S^(hp d dl mZmZ d dlmZ d dlmZmZmZm Z m
Z
d dl Z d dl
Z
ddlmZ ddlmZmZ d Zd
Z G d d e Z G d
de Z G d de Z G d d Zy) )ABCabstractmethod)UserDict)DictListOptionalTupleUnionN )add_start_docstrings )
ConstraintConstraintListStatea
Args:
input_ids (`torch.LongTensor` of shape `(batch_size * num_beams, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using any class inheriting from [`PreTrainedTokenizer`]. See
[`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
next_scores (`torch.FloatTensor` of shape `(batch_size, 2 * num_beams)`):
Current scores of the top `2 * num_beams` non-finished beam hypotheses.
next_tokens (`torch.LongTensor` of shape `(batch_size, 2 * num_beams)`):
`input_ids` of the tokens corresponding to the top `2 * num_beams` non-finished beam hypotheses.
next_indices (`torch.LongTensor` of shape `(batch_size, 2 * num_beams)`):
Beam indices indicating to which beam hypothesis the `next_tokens` correspond.
pad_token_id (`int`, *optional*):
The id of the *padding* token.
eos_token_id (`Union[int, List[int]]`, *optional*):
The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
beam_indices (`torch.LongTensor`, *optional*):
Beam indices indicating to which beam hypothesis each token correspond.
group_index (`int`, *optional*):
The index of the group of beams. Used with [`~PreTrainedModel.group_beam_search`].
Return:
`UserDict`: A dictionary composed of the fields as defined above:
- **next_beam_scores** (`torch.FloatTensor` of shape `(batch_size * num_beams)`) -- Updated scores of all
non-finished beams.
- **next_beam_tokens** (`torch.FloatTensor` of shape `(batch_size * num_beams)`) -- Next tokens to be added
to the non-finished beam_hypotheses.
- **next_beam_indices** (`torch.FloatTensor` of shape `(batch_size * num_beams)`) -- Beam indices
indicating to which beam the next tokens shall be added.
a
Args:
input_ids (`torch.LongTensor` of shape `(batch_size * num_beams, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using any class inheriting from [`PreTrainedTokenizer`]. See
[`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
final_beam_scores (`torch.FloatTensor` of shape `(batch_size * num_beams)`):
The final scores of all non-finished beams.
final_beam_tokens (`torch.FloatTensor` of shape `(batch_size * num_beams)`):
The last tokens to be added to the non-finished beam_hypotheses.
final_beam_indices (`torch.FloatTensor` of shape `(batch_size * num_beams)`):
The beam indices indicating to which beam the `final_beam_tokens` shall be added.
pad_token_id (`int`, *optional*):
The id of the *padding* token.
eos_token_id (`Union[int, List[int]]`, *optional*):
The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
Return:
`torch.LongTensor` of shape `(batch_size * num_return_sequences, sequence_length)`: The generated sequences.
The second dimension (sequence_length) is either equal to `max_length` or shorter if all batches finished early
due to the `eos_token_id`.
c Z e Zd ZdZe ee dej dej dej dej de
ej f
d Z e ee
dej dej dej dej dedej f d Zy
)
BeamScorerz
Abstract base class for all beam scorers that are used for [`~PreTrainedModel.beam_search`] and
[`~PreTrainedModel.beam_sample`].
input_ids next_scores next_tokens next_indicesreturnc t d NzThis is an abstract method.NotImplementedError)selfr r r r kwargss a/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/generation/beam_search.pyprocesszBeamScorer.processa s ""?@@
max_lengthc t d r r )r r r r r r r s r finalizezBeamScorer.finalizem s ""?@@r N)__name__
__module__ __qualname____doc__r r PROCESS_INPUTS_DOCSTRINGtorch
LongTensor FloatTensorr Tensorr FINALIZE_INPUTS_DOCSTRINGintr" r r r r [ s
23A##A &&A %% A
&& A
u|| A 4 A 34 A## A && A %% A
&& A
A
A 5 Ar r c e Zd ZdZ d dededej dee dee e
e f dee d ee d
ee fd Z e
d e
fd
Z d dej dej dej dej dee eej" f dee eee ej" f deej dee dee d ee ej" f fdZ d dej dej dej dej d
edee eej" f dee eee ej" f deej dee d eej fdZy) BeamSearchScorera
[`BeamScorer`] implementing standard beam search decoding.
