README.md

metadata

license: apache-2.0

Tensorflow

• Version: 2.7.0
• TensorRT version: 7.2.2.1
• Docker image: nvcr.io/nvidia/tensorflow:20.12-tf2-py3
• GPU: NVIDIA GeForce 3090
• CUDA: 11.6
• Driver version: 510.54

Optimization	Model size (MB)	MSE	Inference time (s/frame)	Filename
Baseline	6.0925140380859375	0.010881431312199034	0.0016004319190979005	28_04_pilotnet_model.tflite
Dynamic Range Quantization	1.5389328002929688	0.041451400199878044	0.0008851253986358643	28_04_pilotnet_dynamic_quant.tflite
Integer Quantization	1.5389328002929688	0.01102226436099348	0.0008868560791015625	28_04_pilotnet_int_quant.tflite
Integer (float fallback) Quantization	1.5389175415039062	0.011030688516599923	0.0008031470775604248	28_04_pilotnet_intflt_quant.tflite
Float16 Quantization	-	-	-	28_04_pilotnet_float16_quant.tflite
Quantization Aware Training	1.5446319580078125	0.0115418379596583	0.0008456888198852539	28_04_pilotnet_quant_aware.tflite
(random sparse) Weight pruning	6.0925140380859375	0.011697137610230973	0.0016570956707000733	28_04_pilotnet_pruned.tflite
(random sparse) Weight pruning Quantization	1.536590576171875	0.011635421636510991	0.0012711701393127441	28_04_pilotnet_pruned_quan.tflite
Cluster preserving Quantization Aware	1.5446319580078125	0.010546523951115492	0.0008221814632415771	28_04_pilotnet_cqat_model.tflite
Pruning preserving Quantization Aware	1.5446319580078125	0.010758002372154884	0.0008252830505371093	28_04_pilotnet_pqat_model.tflite
Sparsity and cluster preserving quantization aware training (PCQAT)	1.5446319580078125	0.008262857163545972	0.0008286898136138916	28_04_pilotnet_pcqat_model.tflite

TensorRT-Tensorflow

Optimization	Model size (MB)	MSE	Inference time (s/frame)	Folder
Float32 Quantization	0.00390625	0.010798301750717706	0.00038761067390441896	24_04_pilotnet_tftrt_fp32
Float16 Quantization	0.00390625	0.010798278900279191	0.00042218327522277834	24_04_pilotnet_tftrt_fp16
Int8 Quantization	0.00390625	0.04791482252948612	0.0003384373188018799	14_06_pilotnet_tftrt_fp16