o Gi@sdZddlmZddlZddlmZddlmZddlm Z ddl m Z dd l m Z  d!dejdejde dededededede fddZ d"de dede fddZde deeefdd ZdS)#z8 This file contains various functions for CTC decoding. )ListN)intersect_dense_pruned) DenseFsaVec)Fsa shortest_path) RaggedTensor'log_prob log_prob_lendecoding_graph search_beam output_beammin_active_statesmax_active_statessubsampling_factorreturnc Cs|jdks J|j|jdksJ|j|d|dks&J|j|jf|d}tt|t||g tj } t || |dd} t || ||||d} | S)a4Get the decoding lattice from a decoding graph and log_softmax output. Args: log_prob: Output from a log_softmax layer of shape ``(N, T, C)``. log_prob_len: A tensor of shape ``(N,)`` containing number of valid frames from ``log_prob`` before padding. decoding_graph: An Fsa, the decoding graph. It can be either an ``HLG`` or an ``H``. You can use :func:`ctc_topo` to build an ``H``. search_beam: Decoding beam, e.g. 20. Smaller is faster, larger is more exact (less pruning). This is the default value; it may be modified by `min_active_states` and `max_active_states`. output_beam: Beam to prune output, similar to lattice-beam in Kaldi. Relative to best path of output. min_active_states: Minimum number of FSA states that are allowed to be active on any given frame for any given intersection/composition task. This is advisory, in that it will try not to have fewer than this number active. Set it to zero if there is no constraint. max_active_states: Maximum number of FSA states that are allowed to be active on any given frame for any given intersection/composition task. This is advisory, in that it will try not to exceed that but may not always succeed. You can use a very large number if no constraint is needed. subsampling_factor: The subsampling factor of the model. Returns: An FsaVec containing the decoding result. It has axes [utt][state][arc]. rr)allow_truncate)rrrr) ndimshapesizetorchstackarangezeroscputtoint32rr) rrrrrrrr batch_sizesupervision_segment dense_fsa_veclatticer(=/home/ubuntu/.local/lib/python3.10/site-packages/k2/decode.py get_lattices<*   r*Tr'use_double_scorescCst||d}|S)a;Get the best path from a lattice. Args: lattice: The decoding lattice returned by :func:`get_lattice`. use_double_scores: True to use double precision floating point in the computation. False to use single precision. Return: An FsaVec containing linear paths. )r+r)r'r+ best_pathr(r(r)one_best_decodingbs r- best_pathscCst|jtr&|jd}|j|j}|d}|d}t||j}n|jd}t||j}|d}|j dks@J| S)aExtract aux_labels from the best-path FSAs and remove 0s and -1s. Args: best_paths: An Fsa with best_paths.arcs.num_axes() == 3, i.e. containing multiple FSAs, which is expected to be the result of `shortest_path` (otherwise the returned values won't be meaningful). TODO: Also return timestamps of each label. Returns: Returns a list of lists of int, containing the label sequences we decoded. rr) isinstance aux_labelsr remove_values_leqarcsrcompose remove_axisvaluesnum_axestolist)r.r1 aux_shaper(r(r)get_aux_labelsus      r:)r r r r r)T)__doc__typingrrautogradrr&rfsarfsa_algorraggedr Tensorfloatintr*boolr-r:r(r(r(r)sR       S