Adapted in part from [Facebook's XLM beam search
code](https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529).
Reference for the diverse beam search algorithm and implementation [Ashwin Kalyan's DBS
implementation](https://github.com/ashwinkalyan/dbs/blob/master/dbs/beam_utils.lua)
Args:
batch_size (`int`):
Batch Size of `input_ids` for which standard beam search decoding is run in parallel.
num_beams (`int`):
Number of beams for beam search.
device (`torch.device`):
Defines the device type (*e.g.*, `"cpu"` or `"cuda"`) on which this instance of `BeamSearchScorer` will be
allocated.
length_penalty (`float`, *optional*, defaults to 1.0):
Exponential penalty to the length that is used with beam-based generation. It is applied as an exponent to
the sequence length, which in turn is used to divide the score of the sequence. Since the score is the log
likelihood of the sequence (i.e. negative), `length_penalty` > 0.0 promotes longer sequences, while
`length_penalty` < 0.0 encourages shorter sequences.
do_early_stopping (`bool` or `str`, *optional*, defaults to `False`):
Controls the stopping condition for beam-based methods, like beam-search. It accepts the following values:
`True`, where the generation stops as soon as there are `num_beams` complete candidates; `False`, where an
heuristic is applied and the generation stops when is it very unlikely to find better candidates;
`"never"`, where the beam search procedure only stops when there cannot be better candidates (canonical
beam search algorithm).
num_beam_hyps_to_keep (`int`, *optional*, defaults to 1):
The number of beam hypotheses that shall be returned upon calling
[`~transformers.BeamSearchScorer.finalize`].
num_beam_groups (`int`, *optional*, defaults to 1):
Number of groups to divide `num_beams` into in order to ensure diversity among different groups of beams.
See [this paper](https://arxiv.org/pdf/1610.02424.pdf) for more details.
max_length (`int`, *optional*):
The maximum length of the sequence to be generated.
N
batch_size num_beamsdevicelength_penaltydo_early_stoppingnum_beam_hyps_to_keepnum_beam_groupsr c || _ || _ || _ || _ || _ || _ | j | j
z | _ d| _ t || j
z D cg c]/ } t | j | j | j | 1 c} | _
t j t || j
z D cg c] } d c} t j | j | _ t |t r|dk rt# d| d t |t r
||kD s||z dk7 rt# d| d | d
y c c} w c c} w NF)r2 r4 early_stoppingr dtyper3 r
zA`num_beams` has to be an integer strictly greater than 1, but is zG. For `num_beams` == 1, one should make use of `greedy_search` instead.r z`num_beam_groups` has to be an integer smaller or equal than `num_beams` and `num_beams` has to be divisible by `num_beam_groups`, but is z with `num_beams` being .)r2 r3 r4 r5 r6 r7
group_size_is_initrangeBeamHypotheses
_beam_hypsr( tensorbool_done
isinstancer-
ValueError)
r r1 r2 r3 r4 r5 r6 r7 r _s
r __init__zBeamSearchScorer.__init__ sp # ,!2 %:" . ..D,@,@@
:(<(<<=
// # 2 2 # 5 5%
\\ !*t/C/C"C D EqU EUZZ`d`k`k
)S)Y!^ ST]S^ _C C
/3/Oi4OU^apUptuUu ;;J:KKcdmcnnoq
Vv+
