LaNet
LaNet - Unidirectional LSTM architecture for lane identification using Accelerometer data
Reference Codebase for LaNet proposal submitted to ACM BuildSys 2019.
Data for model training available at : https://drive.google.com/file/d/1keAI9dy1iqarq84IJfWamleNDFrEMXkd/view?usp=sharing
Highlights
- Codes for LaNet LSTM arch along with WeightedLSTMCrossEntropyLoss are included.
- Supports Tensorboard visualisations to visualize incremental accuracy observed in each LSTM Cell
- Provides scope for reconfiguration of model parameters including sub-drive length and sub-segment lengths.
Default model configurations
- Sample Length "l" = 800000
- Sub-drive stride "s" = 50000
- Sub-segment length"d" = 50000
- Sub-segment stride "m" = int(50000/2)
- LSTM Hidden dimension "H" = 300
- Learning rate "lr" = 0.005
- Num epochs = 4
- Number of LSTM Layers = 2
- Number of lanes/num_classes = 2
- Accelerometer data "sampling_rate" = 2000
- Batch Size = 512