An efficient, pure Python and Numpy implementation of an LSTM Network.
This is a pure numpy and python implementation of an LSTM network. The example here is for time-series prediction.
Required dependiencies are:
- Numpy
- Pandas (only if importing DataFrames)
- Matplotlib (for visualisation)
The execution file is not commented as of yet, however the LSTM class object file has comments to understand what's happening. This is loosely based on a Gist by Karpathy.
The LSTM cell includes "Peep-hole" connections.
Since the implementation does not use batch-training, the network's convergence is not optimal. This can be seen in the forward projections of a time-series, as the projections have some deviations. However, the hope is that it clearly shows a pure implementation of an LSTM cell and a network to gain a deeper understanding.
Uses Adagrad for training the network.