Problem task: use autoencoders on timeseries data from Green Mobility fleet to see if they fit in a latent space which could help in the binary classification problem: "good road/bad road"
Z-acceleration time sequences are represented as pandas series.
We do some preprocessing: drops ones containing not enough data points, as well as creating pytorch PackedSequences out of them all. PackedSequence is a timeseries which is padded to a certain size for reconstruction purposes but lets torch.RNN layers know not to consider the padding as something to compute on.
We then load this data into an enhanced Dataloader object
The encoder itself is written in RNN_AE,a wrapper class for an encoder and decoder subclass
newAE_conv.ipynb is where you will find most reults: training and validation performance
results of the visualization of the latent space are in the visualizations.ipynb file.
In general, when looking through old commit or old models we made, we tried to keep the conventions
*_enc (RNN_enc, enc, conv1d_enc, etc)
#trains RRN layer using using packed sequences
#returns unpacked
# forward pass unpacks sequences qhen training RNN
*_dec
#calls get_packed method we defined
#returns packed
Initializable like regular torch model (inherits from nn.Module()).
class: AE_conv, uses
couckou
ok :)