tgt_vocab=tgt_vocab,
encoder=encoder,
decoder=decoder,
share_embed=args.dataset != 'TOY',
embed_size=args.num_units,
tie_weights=args.dataset != 'TOY',
embed_initializer=None,
prefix='transformer_')
model.initialize(init=mx.init.Xavier(magnitude=args.magnitude), ctx=ctx)
static_alloc = True
model.hybridize(static_alloc=static_alloc)
logging.info(model)
translator = BeamSearchTranslator(model=model,
beam_size=args.beam_size,
scorer=BeamSearchScorer(alpha=args.lp_alpha,
K=args.lp_k),
max_length=200)
logging.info('Use beam_size={}, alpha={}, K={}'.format(args.beam_size,
args.lp_alpha,
args.lp_k))
label_smoothing = LabelSmoothing(epsilon=args.epsilon, units=len(tgt_vocab))
label_smoothing.hybridize(static_alloc=static_alloc)
loss_function = SoftmaxCEMaskedLoss(sparse_label=False)
loss_function.hybridize(static_alloc=static_alloc)
test_loss_function = SoftmaxCEMaskedLoss()
test_loss_function.hybridize(static_alloc=static_alloc)
num_bi_layers=args.num_bi_layers)
model = NMTModel(src_vocab=src_vocab,
tgt_vocab=tgt_vocab,
encoder=encoder,
decoder=decoder,
one_step_ahead_decoder=one_step_ahead_decoder,
embed_size=args.num_hidden,
prefix='gnmt_')
model.initialize(init=mx.init.Uniform(0.1), ctx=ctx)
static_alloc = True
model.hybridize(static_alloc=static_alloc)
logging.info(model)
translator = BeamSearchTranslator(model=model,
beam_size=args.beam_size,
scorer=nlp.model.BeamSearchScorer(
alpha=args.lp_alpha, K=args.lp_k),
max_length=args.tgt_max_len + 100)
logging.info('Use beam_size={}, alpha={}, K={}'.format(args.beam_size,
args.lp_alpha,
args.lp_k))
loss_function = MaskedSoftmaxCELoss()
loss_function.hybridize(static_alloc=static_alloc)
def evaluate(data_loader):
"""Evaluate given the data loader
Parameters
----------
