Timing the ONNX model, retrain CMS-GNNLSH-TF

[jpata]

Overview

• added a simple timing script based on ONNX
• added some timing benchmarks based on the current CLIC v1.6 model
• include the new cms_pf_multi_particle_gun dataset
• include new smaller versions of the CLIC hit-based datasets: clic_edm_ttbar_hits_pf10k, clic_edm_qq_hits_pf10k

Timing tests

https://github.com/jpata/particleflow/blob/timing/timing.md
Tested on RTX2070S 8192MiB.

batch_size=20 bin_size=256 num_features=17 use_gpu=True num_threads=1
Nelem=256 mean_time=1.96 ms stddev_time=0.52 ms mem_used=782 MB
Nelem=512 mean_time=3.21 ms stddev_time=0.15 ms mem_used=916 MB
Nelem=2560 mean_time=21.48 ms stddev_time=0.14 ms mem_used=1721 MB
Nelem=5120 mean_time=50.09 ms stddev_time=0.19 ms mem_used=2795 MB
Nelem=10240 mean_time=106.15 ms stddev_time=0.63 ms mem_used=4943 MB

[jpata]

Training the CMS model, reverting the pt and energy regression to Huber loss (red curve) from MeanSquaredLogError (blue curve), jet and MET performance are much, much, much better.

Jet and MET response on a small sample of cms_pf_qcd_high_pt: