search

nicholas-leonard / dp

A deep learning library for streamlining research and development using the Torch7 distribution.
Other
343 stars 139 forks source link

Confusion feedback _add output is table not tensor #185

Closed cwellsarnold closed 8 years ago

cwellsarnold commented 8 years ago

Hi all,

I have a bidirectional LSTM network that I am using for sequence tagging. I'm trying to use the Confusion feedback object. However, in the _add function, the code expects output to be a tensor, but in my case it is a table. This then fails at the output:view() and output:size(1) calls on line 60. When I run my code, the table is equal to my context_size. So, if I have a batch size of 512, six classes, and a context_size of 5, I get a table of five 512x6 tensors.

Perhaps I am configuring my network wrong? Or am I not understanding context_size and rho? I think they are the same thing. Or should context_size=1 for LSTMs and rho = the context window? Even when I try this, the output object in Confusion is still a table of size 1.

My code uses a logsoftmax output, like the NN tutorial (https://github.com/nicholas-leonard/dp/blob/4616f3a2eca5be0cf1500afe5a0a156e277a9c53/doc/neuralnetworktutorial.md), so perhaps it is a different layer?

Thanks for any help!

cwellsarnold commented 8 years ago

Related follow-up question, when I try to access the confusion matrix in the epoch_callback function using report.optimizer.feedback.confusion.matrix, every element in the matrix is zero. report.optimizer.feedback.confusion.per_class.accuracy is also a vector of 0s. This happens in the NN tutorial as well. How should one correctly retrieve the confusion matrix for display?

Thanks again!

cwellsarnold commented 8 years ago

Figured it out. The first element in the output table for a recurrent model holds the predictions that correspond to batch:targets():forward('b').

I'm printing the confusion matrix from confusion.lua rather than my code, but it seems to work fine.

nicholas-leonard commented 8 years ago

:+1: