TLSTM original (tensorflow 1.2 & python 2.7): https://github.com/illidanlab/T-LSTM
LRP for LSTM original: https://github.com/ArrasL/LRP_for_LSTM
This code is tested with python==3.8, tensorflow-gpu==2.7, numpy==1.23, and cuda/cudnn 11.2/8.1
- Python 3.X
- Tensorflow 2.X
- CUDA 10.X or 11.X for gpu settings depends on your hardware
Since this code is only a cell of TLSTM, you need an RNN layer in order to use it.
# ex) multi-label classification
import tensorflow as tf
from tensorflow.keras import layers, Model
from model import ModelConstruct
from nplrp import LRPforTLSTM
# TLSTM
model = ModelConstruct(...)
# LRP for TLSTM
lrp = LRPforTLSTM(model=model)
lrp.set_input(...)
lrp.lrp(...)Note that this is a time-aware LSTM thus additional time input needs to be prepared and the time_input option needs to be set to True. It may work as a vanilla LSTM otherwise.
input shape: (batch_size, seq_len, dim)
time input shape: (batch_size, seq_len, 1)
input = tf.concat([time_input, input], axis=-1)
Refer to the code for further details.