3
EgG0 @ sn d dl mZ G dd dZdd Zdd Zdd d
Zd d Zd
d ZdddZdddZ dddZ
dd Z dS ) )brewc @ s e Zd Zeed\ZZZZdS )
AttentionType N) __name__
__module__ __qualname__tuplerangeZRegularZ RecurrentZDotZ SoftCoverage r
r
C/usr/local/lib64/python3.6/site-packages/caffe2/python/attention.pyr s r c C s dj t| t|S )Nz{}/{})formatstr)scopenamer
r
r s s r c C sB t j| ||gt|d}| jj||t|dgdd|gd\}}|S )N"attention_weighted_encoder_contextZ,attention_weighted_encoder_context_old_shape )shape)r Z
batch_mat_mulr netZReshape)modelencoder_outputs_transposedencoder_output_dimattention_weights_3dr r _r
r
r _calc_weighted_context s
r Nc C s> |d k r | j j||gdgdd}tj| |t|dddd}|S )NZmasked_attention_logitssequence)moder ZCUDNNr )Zengineaxis)r Z SequenceMaskr Zsoftmaxr )r attention_logits_transposedr encoder_lengthsr r
r
r _calc_attention_weights0 s r! c C sN | j j||}tj| |t|d|dddd}tj| |t|ddddgd}|S ) Nattention_logitsr T)dim_indim_outr Z freeze_biasr r )Zaxes)r ZTanhr fcr transpose)r "decoder_hidden_encoder_outputs_sumr r r" r r
r
r %_calc_attention_logits_from_sum_matchI s" r) c C s4 t j| |t||||dd}| jj||dgd}|S )Nr# )r$ r% r r )dims)r r& r r Squeeze)r inputr$ r% r r outputr
r
r _apply_fc_weight_for_sum_matchk s
r. c C s t | ||||dd} t | ||||dd}
| jj| |
gt|d} | jj|| gt|ddd} t| | ||d}
t| |
||d }t| ||||d
}||| gfS ) N weighted_prev_attention_context)r r, r$ r% r r weighted_decoder_hidden_state&decoder_hidden_encoder_outputs_sum_tmpr( r ) broadcast)r r( r r )r r r r )r r r r r )r. r Addr r) r! r )r r r weighted_encoder_outputsdecoder_hidden_state_tdecoder_hidden_state_dimZ)attention_weighted_encoder_context_t_prevr r r/ r0 r1 r( r r r r
r
r apply_recurrent_attention sR r7 c
C sp t | ||||dd}| jj||gt|dddd} t| | ||d}
t| |
||d} t| ||| |d} | | | gfS ) Nr0 )r r, r$ r% r r r( r )r2 Z
use_grad_hack)r r( r r )r r r r )r r r r r )r. r r3 r r) r! r )
r r r r4 r5 r6 r r r0 r( r r r r
r
r apply_regular_attention s<
r8 c
C s ||kr&t j| |t|d||dd}n|}| jj|t|ddgd}| jj||dgd} | jj|| gt|ddd }
t| |
||d
} t| ||| |d } | | g fS ) Nr0 r# )r$ r% r &squeezed_weighted_decoder_hidden_stater )r* r" r )Ztrans_a)r r r r )r r r r r ) r r& r r r+ Z
ExpandDimsZ BatchMatMulr! r )
r r r r5 r6 r r r0 r9 Z1expanddims_squeezed_weighted_decoder_hidden_stater r r r
r
r apply_dot_attention sD
r: c
C s
t | ||||dd}
| jj||
gt|ddd} | jj|t|ddgd} tj| | t|d }
| jj| |
gt|d
ddd }| jj| |gt|d }t| |||d
}t | |||d}t
| ||||d}| jj|t|ddgd}| jj||gt|ddd}|||g|fS )Nr0 )r r, r$ r% r r r1 r )r2 coverage_t_prev_2dr )r* coverage_t_prev_transposedscaled_coverage_weights)r2 r r( )r r( r r )r r r r )r r r r r attention_weights_2dr#
coverage_t) r. r r3 r r+ r r' ZMulr) r! r )r r r r4 r5 r6 r r Zcoverage_t_prevZcoverage_weightsr0 r1 r; r< r= r( r r r r> r? r
r
r apply_soft_coverage_attentionF sn
r@ )N)N)N)N) Z
caffe2.pythonr r r r r! r) r. r7 r8 r: r@ r
r
r
r