Fs *4E Er c 6 | j j S NrE allr s r is_donezBeamSearchScorer.is_done zz~~ r r r r r pad_token_id eos_token_id beam_indices group_indexdecoder_prompt_lenc
|j d dz }
t | j | j z } | |j d | j z k( sa| j dkD r)t d|j d d| j d t d|j d d| j d |j } t j | | j f|j | }
t j | | j f|j | }t j | | j f|j | }|Bt |t j s(t |t r|g}t j | }t | D ] }|| j z |z }| j | ri| j t | j | k rt d
| j d || t d d|
|d d f< |||d d f< d||d d f< d}t# t% || || || D ] \ }\ }}}|| j z |z }|y|j' |v rg|| j k\ }|rB| || }||fz }nd }| j | j) || j+ |j' ||
| z
n||
||f< ||||f< ||||f< |dz
}|| j k( s n || j k r(t d| j d|| d| d|| d | j | xs? | j | j- || j/ j' |
| | j |< t1 |
j3 d |j3 d |j3 d d S )Nr
r A group beam size of 0 is used as the input, but a group beam size of is expected by the beam scorer.A beam size of * is used as the input, but a beam size of r; #Batch can only be done if at least beams have been generatedMGenerated beams >= num_beams -> eos_token_id and pad_token have to be definedrS
generated_lenAt most tokens in can be equal to `eos_token_id:
`. Make sure are corrected.next_beam_scoresnext_beam_tokensnext_beam_indices)shapelenrB r7 r> rG r3 r( zerosr< rF r+ r- rC r@ rE r2 enumeratezipitemaddclonerO maxr view)r r r r r rQ rR rS rT rU cur_lenr1 r3 rh ri rj batch_idxbatch_group_idxbeam_idxbeam_token_rank
next_token
next_score
next_indexbatch_beam_idx&is_beam_token_worse_than_top_num_beams
beam_indexs r r zBeamSearchScorer.process s= //"%) )T-A-AA
yq 1T__ DE##a' +IOOA,>+? @ # //OQ
!%iooa&8%99c''GI
!! ;;
DOO'DKL]L]flm ;;
DOO'DKL]L]flm !KKT__(E\M_M_hn o #J|U\\,R,, ,~ << 5L z* 9 I '$*>*> > LOzz/*>>C(H$II$'J4>>JZZt%uvv '<+?$%tuu12 A .1= A .23!)Q,/ HIRK *K ,BLQZD[\J
E!E*j* "+T__!2C%C
%)
OOO488!.1779")%/&-0B&B 9 =G$Y%89PJDJJ 'o9 v $4$9$9"$=$4$9$9"$=%6%;%;B%?
r final_beam_scoresfinal_beam_tokensfinal_beam_indicesc
t | j | j z }
|Bt |t j
s(t |t r|g}t j | }t | j D ] \ } } | j | rt | j D ]\ }
| | j z |
z }|| j }|| }||| nd }|j d | z
}| j |||| ^ |j |
| j z }g }g }t j" |
| j z | j$ t j& }t |
D ] }| j || j z |dz | j z }|D cg c] } | j( D ] }| }} }t+ |d }t | j D ]v } |j- } | d } | d } | d } t | || j |z | z < |j/ | |j/ | | ||| j z | z < x |j1 j dz }!| t3 |!| n|!}"|j |
| j z |" }#t | dkD r%|d |j |
| j z |" }$nd }$|j3 j |j1 j k7 r | t5 d |#j7 | |$|$j7 d t t9 || D ]Q \ }\ }%}&|%|#|d || f< |$%t j |& |$|d t |& f< || |"k sE|d |#||| f< S t; |#||$d
S c c}} w ) NrW r` r3 r< r
c | d S Nr r. xs r z+BeamSearchScorer.finalize..l s
! r keyr r `pad_token_id` has to be defined sequencessequence_scoresrS ) rl rB r7 rF r( r+ r- rC rn rE r@ r> rp rk rq newr6 rm r3 float32beamssortedpopappendrs minrG fill_ro r )'r r r r r r rQ rR rS rU r1 rw beam_hypindex_per_groupr} final_score final_tokensr ra sent_lengthsbest best_indices best_scoresibeam_hyps_in_batchbeamcandidate_beams sorted_hypsjbest_hyp_tuple
best_scorebest_hyp
best_indexsent_lengths_max sent_max_lendecodedindiceshypobest_idxs' r r" zBeamSearchScorer.finalize@ s )T-A-AA
#J|U\\,R,, ,~ << 5L *34??)C n %OXzz/* $) #9
n!04??!B_!T / ? D D F ( 8 =I=U\. 9[_
, 2 22 69K K
\;Z_l m
n
n !}}Z$2L2L%LM kk*t/I/I"IRVR]R]ejerers z" MA!% T5I5I1IQQRUVZVjVjLj!k 4F bS[SaSa b4t bt bO b nEK4556
M!, !2 +A .
