CTC

A pure pytorch implementation of CTC (Connectionist Temoral Classification) objective. Equivalent to its Theano counterpart dandelion.objective.CTC_Logscale() The cost() function of CTC_Log class returns the average NLL over a batch samples given query sequences and score matrices.

class CTC_Log(eps=1E-12, inf=1E12, align='pre')

• align: {'pre' or 'post'}, indicating how samples in a batch are aligned.

.cost(queryseq, scorematrix, queryseq_mask=None, scorematrix_mask=None, blank_symbol=None)

Compute CTC cost, using only the forward pass

• queryseq: (B, L) LongTensor
• scorematrix: (B, T, C+1) FloatTensor
• queryseq_mask: (B, L) ByteTensor
• scorematrix_mask: (B, T) ByteTensor
• blank_symbol: scalar, = C by default Integer
• return: negative log likelihood averaged over a batch

  ---