
-
(As) Real-time (as possible) inference
-
For high energy physics
Why this project matters?
Define TMVA's Deep Learning Roadmap
Benchmark Setting
- Intel i7-7820X CPU @ 3.60GHz (8 PCores/16 LCores)
- 32 GB DDR4 RAM
- Non-optimized TF v1.9rc1 and Keras
- MKL Optimized TF v1.8 and Keras
- TF v1.8 C++ API
- TMVA from ROOT 6.14
Inference Overhead
Inference Overhead
TMVA is 1-2 orders of magnitude faster on small batches and nets
Dependency of inference time on model size (TMVA / TF C++ inf. time)
Dependency of inference time on model size (TMVA / TF C++ inf. time)
TMVA MLPs work best with 1-4 layers and narrow layers
Dependency of inference time on batch size
2 Layers, 128 Neurons/Layer
Dependency of inference time on batch size
3 Layers, 256 Neurons/Layer
For Batch Size > 32 TMVA isn't the best choice
Thank you!
Questions?
Benchmarking TMVA package against TensorFlow on event-by-event inference performance on multi-layered perceptrons for HEP
By Alexandru Burlacu