)!, +A .
CFx= T 7 7! ;a ?@ H% ##J/BL A : : :Q >?
M M( (++-224q8 A h,0GA( a (( )".3ll8.D?S]?*+A -.:1o<?*+ > $#. '
_ cs &O g ?Fr
r
N)NNNr r NNNr )r# r$ r% r& r- r( r3 r floatr
rD strrI propertyrO r) r* r+ r r r r r" r. r r r0 r0 { ss $V +.8=/0)*$(// / /
! / $E$)$45
/ (}/ "# / SM/b <@FJ37%&,-g
##g
&&g
%% g
&& g
uS%,,%6 78
g
uS$s)U\\%A BCg
u / /0g
c]g
%SMg
c5<< g
` <@FJ37,-`
##`
!,,`
!++ `
" , , `
`
uS%,,%6 78`
uS$s)U\\%A BC`
u / /0`
%SM`
u `
r r0 c t e Zd ZdZ d&dededee dej de e
de e e e
f d e e d
e e d e e fd Zed
e fd Zd Zd Z d'dej& dej( dej& dej& dej( de e eej* f de e eee ej* f de ej& de e d
eej* fdZ d(dedej& d ej( d ej( d ej& d ej& d e fd!Z d'dej& d"ej( d#ej& d$ej& d ede e eej* f de e eee ej* f de ej& de e d
eej& fd%Zy))ConstrainedBeamSearchScorera
[`BeamScorer`] implementing constrained beam search decoding.
Args:
batch_size (`int`):
Batch Size of `input_ids` for which standard beam search decoding is run in parallel.
num_beams (`int`):
Number of beams for beam search.
constraints (`List[Constraint]`):
A list of positive constraints represented as `Constraint` objects that must be fulfilled in the generation
output. For more information, the documentation of [`Constraint`] should be read.
device (`torch.device`):
Defines the device type (*e.g.*, `"cpu"` or `"cuda"`) on which this instance of `BeamSearchScorer` will be
allocated.
length_penalty (`float`, *optional*, defaults to 1.0):
Exponential penalty to the length that is used with beam-based generation. It is applied as an exponent to
the sequence length, which in turn is used to divide the score of the sequence. Since the score is the log
likelihood of the sequence (i.e. negative), `length_penalty` > 0.0 promotes longer sequences, while
`length_penalty` < 0.0 encourages shorter sequences.
do_early_stopping (`bool` or `str`, *optional*, defaults to `False`):
Controls the stopping condition for beam-based methods, like beam-search. It accepts the following values:
`True`, where the generation stops as soon as there are `num_beams` complete candidates; `False`, where an
heuristic is applied and the generation stops when is it very unlikely to find better candidates;
`"never"`, where the beam search procedure only stops when there cannot be better candidates (canonical
beam search algorithm).
num_beam_hyps_to_keep (`int`, *optional*, defaults to 1):
The number of beam hypotheses that shall be returned upon calling
[`~transformers.BeamSearchScorer.finalize`].
num_beam_groups (`int`, *optional*, defaults to 1):
Number of groups to divide `num_beams` into in order to ensure diversity among different groups of beams.
See [this paper](https://arxiv.org/pdf/1610.02424.pdf) for more details.
max_length (`int`, *optional*):
The maximum length of the sequence to be generated.
