Coding Period: Week 6

1 minute read


After presenting the simulation videos, we decided to analyze the failure cases. I will also combine the best performing models in single video for presentation and comparison. Furthermore, some PR and issues are created. Most importantly, I worked on using TensorRT to optimize the tensorflow models. Using TensorRT (here TF-TRT) is not typical and therefore I have shared many resources in the blog to understand it.


  • Record F1 car crashing videos and their stats
  • Record stats for Dynamic range quantization model on Montreal circuit
  • Combine the best performing simulation in one video with explainations
  • Updata PR #67 with results table, inference script and weight links
  • Prepare TensorRT scripts for optimization (float32, float16 and int8 modes)
  • Evaluate TensorRT in offline mode

Related to use BehaviorMetrics repository:

Updated PR:

Repository for new script:

The execution

Video demonstration

I used my personal computer with a NVIDIA GeForce GTX 1050/PCIe/SSE2 GPU with Intel® Core™ i7-7700HQ CPU @ 2.80GHz × 8 CPU, 8 GB RAM and batch size of 1 for simulation.

The crash videos are avaiable on google drive -

Final video

Optimization with TensorRT

I used the server with a 8 GB NVIDIA GeForce GTX 1080/PCIe/SSE2 GPU and batch size of 1 (for inference).

There are some useful resources, I have listed below:

The scripts are available here -

Result table

I am using tensorflow graphs (low-level API) for inference here.

Method Model size (MB) MSE Inference time (s)
Baseline 196 0.041032556329194385 0.0012623071670532227
Precision fp32 260 0.04103255125749467 0.0013057808876037597
Precision fp16 260 0.04103255125749467 0.0021804444789886475
Precision int8 260 0.04103255125749467 0.0011799652576446533


  • Conversion with int8 precision has best inference time.
  • There is no loss in performance (MSE is same) of the optimized models.
  • The model size has increased for optimized models.
  • As we know there is no guarantee to receive better performing models with TensorRT, we can expect the results we have with PilotNet.