Train a PyTorch LSTM Model on Existing Dataset #4
Description
We need to process data about the pilot and their relevant joints in order to predict their ongoing trajectory. This study shows very promising results in predicting the accuracy of human gait using a LSTM Model https://www.mdpi.com/2076-0825/11/3/73.
Our very first goal in this process is to train a model predicting human gait, only including walking data. We won't have access to sensors initially, so we may have to simulate data, or find existing datasets online.
Success Criteria:
- Most important thing is the accuracy, we would ike to aim for 90% or higher accuracy in trajectory prediction
- Real-time performance, there are many ways to improve performance if the initial model is not performant enough but it is crucial. We likely will need sub 10-20ms inference time
Have a python script that can be used like this:
train.py --data /../ --params: {TBD ie input_size, hidden_size} which results in something like a model.pt file that can be used and tested.
We will most likely have IMU sensors placed on hip, knee, and ankle joint and have access to whatever they can give us. The study above uses joint angle for each joint - we should experiment with different combinations of available data and validate results.
It's pretty important to train a pytorch (.pt) model, as it should interface easily with the C++ code the embedded team is. writing