Edit distance between two sequences reference (ref) and hypothesis (hyp). Returns edit distance, number of insertions, deletions and substitu

Based off Awni's runAll.py

Loops through node list, returns on first encounter with free gpu. Returns node name and device id if free, -1,-1 otherwise.

Gradient check for model. model should at least implement two methods: costAndGradVec(param, *args) - takes a vectorized parameter a

Gradient check for model. model should at least implement two methods: costAndGradVec(param, *args) - takes a vectorized parameter a

Initialize parameters using 6/sqrt(fanin+fanout)

Initialize parameters using 6/sqrt(fanin+fanout)

Initialize parameters using 6/sqrt(fanin+fanout)

Read in data

Runs stochastic gradient descent with model as objective. Expects data in n x m matrix where n is feature dimension and m is number

Runs stochastic gradient descent with model as objective. Uses single utterances instead of minibatches

Sets view of gpu memory to be correct size for utterance.

Sets view of gpu memory to be correct size for utterance.

Sets view of gpu memory to be correct size for utterance.

Reads swbd transcripts and builds swbd k to list of integer labels mapping.

Writes header for each utterance in Kaldi Style, assumes data written in float32 and C order.

Takes utterance to alignment mapping and splits it up into alignment files according to file structure of training set.

filedir_feat: directory for feature and key files filedir_ali: directory for alignment files. Assumed same as filedir if not given

temporalLayer indicates which layer is recurrent. <= 0 indicates no recurrernce

Compute bigrams with smoothing. Save in bigrams.bin.

Wrapper function used for SFO optimizer.

Vectorized version of costAndGrad

Vectorized version of CTC cost data is a single utterance. Each column is a time index [0...T]