Abstract
Research and implement a pipeline for predicting the damages caused to a car after an accident, using sensors from the car, combines sequences based modelling (LSTM) as well as CNN on time series.
Challenges
- Data collection: there was a problem with the labeling of the data so I spent one week creating a data parser to correctly label the data.
- Data preprocessing: most of the data was stored in thousands of csv files.
- Time series modelling: I had to read and implement a few research papers (which I haven’t done in the past).
- Efficiency: one of the goals was to explore the use of different deep learning architectures that are usually not used for time series but could significantly shorten training time of the model.
- Evaluation metric: since it was a regression problem, I had to come up with my own evaluation metrics, in addition to the MSE of the models.
Achievements (according to KPIs)
- Train different models on an google cloud machine
- Rank every model according to the evaluation metric I built.
- I was able to build the entire project using pytorch classes, the preprocessing, modelling, training and evaluation.
