A unified framework of perturbation and gradient-based attribution methods for Deep Neural Networks interpretability. DeepExplain also includes support for Shapley Values sampling. (ICLR 2018)
I am trying to use DeepExplain on DeepSpeech and I ran into the "None values not supported" error even though I am using input and target tensors that are part of a standard LSTM network. Is there any common pitfall I might have fallen into?
To give you a bit more context, DeepSpeech performs Speech to Text transcription and its architecture is split in 3 main parts:
A feature extraction part (MFCC) that computes features used as input for the Neural Network
A LSTM NN that outputs a character per time frame
A language models that transforms the previous output in a well formed text
In my case I am working on the second part only, trying to compute attributions with target_tensor = NN output (softmax) and input_tensor = NN input (MFCC features). I am pretty sure that my target and input tensors are connected in the backpropagation and that I am not making the error mentioned in the "NLP / Embedding lookups" section of the readme.
Is there a known reason for this error to happen in this case? Could it be caused by the network architecture (LSTM) or the fact that the network has two input tensors?
Hello,
I am trying to use DeepExplain on DeepSpeech and I ran into the "None values not supported" error even though I am using input and target tensors that are part of a standard LSTM network. Is there any common pitfall I might have fallen into?
To give you a bit more context, DeepSpeech performs Speech to Text transcription and its architecture is split in 3 main parts:
Some details are provided here: https://deepspeech.readthedocs.io/en/v0.6.1/DeepSpeech.html
In my case I am working on the second part only, trying to compute attributions with target_tensor = NN output (softmax) and input_tensor = NN input (MFCC features). I am pretty sure that my target and input tensors are connected in the backpropagation and that I am not making the error mentioned in the "NLP / Embedding lookups" section of the readme.
Is there a known reason for this error to happen in this case? Could it be caused by the network architecture (LSTM) or the fact that the network has two input tensors?
Thank you and best, Romain