Python utils.load_ensemble_for_inference方法代码示例

# 需要导入模块: from fairseq import utils [as 别名]
# 或者: from fairseq.utils import load_ensemble_for_inference [as 别名]
def __init__(self, path, data, use_cpu=True):
        # Create the language modeling task.
        self.args = FluencyArgs(path, data)
        self.task = tasks.setup_task(self.args)
        self.use_cuda = torch.cuda.is_available and not use_cpu

        # Load language model ensemble.
        models, model_args = utils.load_ensemble_for_inference(self.args.path.split(':'), self.task)
        self.models = models
        self.model_args = model_args

        # Optimize ensemble for generation.
        for model in self.models:
            model.make_generation_fast_()
            if self.use_cuda and self.model_args.fp16:
                model.half()
        
        # Create the sequence scorer.
        self.scorer = SequenceScorer(self.models, self.task.target_dictionary)
        if self.use_cuda:
            self.scorer.cuda() 

# 需要导入模块: from fairseq import utils [as 别名]
# 或者: from fairseq.utils import load_ensemble_for_inference [as 别名]
def setup_asr(args, logger):
    check_args(args)
    import_user_module(args)

    if args.max_tokens is None and args.max_sentences is None:
        args.max_tokens = 30000
    logger.info(args)

    use_cuda = torch.cuda.is_available() and not args.cpu

    # Load dataset splits
    task = tasks.setup_task(args)

    # Set dictionary
    tgt_dict = task.target_dictionary

    if args.ctc or args.rnnt:
        tgt_dict.add_symbol("<ctc_blank>")
        if args.ctc:
            logger.info("| decoding a ctc model")
        if args.rnnt:
            logger.info("| decoding a rnnt model")

    # Load ensemble
    logger.info("| loading model(s) from {}".format(args.path))
    models, _model_args = load_ensemble_for_inference(
        args.path.split(":"),
        task,
        model_arg_overrides=eval(args.model_overrides),  # noqa
    )
    optimize_models(args, use_cuda, models)

    # Initialize generator
    generator = task.build_generator(args)

    sp = spm.SentencePieceProcessor()
    sp.Load(os.path.join(args.data, "spm.model"))
    return task, generator, models, sp, tgt_dict 

# 需要导入模块: from fairseq import utils [as 别名]
# 或者: from fairseq.utils import load_ensemble_for_inference [as 别名]
def __init__(self, parsed_args):
        self.args = parsed_args
        import_user_module(parsed_args)
        assert parsed_args.path is not None, '--path required for evaluation'

        print(parsed_args)

        self.use_cuda = torch.cuda.is_available() and not parsed_args.cpu

        self.task = tasks.setup_task(parsed_args)

        # Load ensemble
        print('| loading model(s) from {}'.format(parsed_args.path))
        self.models, args = utils.load_ensemble_for_inference(
            parsed_args.path.split(':'), self.task, model_arg_overrides=eval(parsed_args.model_overrides),
        )

        for model in self.models:
            model.make_generation_fast_()
            if self.use_cuda:
                model.cuda()

        for arg in vars(parsed_args).keys():
            if arg not in {'self_target', 'future_target', 'past_target', 'tokens_per_sample',
                           'output_size_dictionary'}:
                setattr(args, arg, getattr(parsed_args, arg))
        self.task = tasks.setup_task(args)

        self.gen_timer = StopwatchMeter()
        self.scorer = SequenceScorer(self.task.target_dictionary) 

# 需要导入模块: from fairseq import utils [as 别名]
# 或者: from fairseq.utils import load_ensemble_for_inference [as 别名]
def main(args):
    assert args.path is not None, '--path required for evaluation!'

    if args.tokens_per_sample is None:
        args.tokens_per_sample = 1024
    print(args)

    use_cuda = torch.cuda.is_available() and not args.cpu

    # Load dataset splits
    task = tasks.setup_task(args)
    task.load_dataset(args.gen_subset)
    print('| {} {} {} examples'.format(args.data, args.gen_subset, len(task.dataset(args.gen_subset))))

    # Load ensemble
    print('| loading model(s) from {}'.format(args.path))
    models, _ = utils.load_ensemble_for_inference(args.path.split(':'), task)

    # Optimize ensemble for generation and set the source and dest dicts on the model (required by scorer)
    for model in models:
        model.make_generation_fast_()

    itr = data.EpochBatchIterator(
        dataset=task.dataset(args.gen_subset),
        max_sentences=args.max_sentences or 4,
        max_positions=model.max_positions(),
        num_shards=args.num_shards,
        shard_id=args.shard_id,
    ).next_epoch_itr(shuffle=False)

    gen_timer = StopwatchMeter()
    scorer = SequenceScorer(models, task.target_dictionary)
    if use_cuda:
        scorer.cuda()

    score_sum = 0.
    count = 0
    with progress_bar.build_progress_bar(args, itr) as t:
        results = scorer.score_batched_itr(t, cuda=use_cuda, timer=gen_timer)
        wps_meter = TimeMeter()
        for _, src_tokens, __, hypos in results:
            for hypo in hypos:
                pos_scores = hypo['positional_scores']
                inf_scores = pos_scores.eq(float('inf')) | pos_scores.eq(float('-inf'))
                if inf_scores.any():
                    print('| Skipping tokens with inf scores:',
                          task.target_dictionary.string(hypo['tokens'][inf_scores.nonzero()]))
                    pos_scores = pos_scores[(~inf_scores).nonzero()]
                score_sum += pos_scores.sum()
                count += pos_scores.numel()
            wps_meter.update(src_tokens.size(0))
            t.log({'wps': round(wps_meter.avg)})

    avg_nll_loss = -score_sum / count
    print('| Evaluated {} tokens in {:.1f}s ({:.2f} tokens/s)'.format(gen_timer.n, gen_timer.sum, 1. / gen_timer.avg))
    print('| Loss: {:.4f}, Perplexity: {:.2f}'.format(avg_nll_loss, np.exp(avg_nll_loss))) 

