iNNvestigate 2.0: LRP-Network-Canonization Module

[leanderweber]

Special Model Types to Consider

• LSTM
• ResNet
• DenseNet
• ...

Mixed Ideas

• Replace Pooling Layer with Conv2D
• ...

[miriamelia]

I get NotImplementedError:

LRP is only tested for convolutional neural networks. Check triggered by layers: [<keras.layers.core.tf_op_layer.TFOpLambda object at 0x2c36c0370>, <keras.layers.core.tf_op_layer.TFOpLambda object at 0x2c3879eb0>, <keras.layers.core.tf_op_layer.TFOpLambda object at 0x2c388da30>, <keras.layers.core.tf_op_layer.TFOpLambda object at 0x2c36c0310>, <keras.layers.wrappers.Bidirectional object at 0x2c38a69d0>, <keras.layers.wrappers.Bidirectional object at 0x2c38ce4f0>]

even though I have version 2.0.1:

conda list
innvestigate 2.0.1 pypi_0 pypi

... so lrp is not implemented for lstms?

[adrhill]

Yes, iNNvestigate doesn't support LSTM layers.

[miriamelia]

okay :)
found this implementation, maybe one day it could be implemented in your framework
https://github.com/ArrasL/LRP_for_LSTM