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) detokenizer = SacreMosesDetokenizer() def evaluate(data_loader, context=ctx[0]): """Evaluate given the data loader Parameters ---------- data_loader : DataLoader
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=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 = SoftmaxCEMaskedLoss() loss_function.hybridize(static_alloc=static_alloc) def evaluate(data_loader): """Evaluate given the data loader Parameters ---------- data_loader : DataLoader Returns ------- avg_loss : float Average loss real_translation_out : list of list of str
embed_size=args.num_hidden, prefix='gnmt_') model.initialize(init=mx.init.Uniform(0.1), ctx=ctx) model.hybridize() logging.info(model) translator = BeamSearchTranslator(model=model, beam_size=args.beam_size, scorer=BeamSearchScorer(alpha=args.lp_alpha, K=args.lp_k), max_length=args.tgt_max_len) logging.info('Use beam_size={}, alpha={}, K={}'.format(args.beam_size, args.lp_alpha, args.lp_k)) loss_function = SoftmaxCEMaskedLoss() loss_function.hybridize() def evaluate(data_loader): """Evaluate given the data loader Parameters ---------- data_loader : DataLoader Returns ------- avg_loss : float Average loss real_translation_out : list of list of str
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=nlp.model.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) detokenizer = nlp.data.SacreMosesDetokenizer() def evaluate(data_loader, context=ctx[0]): """Evaluate given the data loader Parameters ---------- data_loader : DataLoader
num_layers=args.num_layers, num_bi_layers=args.num_bi_layers) model = NMTModel(src_vocab=src_vocab, tgt_vocab=tgt_vocab, encoder=encoder, decoder=decoder, embed_size=args.num_hidden, prefix='gnmt_') model.initialize(init=mx.init.Uniform(0.1), ctx=ctx) model.hybridize() logging.info(model) translator = BeamSearchTranslator(model=model, beam_size=args.beam_size, scorer=BeamSearchScorer(alpha=args.lp_alpha, K=args.lp_k), max_length=args.tgt_max_len) logging.info('Use beam_size={}, alpha={}, K={}'.format(args.beam_size, args.lp_alpha, args.lp_k)) loss_function = SoftmaxCEMaskedLoss() loss_function.hybridize() def evaluate(data_loader): """Evaluate given the data loader Parameters ---------- data_loader : DataLoader Returns ------- avg_loss : float Average loss real_translation_out : list of list of str
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) detokenizer = NLTKMosesDetokenizer() def evaluate(data_loader, context=ctx[0]): """Evaluate given the data loader Parameters ---------- data_loader : DataLoader