def param_init_gru(options,
params,
prefix='gru',
nin=None,
dim=None,
hiero=False):
if nin is None:
nin = options['dim_proj']
if dim is None:
dim = options['dim_proj']
if not hiero:
W = numpy.concatenate([norm_weight(nin, dim),
norm_weight(nin, dim)], axis=1)
params[prfx(prefix, 'W')] = W
params[prfx(prefix, 'b')] = numpy.zeros((2 * dim,)).astype('float32')
U = numpy.concatenate([ortho_weight(dim),
ortho_weight(dim)], axis=1)
params[prfx(prefix, 'U')] = U
Wx = norm_weight(nin, dim)
params[prfx(prefix, 'Wx')] = Wx
Ux = ortho_weight(dim)
params[prfx(prefix, 'Ux')] = Ux
params[prfx(prefix, 'bx')] = numpy.zeros((dim,)).astype('float32')
return params
def param_init_lstm_tied(options,
params,
prefix='lstm_tied',
nin=None,
dim=None):
if nin is None:
nin = options['dim_proj']
if dim is None:
dim = options['dim_proj']
W = numpy.concatenate([norm_weight(nin, dim),
norm_weight(nin, dim),
norm_weight(nin, dim)], axis=1)
params[prfx(prefix, 'W')] = W
U = numpy.concatenate([ortho_weight(dim),
ortho_weight(dim),
ortho_weight(dim)], axis=1)
params[prfx(prefix, 'U')] = U
params[prfx(prefix, 'b')] = numpy.zeros((3 * dim,)).astype('float32')
return params
def param_init_fflayer(options,
params,
prefix='ff',
nin=None,
nout=None,
ortho=True,
use_bias=True):
if nin is None:
nin = options['dim_proj']
if nout is None:
nout = options['dim_proj']
params[prfx(prefix, 'W')] = norm_weight(nin, nout, scale=0.01, ortho=ortho)
if use_bias:
params[prfx(prefix, 'b')] = numpy.zeros((nout,)).astype('float32')
return params
def param_init_gru_cond(options,
params,
prefix='gru_cond',
nin=None,
dim=None,
dimctx=None):
if nin is None:
nin = options['dim']
if dim is None:
dim = options['dim']
if dimctx is None:
dimctx = options['dim']
params = param_init_gru(options,
params,
prefix,
nin=nin,
dim=dim)
# context to LSTM
Wc = norm_weight(dimctx, dim*2)
params[prfx(prefix, 'Wc')] = Wc
Wcx = norm_weight(dimctx, dim)
params[prfx(prefix, 'Wcx')] = Wcx
# attention: prev -> hidden
Wi_att = norm_weight(nin, dimctx)
params[prfx(prefix, 'Wi_att')] = Wi_att
# attention: context -> hidden
Wc_att = norm_weight(dimctx)
params[prfx(prefix, 'Wc_att')] = Wc_att
# attention: LSTM -> hidden
Wd_att = norm_weight(dim, dimctx)
params[prfx(prefix, 'Wd_att')] = Wd_att
# attention: hidden bias
b_att = numpy.zeros((dimctx,)).astype('float32')
params[prfx(prefix, 'b_att')] = b_att
# attention:
U_att = norm_weight(dimctx, 1)
params[prfx(prefix, 'U_att')] = U_att
c_att = numpy.zeros((1,)).astype('float32')
params[prfx(prefix, 'c_tt')] = c_att
return params
def init_params(options):
params = OrderedDict()
# embedding
params['Wemb_word'] = norm_weight(options['n_words_q'],
options['dim_word_desc'])
mult = 2
if options['ms_nlayers'] > 1 and (options['encoder_desc'] == 'lstm_ms' or \
options['encoder_desc'] == 'lstm_max_ms'):
mult = options['ms_nlayers']
if options['use_bidir']:
mult *= 2
if options['use_dq_sims']:
params['ff_att_bi_dq'] = \
norm_weight(mult * options['dim'],
mult * options['dim'])
params['ff_att_proj'] = norm_vec(options['dim'])
# encoder: bidirectional RNN
params = get_layer(options['encoder_desc_word'])[0](options,
params,
prefix='encoder_desc_word',
nin=options['dim_word_desc'],
dim=options['dim'])
params = get_layer(options['encoder_q'])[0](options,
params,
prefix='encoder_q',
nin=options['dim_word_q'],
dim=options['dim'])
if options['use_bidir']:
params = get_layer(options['encoder_desc_word'])[0](options,
params,
prefix='encoder_desc_word_r',
nin=options['dim_word_desc'],
dim=options['dim'])
params = get_layer(options['encoder_q'])[0](options,
params,
prefix='encoder_q_r',
nin=options['dim_word_q'],
dim=options['dim'])
if options['use_sent_reps']:
params['Wemb_sent'] = norm_weight(mult * options['dim'],
mult * options['dim'])
# encoder: bidirectional RNN
params = get_layer(options['encoder_desc_sent'])[0](options,
params,
prefix='encoder_desc_sent',
nin=mult * options['dim'],
dim=options['dim'])
if options['use_bidir']:
params = get_layer(options['encoder_desc_sent'])[0](options,
params,
prefix='encoder_desc_sent_r',
nin=mult * options['dim'],
dim=options['dim'])
ctxdim = mult * options['dim']
logger.info("context dimensions is %d" % ctxdim)
params = get_layer('ff')[0](options, params,
prefix='ff_att_ctx',
nin=ctxdim,
nout=options['dim'])
# readout
params = get_layer('ff')[0](options, params,
prefix='ff_att_q',
nin=ctxdim,
nout=options['dim'],
use_bias=False,
ortho=False)
if options['use_desc_skip_c_g']:
# readout for mean pooled desc
params = get_layer('ff')[0](options, params,
prefix='ff_out_mean_d',
nin=ctxdim,
nout=options['dim_word_ans'],
use_bias=False,
ortho=False)
params = get_layer('ff')[0](options, params,
prefix='ff_out_q',
nin=ctxdim,
nout=options['dim_word_ans'],
ortho=False)
params = get_layer('ff')[0](options, params,
prefix='ff_out_ctx',
nin=ctxdim,
nout=options['dim_word_ans'],
use_bias=False,
ortho=False)
params = get_layer('ff')[0](options, params,
prefix='ff_logit',
nin=options['dim_word_ans'],
nout=options['n_words_ans'],
ortho=False)
return params
