o Gi(@sddlmZddlmZddlZddlZddlZddlmZddlmZddlmZddlm Z dd l m Z dd lm Z Gd d d e ZGd dde ZdS))List)TupleN)Fsa) RaggedShape) RaggedTensor)Tensor) index_select)RnntDecodingConfigc@s,eZdZdeddfddZdefddZdS)RnntDecodingStreamfsareturnNcCs ||_t|j|_|j|_dS)a Create a new rnnt decoding stream. Every sequence(wave data) needs a decoding stream, this function is expected to be called when a new sequence comes. We support different decoding graphs for different streams. Args: graph: The decoding graph used in this stream. Returns: A rnnt decoding stream object, which will be combined into :class:`RnntDecodingStreams` to do decoding together with other sequences in parallel. N)r _k2create_rnnt_decoding_streamarcsstreamdevice)selfr rB/home/ubuntu/.local/lib/python3.10/site-packages/k2/rnnt_decode.py__init__!s zRnntDecodingStream.__init__cCs|jd|jdS)aReturn a string representation of this object For visualization and debug only. z , device :  )rrrrrr__str__5szRnntDecodingStream.__str__)__name__ __module__ __qualname__rrstrrrrrrr sr c@seZdZdZdeededdfddZdefdd Z de e e ffd d Z d e ddfd dZdddZ ddeededefddZdS)RnntDecodingStreamszSee https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py # noqa for how this class is used in RNN-T decoding. src_streamsconfigr NcCsPt|dksJt||_||_|jdj|_dd|jD}t|||_dS)a! Combines multiple RnntDecodingStream objects to create a RnntDecodingStreams object, then all these RnntDecodingStreams can do decoding in parallel. Args: src_streams: A list of RnntDecodingStream object to be combined. config: A configuration object which contains decoding parameters like `vocab-size`, `decoder_history_len`, `beam`, `max_states`, `max_contexts` etc. Returns: Return a RnntDecodingStreams object. rcSsg|]}|jqSr)r).0xrrr Ysz0RnntDecodingStreams.__init__..N)len num_streamsr rrrstreams)rr r!r'rrrrBs  zRnntDecodingStreams.__init__cCs<d|jd}t|jD]}|d|d|j|7}q |S)rznum_streams : rzstream[z] : )r&ranger )rsirrrr\szRnntDecodingStreams.__str__cCs |jS)a= This function must be called prior to evaluating the joiner network for a particular frame. It tells the calling code for which contexts it must evaluate the joiner network. Returns: Return a two-element tuple containing a RaggedShape and a tensor. shape: A RaggedShape with 2 axes, representing [stream][context]. contexts: A tensor of shape [tot_contexts][decoder_history_len], where tot_contexts == shape->TotSize(1) and decoder_history_len comes from the config, it represents the number of symbols in the context of the decoder network (assumed to be finite). It contains the token ids into the vocabulary(i.e. `0 <= value < vocab_size`). Its dtype is torch.int32. )r' get_contextsrrrrr+fs z RnntDecodingStreams.get_contextslogprobscCs|j|dS)ad Advance decoding streams by one frame. Args: logprobs: A tensor of shape [tot_contexts][num_symbols], containing log-probs of symbols given the contexts output by `get_contexts()`. It satisfies `logprobs.Dim0() == shape.TotSize(1)`, shape is returned by `get_contexts()`. N)r'advance)rr,rrrr-|s zRnntDecodingStreams.advancecCs|jdS)aX Terminate the decoding process of current RnntDecodingStreams object. It will update the decoding states and store the decoding results currently got to each of the individual streams. Note: We can not decode with this object anymore after calling terminate_and_flush_to_streams(). N)r'terminate_and_flush_to_streamsrrrrr.s z2RnntDecodingStreams.terminate_and_flush_to_streamsF num_frames allow_partialcCst||jks J|j||\}}t|}t}t|jD]O}|j|j}|j ddD]@\} } | |vrld} t | t rHt | | } | j} d} nt | tjsPJ| jtjksXJ| jdks_Jtj} d} | | | d|| <q,q|D]\} }t}d}t|jD]}|j|j}|j}||j}||||}||}t|| rt|| } |ddkrt | t sJt| || d }nCt | tsJ| jdksJ| jtjksJ| j|ddd \}}n"|ddkrtj| g||d |d }nttj|df|d |d }||q|ddkrt|}ntj j|dd }t!|| |qrt|jD]}|j|j}|"D] \} } t!|| | q1q%|S)a Generate the lattice Fsa currently got. Note: The attributes of the generated lattice is a union of the attributes of all the decoding graphs. For example, if `self` contains three individual stream, each stream has its own decoding graphs, graph[0] has attributes attr1, attr2; graph[1] has attributes attr1, attr3; graph[2] has attributes attr3, attr4; then the generated lattice has attributes attr1, attr2, attr3, attr4. Args: num_frames: A List containing the number of frames we want to gather for each stream (note: the frames we have ever received for the corresponding stream). It MUST satisfy `len(num_frames) == self.num_streams`. allow_partial: If true and there is no final state active, we will treat all the states on the last frame to be final state. If false, we only care about the real final state in the decoding graph on the last frame when generating lattice. Default False. Returns: Return the lattice Fsa with all the attributes propagated. The returned Fsa has 3 axes with `fsa.dim0==self.num_streams`. F)include_scoresrrr)fillerdtype tensor_typer5) default_value)axisneed_value_indexesr4)r4r)r7)#r%r&r' format_outputrdictr(r r named_tensor_attr isinstancerfloat get_fillerr4k2rtorchint32num_axesitemslistrnum_arcshasattrgetattrr indextensoremptyappendcatraggedsetattrnamed_non_tensor_attr)rr/r0 ragged_arcsout_mapr tensor_attr_infor*srcnamevaluer3r4r5infovaluesstartrrEarc_map new_value_rrrr9s             z!RnntDecodingStreams.format_output)r N)F)rrr__doc__rr r rrrrrrr+r-r.intboolrr9rrrrr=s*   r)typingrrr?r@rrrrropsr r objectr rrrrrs