Nr1 r2 constraintsr3 r4 r5 r6 r7 r c
|| _ || _ || _ || _ || _ || _ | j | j
z | _ || _ d| _ t | D
cg c]/ }
t | j | j | j | 1 c}
| _ t j t | D
cg c] }
d c}
t j | j | _ t! |t" r|dk rt% d| d t! |t" r
||kD s||z dk7 rt% d| d | d
y c c}
w c c}
w r9 )r2 r3 r4 r5 r6 r7 r> r r? r@ rA rB r( rC rD rE rF r- rG ) r r1 r2 r r3 r4 r5 r6 r7 r rH s r rI z$ConstrainedBeamSearchScorer.__init__ s] # ,!2 %:" . ..D,@,@@&
:&
.. # 2 2 # 5 5%
\\%
2C"DQ5"DEJJ_c_j_jk
)S)Y!^ ST]S^ _C C
/3/Oi4OU^apUptuUu ;;J:KKcdmcnnoq
Vv#
#Es $4E < Er c 6 | j j S rK rL rN s r rO z#ConstrainedBeamSearchScorer.is_done rP r c t | D cg c]3 }t | j D cg c] }|j c} 5 c}}S c c}w c c}}w rK )r@ r r copy)r nrH
constraints r make_constraint_statesz2ConstrainedBeamSearchScorer.make_constraint_states s> ejklemn`a#IYIY$Z:Z__%6$Z[nn$Zns AA A Ac d | j d d }|j | |j S )Nr
r )r reset completed)r sequence new_states r check_completes_constraintsz7ConstrainedBeamSearchScorer.check_completes_constraints s/ //215 !"""r r r r r scores_for_all_vocabrQ rR rS rU c
|j d dz }
t | j } | |j d | j z k( sa| j dkD r)t d|j d d| j d t d|j d d| j d |j } t j | | j f|j | }
t j | | j f|j | }t j | | j f|j | }|Bt |t j s(t |t r|g}t j | }t | j D ]& \ }}| j | r\| j t | k rt d | j d || t d
d|
|d
d
f< |||d
d
f< d||d
d
f< rd}t t# || || || D ] \ }\ }}}|| j z |z }||j% |v r~|| j k\ }|rB| j' || j) }|rb| || }||fz }nd
}|j+ || j- |j% ||
| z
n||
||f< ||||f< ||||f< |dz
}|| j k( s n | j/ ||||
| || || \ }} } ||
|< | ||< | ||< || j k r(t d| j d|| d| d|| d | j | xs2 |j1 || j3 j% |
| | j |< ) t5 |
j7 d |j7 d |j7 d d S )al
Args:
input_ids (`torch.LongTensor` of shape `(batch_size * num_beams, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using any class inheriting from [`PreTrainedTokenizer`]. See
[`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
next_scores (`torch.FloatTensor` of shape `(batch_size, 2 * num_beams)`):
Current scores of the top `2 * num_beams` non-finished beam hypotheses.
next_tokens (`torch.LongTensor` of shape `(batch_size, 2 * num_beams)`):
`input_ids` of the tokens corresponding to the top `2 * num_beams` non-finished beam hypotheses.
next_indices (`torch.LongTensor` of shape `(batch_size, 2 * num_beams)`):
Beam indices indicating to which beam hypothesis the `next_tokens` correspond.
scores_for_all_vocab (`torch.FloatTensor` of shape `(batch_size * num_beams, sequence_length)`):
The scores of all tokens in the vocabulary for each of the beam hypotheses.
pad_token_id (`int`, *optional*):
The id of the *padding* token.
eos_token_id (`Union[int, List[int]]`, *optional*):
The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
beam_indices (`torch.LongTensor`, *optional*):
Beam indices indicating to which beam hypothesis each token correspond.
decoder_prompt_len (`int`, *optional*):
The length of prompt that is included in the input to decoder.
Return:
`UserDict`: A dictionary composed of the fields as defined above:
- **next_beam_scores** (`torch.FloatTensor` of shape `(batch_size * num_beams)`) -- Updated scores of
all
non-finished beams.
- **next_beam_tokens** (`torch.FloatTensor` of shape `(batch_size * num_beams)`) -- Next tokens to be
added
to the non-finished beam_hypotheses.
- **next_beam_indices** (`torch.FloatTensor` of shape `(batch_size * num_beams)`) -- Beam indices
indicating to which beam the next tokens shall be added.
