Coding Period: Week 1
Preliminaries
During the community bonding period, I prepared the tools for use in coding period. The tools now support training, evaluation and comparison of models in various ways. I further made changes to my PR’s according to the discussion with the mentors. Next, I will start the work on optimizing the models. I will start with models trained in TensorFlow library TF model link. This week focus is on the architectures - PilotNet and DeepestLSTMTinyPilotNet. After setting a baseline on my personal computer, I will start experiments to optimize the models with TensorRT [3]. I also received access to Nvidia 1080 GPU machine, which would be used later for experiments.
Objectives
- Update previous PR’s according to requested changes.
- Test access to GPU server.
- PilotNet baseline - Calculate loss (regression), inference time (script and BehaviorNet)
- DeepestLSTMTinyPilotNet baseline - Calculate loss (regression), inference time (script and BehaviorNet)
- PilotNet optimization experiment with TensorRT
- DeepestLSTMTinyPilotNet optimization experiment with TensorRT
Issues and Pull requests.
- While working on baseline I came across issue Missing argument: learning_rate #55
The execution
I follow the data split defined in paper Memory based neural networks for end-to-end autonomous driving. The baseline was calculate using dataset based on circuits - Simple circuit, Montmeló and Montreal. I came across an error regarding cuda Could not load dynamic library 'libcudart.so.11.0'; dlerror: libcudart.so.11.0. It took me quite some time to solve the issue and install new cuda package, which was too time consuming. I need to reinstall the complete GPU (cuda) environment from scratch and found suitable instructions in the blog and some independent articles for Ubuntu 18.04 and Cuda 11.4.
A preliminary table to record baseline performances on only simple circuit dataset:
| Model | Loss | MSE | Mean Absolute Error | Inference time |
|---|---|---|---|---|
| PilotNet | 0.014 | 0.014 | 0.077 | 0.0735 sec |
Due to the major blockage with cuda environment, the other unfinished goals will be completed in next week. My code for evaluating the performance of PilotNet can be found in a separate branch baseline-exp.
References
[1] https://github.com/JdeRobot/BehaviorMetrics
[2] https://github.com/JdeRobot/DeepLearningStudio
[3] https://developer.nvidia.com/tensorrt