o Gi%@sNddlmZddlZddlZddlmZGdddeZdedefd d ZdS) )UnionN)Fsac seZdZ ddejdejdeddfddZedejfd d Ze d e j d ejddffd d Z defddZ defddZedejfddZdefddZdefddZdeejefddfddZZS) DenseFsaVecr log_probssupervision_segmentsallow_truncatereturnNcCs>|jdksJ|jtjksJ|jdksJ|jtjksJ|jjdks&J|dks,J|j\}}}tdg}g}g} d} | D]_} | \} } }d| krT|ksWJJd| krb|kseJJ|dkskJ| |||ksuJ| |}t | ||}|| }| t | ||| || |7} | | | d7} qB|j}t ||}tj| |d|j|d}|d|d||ddddf<td |dddf<tjdgtd g||d || <tj|dddtjd }t| d|dd<|dd|ddd|_||}t|||_||_dS) aConstruct a DenseFsaVec from neural net log-softmax outputs. Args: log_probs: A 3-D tensor of dtype `torch.float32` with shape `(N, T, C)`, where `N` is the number of sequences, `T` the maximum input length, and `C` the number of output classes. supervision_segments: A 2-D **CPU** tensor of dtype `torch.int32` with 3 columns. Each row contains information for a supervision segment. Column 0 is the `sequence_index` indicating which sequence this segment comes from; column 1 specifies the `start_frame` of this segment within the sequence; column 2 contains the `duration` of this segment. Note: - `0 < start_frame + duration <= T + allow_truncate` - `0 <= start_frame < T` - `duration > 0` Caution: If the resulting dense fsa vec is used as an input to `k2.intersect_dense`, then the last column, i.e., the duration column, has to be sorted in **decreasing** order. That is, the first supervision_segment (the first row) has the largest duration. Otherwise, you don't need to sort the last column. `k2.intersect_dense` is often used in the training stage, so you should usually sort dense fsa vecs by its duration in training. `k2.intersect_dense_pruned` is usually used in the decoding stage, so you don't need to sort dense fsa vecs in decoding. allow_truncate: If not zero, it truncates at most this number of frames from duration in case start_frame + duration > T. cpurr)dtypedeviceNz-inf)r)rr )ndimr torchfloat32int32rtypeshapetensortolistminappendarangecattoemptyreshape index_selectfloatzerossize _duration_k2r dense_fsa_vecscores)selfrrrNTC placeholderindexeslast_frame_indexescursegment segment_index start_framedurationoffset end_framerr& row_splitsr6D/home/ubuntu/.local/lib/python3.10/site-packages/k2/dense_fsa_vec.py__init__+sT)         $   zDenseFsaVec.__init__cCsFt|ds |jd}|dd|ddd}||_|jS)z2Return the duration (on CPU) of each seq. r#rNr)hasattrr%rr5r r#)r'r5r2r6r6r7r2s  zDenseFsaVec.durationr%r&cs(||}tt|||_||_|S)ajConstruct a DenseFsaVec from `_k2.DenseFsaVec` and `scores`. Note: It is intended for internal use. Users will normally not use it. Args: dense_fsa_vec: An instance of `_k2.DenseFsaVec`. scores: The `scores` of `_k2.DenseFsaVec` for back propagation. Return: An instance of DenseFsaVec. )__new__superrr8r%r&)clsr%r&ans __class__r6r7_from_dense_fsa_vecs zDenseFsaVec._from_dense_fsa_veccC |jS)z&Return number of supervision segments.)r%dim0r'r6r6r7rBs zDenseFsaVec.dim0cCrAN)r%to_strrCr6r6r7__str__s zDenseFsaVec.__str__cCs|jjSrD)r&rrCr6r6r7rszDenseFsaVec.devicecC |jjdkS)zReturn true if this DenseFsaVec is on CPU. Returns: True if the DenseFsaVec is on CPU; False otherwise. r rrrCr6r6r7is_cpu zDenseFsaVec.is_cpucCrG)zReturn true if this DenseFsaVec is on GPU. Returns: True if the DenseFsaVec is on GPU; False otherwise. cudarHrCr6r6r7is_cudarJzDenseFsaVec.is_cudarcCsVt|tr t|}|jdvsJ||jjkr|S|j|}|j|}t ||S)aMove the DenseFsaVec onto a given device. Args: device: An instance of `torch.device` or a string that can be used to construct a `torch.device`, e.g., 'cpu', 'cuda:0'. It supports only cpu and cuda devices. Returns: Returns a new DenseFsaVec which is this object copied to the given device (or this object itself, if the device was the same). )r rK) isinstancestrrrrr&rr%rr@)r'rr&r%r6r6r7rs     zDenseFsaVec.to)r)__name__ __module__ __qualname__rTensorintr8propertyr2 classmethodr$rr@rBrNrFrboolrIrLrr __classcell__r6r6r>r7rs4 h $rr%r cCst|j}t|S)aConvert a DenseFsaVec to an FsaVec. Caution: Intended for use in testing/debug mode only. This operation is NOT differentiable. Args: dense_fsa_vec: DenseFsaVec to convert. Returns: The converted FsaVec . )r$convert_dense_to_fsa_vecr%r)r% ragged_arcr6r6r7rXs rX) typingrrr$fsarobjectrrXr6r6r6r7s  P