rW r
r rX rY rZ r[ r\ r; Nr] r^ r_ r` rb rc rd re rf rg ) rk rl rB r> r7 rG r3 r( rm r< rF r+ r- rC rn rE r2 ro rp r tolistrq rr step_sentence_constraintrO rs r rt ) r r r r r r rQ rR rS rU ru r1 r3 rh ri rj rv r rx ry rz r{ r| r} r~ completes_constraintr
new_scores
new_tokens new_indicess r r z#ConstrainedBeamSearchScorer.process s h //"%) )
yq 1T__ DE##a' +IOOA,>+? @ # //OQ
!%iooa&8%99c''GI
!! ;;
DOO'DKL]L]flm ;;
DOO'DKL]L]flm !KKT__(E\M_M_hn o #J|U\\,R,, ,~ << 5L#,T__#= H Ixzz)$>>CM1$'J4>>JZZt%uvv '<+?$%tuu12 A .1= A .23!)Q,/ HIRK *K ,BLQZD[\J "
E!E*j* "+T__!P>DJJy !MH T $4$9$9"$=$4$9$9"$=%6%;%;B%?
r rv vocab_scoressent_beam_scoressent_beam_tokenssent_beam_indices
push_progressc B |j d }|j } | j | }
| j | } ||z |dz |z }
} || |
}|| |
}t j || |j d fd }|j
g g g g d}t | D ]\ \ }}|
| }|j || j
| | }|j |j
|j s$t j |j t j | }|D ] }|j d }|j |j
t j ||j d fd j
}||d vsq|d j | |d
j | |z |d j | |d j || j! | |d
j | |st j" || d \ }}t j ||j d fd }| | }|j
}|j | ||d vs|d j | |d
j | |d j | |d j | |d
j | _ t% |d
dkD rt j |d
| } t j& |d j) | } t j& |d j) | } |
|d
z } t j || fd } t j || fd }!t j | D "cg c] }"|"j+ c}"| }#|#dz |!z }$|$j- d j. }%|#|% }&d}'|&d }(g })|&D ]" }*|*|(k( r|'dz
}'nd}'|*}(|)j |' $ t j t1 j2 |)d }+|%|+ d | }%|!|% }| |% }t j || f |% }|||fS c c}"w )Nr r
rW )dim)new_seqs
new_statesr r r r; T)statefulr r r r r )r3 d )
descending mergesort)kind)sizer3 r r( cat unsqueezer rn r r rC advancelongr rq r takers rl stacktoget_banksortr npargsort),r rv r r r r r r orig_lenr3 topk_contraint_statesadvance_constraint_statessidxeidxthis_batch_input_idsthis_batch_token_scoresfull_hypotheses track_newseq_idxpre_seq
topk_state
advance_stateadvance_tokens
advance_tokenr advance_seq new_score new_tokenr r r
all_states
all_tokens
all_scoresone all_bankszippedr sorted_bankscountercur_bank
incrementsbank rearrangerss, r r z4ConstrainedBeamSearchScorer.step_sentence_constraint s ( %))!,")) !% ; ;H E $($?$?$I! )IMX+Ed (d 3 ".tD"9 ))Y/@%ACSC]C]^`Ca$bhjk (..0
!**> ? ? B GW /w7J _W 5 < < > ? 5g >M 0 1 **!& m.C.C.EUZZ`f!g%3 BM - 2 2D 2 AI MM-"6"6"89"'))Wm6M6Ma6P,QSU"V"]"]"_K")J*??!*-44[A!-077wG!,/66}E!,/667Nw7W7\7\]j7kl!,/66yA B ( (-yy1H1QST'U$ 9 #ii)2E2Ea2H(I2 N 9' B
) 0 0 2 ##K0 i
&;; j)00= m,33G< l+229= l+229= l+22=A? BB y ' (1 , ,,y '?OK Y|%<=@@HJ Y|%<=@@HJ.<1HHJ $4j#A2FJ $4j#A2FJ
%Kclln%KTZ[I _z1F kkTk2::G$W-L G#AHJ$
+8#qLG G #H!!'*
+ ,,rzz*;'OPK k*9H5G )' 2 )' 2 % +.j s
qt r r r r
r r r ) rl rB rF r( r+ r- rC rn rE r@ r2 rp r r rk rq r r6 r rm r3 r r r r rs r rG r ro r )&r r r r r r rQ rR rS rU r1 rv r ids_collectbeam_idr} r r r r ra r r r r r r r r r r r r r r r r r s& r r" z$ConstrainedBeamSearchScorer.finalize- st )
#J|U\\,R,, ,~ << 5L $-T__#= Ixzz)$
K 0
0!*T^^!;g!E / ? D D F ( 8 '+'G'G H[H[H]'^$'AMAY n!=_cJ$0$6$6r$:=O$OM LL {cpLq &&w/
0 ; $"<"<< $T^^ 4 G k1)2T^^)Cg)M&7&G&L&L&N '0 '@ (4(:(:2(>AS(S
\;m \;'4+E+EE / D !}}Z$2L2L%LM kk*t/I/I"IRVR]R]ejerers %T__5 MKAx ^DK4556
M!, !2 +A .
