theano.tensor.nnet.ctc – Connectionist Temporal Classification (CTC) loss¶
Usage of connectionist temporal classification (CTC) loss Op, requires that
the warp-ctc library is
available. In case the warp-ctc library is not in your compiler’s library path,
config.ctc.root configuration option must be appropriately set to the
directory containing the warp-ctc library files.
This interface is the prefered interface. It will be moved automatically to the GPU.
Unfortunately, Windows platforms are not yet supported by the underlying library.
ctc(activations, labels, input_lengths)¶
Compute CTC loss function.
Using the loss function requires that the Baidu’s warp-ctc library be installed. If the warp-ctc library is not on the compiler’s default library path, the configuration variable
config.ctc.rootmust be properly set.
- activations – Three-dimensional tensor, which has a shape of (t, m, p), where t is the time index, m is the minibatch index, and p is the index over the probabilities of each symbol in the alphabet. The memory layout is assumed to be in C-order, which consists in the slowest to the fastest changing dimension, from left to right. In this case, p is the fastest changing dimension.
- labels – A 2-D tensor of all the labels for the minibatch. In each row, there is a sequence of target labels. Negative values are assumed to be padding, and thus are ignored. Blank symbol is assumed to have index 0 in the alphabet.
- input_lengths – A 1-D tensor with the number of time steps for each sequence in the minibatch.
Cost of each example in the minibatch.
CTC loss function wrapper.
Using the wrapper requires that Baidu’s warp-ctc library is installed. If the warp-ctc library is not on your compiler’s default library path, you must set the configuration variable
Parameters: compute_grad – If set to True, enables the computation of gradients of the CTC loss function.