OptimizedPilotNet

These models contain a PilotNet baseline model for end-to-end control of an autonomous ego vehicle in CARLA and its optimized versions for both PyTorch and Tensorflow.

• GPU: NVIDIA GeForce 3090
• CUDA: 11.6
• Driver version: 510.54
• Input shape (200,66,3)

Tensorflow

• Tensorflow version: 2.7.0 / 2.11.0
• TensorRT version: 7.2.2.1
• Docker image: nvcr.io/nvidia/tensorflow:20.12-tf2-py3
• nvidia-tensorrt: 7.2.2.1

Optimization	Model size (MB)	MSE	Inference time (s/frame)	Filename
Original	19	0.018	0.022	pilotnet.h5
Baseline	6.0925140380859375	0.010881431312199034	0.0016004319190979005	pilotnet_model.tflite
Dynamic Range Quantization	1.5377578735351562	0.010803998294344926	0.0008851253986358643	pilotnet_dynamic_quant.tflite
Integer Quantization	1.5389328002929688	0.01102226436099348	0.0008868560791015625	pilotnet_int_quant.tflite
Integer (float fallback) Quantization	1.5389175415039062	0.0008868560791015625	0.0008031470775604248	pilotnet_intflt_quant.tflite
Float16 Quantization	3.0508956909179688	0.010804510797606127	0.0013616561889648437	pilotnet_float16_quant.tflite
Quantization Aware Training	1.5446319580078125	0.0115418379596583	0.0008456888198852539	pilotnet_quant_aware.tflite
(random sparse) Weight pruning	6.0925140380859375	0.011697137610230973	0.0016570956707000733	pilotnet_pruned.tflite
(random sparse) Weight pruning Quantization	1.536590576171875	0.011635421636510991	0.0012711701393127441	pilotnet_pruned_quan.tflite
Cluster preserving Quantization Aware	1.5446319580078125	0.010546523951115492	0.0008221814632415771	pilotnet_cqat_model.tflite
Pruning preserving Quantization Aware	1.5446319580078125	0.010758002372154884	0.0008252830505371093	pilotnet_pqat_model.tflite
Sparsity and cluster preserving quantization aware training (PCQAT)	1.5446319580078125	0.008262857163545972	0.0008286898136138916	pilotnet_pcqat_model.tflite

TensorRT-Tensorflow:

To do inference:

  pip install nvidia-tensorrt===7.2.2.1
  python3 -c "import tensorrt; print(tensorrt.__version__); assert tensorrt.Builder(tensorrt.Logger())"
  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/python3.8/site-packages/tensorrt
  python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

Optimization	Model size (MB)	MSE	Inference time (s/frame)	Folder
Float32 Quantization	0.00390625	0.010798301750717706	0.00038761067390441896	pilotnet_tftrt_fp32
Float16 Quantization	0.00390625	0.010798278900279191	0.00042218327522277834	pilotnet_tftrt_fp16
Int8 Quantization	0.00390625	0.04791482252948612	0.0003384373188018799	pilotnet_tftrt_int8

---

PyTorch

• PyTorch version: 1.13.1+cu116
• TensorRT version: 8.5.5
• Docker image: nvcr.io/nvidia/pytorch:22.12-py3
• torch-tensorrt: 1.3.0

Optimization	Model size (MB)	MSE	Inference time (s/frame)	Filename
Original	6.1217	0.03524	-	pilotnet_model.pth
Dynamic Range Quantization	1.9493608474731445	0.012065857842182075	0.001480283498764038	24_05_dynamic_quan.pth
Static Quantization	1.6071176528930664	0.012072610909984047	0.0007314345836639404	24_05_static_quan.pth
Quantization Aware Training	1.6069536209106445	0.01109830549109022	0.0011710402965545653	24_05_quan_aware .pth
Local Prune	6.122584342956543	0.010850968803449539	0.0014387350082397461	24_05_local_prune.pth
Global Prune	6.122775077819824	0.010964057565769462	0.0014179635047912597	24_05_global_prune.pth
Prune + Quantization	1.6067094802856445	0.010949893930274941	0.0011728739738464356	24_05_prune_quan.pth

TensorRT-PyTorch:

To do inference:

  pip install torch-tensorrt==1.3.0

Optimization	Model size (MB)	MSE	Inference time (s/frame)	Filename
Float32 Quantization	6.121363639831543	0.009570527376262128	0.0002284455299377441	trt_mod_float32.jit.pt
Float16 Quantization	6.121363639831543	0.009571507916721152	0.000250823974609375	trt_mod_float16.jit.pt
Int8 Quantization	6.181861877441406	0.00969304293365875	0.0002463934421539307	trt_mod_int8.jit.pt