)!, +A .
CFx= T 7 7! ;a ?@ H% ##J/BL A : : :Q >?
M M$ (++-224q8 A h,0GA( a (( )".3ll8.D?S]?*+A -.:1o<?*+ > $#. '
r r r )F)r# r$ r% r& r- r r r( r3 r r r
rD r rI r rO r r r) r* r+ r r r r" r. r r r r s "T +.8=/0)*$(++ + *% +
+ !
+ $E$)$45+ (}+ "# + SM+Z o# <@FJ37,-]
##]
&&]
%% ]
&& ]
$ / /
]
uS%,,%6 78]
uS$s)U\\%A BC]
u / /0]
%SM]
u|| ]
N $KEKE ##KE '' KE
++ KE **
KE !++KE KEh <@FJ37,-r
##r
!,,r
!++ r
" , , r
r
uS%,,%6 78r
uS$s)U\\%A BCr
u / /0r
%SMr
u r
r r c
e Zd Zddedededee fdZd Z dde j d ed
ee j d ee fd Z dd
ededee defdZ y)rA Nr2 r4 r: r c || _ || _ || _ || _ g | _ d| _ t
| j t s| j t d yy)z7
Initialize n-best list of hypotheses.
g eANzWhen `do_early_stopping` is set to a string, `max_length` must be defined. Ensure it is passed to the BeamScorer class instance at initialization time.) r4 r: r r2 r worst_scorerF rD rG )r r2 r4 r: r s r rI zBeamHypotheses.__init__ sf - ,$"
$--t4 9P E
:Q4r c , t | j S )z3
Number of hypotheses in the list.
)rl r rN s r __len__zBeamHypotheses.__len__ s 4:: r hyp sum_logprobsrS ra c
|||| j z z }n ||j d | j z z }t | | j k s|| j kD r| j
j
|||f t | | j kD rZt t | j
D cg c]
\ }\ }}}||f c}}} } | j
| d d = | d d | _ yt || j | _ yyc c}}}w )z3
Add a new hypothesis to the list.
NrW r r
)
r4 rk rl r2 r r r r rn r )
r r r rS ra scoreidxsrH sorted_next_scoress
r rq zBeamHypotheses.add s $ M43F3F$FGE !CIIbMT5H5H$HIE t9t~~ %1A1A)A JJ uc< 8 94y4>>)%+IVZV`V`La,b,b.#y1aaX,b%c"JJ 1! 4Q 78#5a#8#; #&ud.>.>#? *B -cs 0D best_sum_logprobsru rU r c t | | j k ry| j du ry| j du r&|||z
| j z z }| j |k\ }|S | j dkD r:| j
|k r t
d || j
|z
| j z z }n|||z
| j z z }| j |k\ }|S )z
If there are enough hypotheses and that none of the hypotheses being generated can become better than the worst
one in the heap, then we are done with this sentence.
FTg z3max_length is not larger than decoder prompt length)rl r2 r: r4 r r rG )r r ru rU highest_attainable_scorerets r rO zBeamHypotheses.is_done s t9t~~ % $ &
E
)'8GFX>CJ ""S(??&88$%Z[[% ;M)MRVReRe(ee )
,=J\@\aeatat?t+t(""&>>CJr rK )NN)r )
r# r$ r% r- r rD r rI r
r( r) rq rO r. r r rA rA s # u d `hil`m " 48'+ @
@ @ u / /0 @
} @2" " "RZ[^R_ "hl "r rA )abcr r collectionsr typingr r r r r
numpyr r( utilsr beam_constraintsr r r' r, r r0 r rA r. r r r si $ 5 5 ( =" H 8 A A@e
z e
P |
* |
~S Sr