# 需要导入模块: from fairseq import utils [as 别名]
# 或者: from fairseq.utils import load_ensemble_for_inference [as 别名]
def build_model(cls, args, task):
        trained_encoder, trained_decoder = None, None
        pretrained = eval(args.pretrained)
        if pretrained:
            print("| Loading pretrained model")
            trained_model = utils.load_ensemble_for_inference(
                # not actually for inference, but loads pretrained model parameters
                filenames=[args.pretrained_checkpoint],
                task=task,
            )[0][0]
            trained_decoder = list(trained_model.children())[1]
            trained_encoder = list(trained_model.children())[0]

            # freeze pretrained model
            for param in trained_decoder.parameters():
                param.requires_grad = False
            for param in trained_encoder.parameters():
                param.requires_grad = False

        """Build a new model instance."""
        encoder = FConvEncoder(
            task.source_dictionary,
            embed_dim=args.encoder_embed_dim,
            convolutions=eval(args.encoder_layers),
            dropout=args.dropout,
            max_positions=args.max_source_positions,
            attention=eval(args.encoder_attention),
            attention_nheads=args.encoder_attention_nheads
        )

        decoder = FConvDecoder(
            task.target_dictionary,
            embed_dim=args.decoder_embed_dim,
            convolutions=eval(args.decoder_layers),
            out_embed_dim=args.decoder_out_embed_dim,
            attention=eval(args.decoder_attention),
            dropout=args.dropout,
            max_positions=args.max_target_positions,
            selfattention=eval(args.self_attention),
            attention_nheads=args.multihead_attention_nheads,
            selfattention_nheads=args.multihead_self_attention_nheads,
            project_input=eval(args.project_input),
            gated_attention=eval(args.gated_attention),
            downsample=eval(args.downsample),
            pretrained=pretrained,
            trained_decoder=trained_decoder
        )
        model = FConvModelSelfAtt(encoder, decoder, trained_encoder)

        return model 

# 需要导入模块: from fairseq import utils [as 别名]
# 或者: from fairseq.utils import load_ensemble_for_inference [as 别名]
def build_model(cls, args, task):
        trained_encoder, trained_decoder = None, None
        pretrained = eval(args.pretrained)
        if pretrained:
            print("| loading pretrained model")
            trained_model = utils.load_ensemble_for_inference(
                # not actually for inference, but loads pretrained model parameters
                filenames=[args.pretrained_checkpoint],
                task=task,
            )[0][0]
            trained_decoder = list(trained_model.children())[1]
            trained_encoder = list(trained_model.children())[0]

            # freeze pretrained model
            for param in trained_decoder.parameters():
                param.requires_grad = False
            for param in trained_encoder.parameters():
                param.requires_grad = False

        """Build a new model instance."""
        encoder = FConvEncoder(
            task.source_dictionary,
            embed_dim=args.encoder_embed_dim,
            convolutions=eval(args.encoder_layers),
            dropout=args.dropout,
            max_positions=args.max_source_positions,
            attention=eval(args.encoder_attention),
            attention_nheads=args.encoder_attention_nheads
        )

        decoder = FConvDecoder(
            task.target_dictionary,
            embed_dim=args.decoder_embed_dim,
            convolutions=eval(args.decoder_layers),
            out_embed_dim=args.decoder_out_embed_dim,
            attention=eval(args.decoder_attention),
            dropout=args.dropout,
            max_positions=args.max_target_positions,
            selfattention=eval(args.self_attention),
            attention_nheads=args.multihead_attention_nheads,
            selfattention_nheads=args.multihead_self_attention_nheads,
            project_input=eval(args.project_input),
            gated_attention=eval(args.gated_attention),
            downsample=eval(args.downsample),
            pretrained=pretrained,
            trained_decoder=trained_decoder
        )
        model = FConvModelSelfAtt(encoder, decoder, trained_encoder)

        return model 

# 需要导入模块: from fairseq import utils [as 别名]
# 或者: from fairseq.utils import load_ensemble_for_inference [as 别名]
def main(args):
    print(args)

    use_cuda = torch.cuda.is_available() and not args.cpu

    # Load ensemble
    print('| loading model(s) from {}'.format(', '.join(args.path)))
    models, model_args = utils.load_ensemble_for_inference(args.path, data_dir=args.data)
    src_dict, dst_dict = models[0].src_dict, models[0].dst_dict

    print('| [{}] dictionary: {} types'.format(model_args.source_lang, len(src_dict)))
    print('| [{}] dictionary: {} types'.format(model_args.target_lang, len(dst_dict)))

    # Optimize ensemble for generation
    for model in models:
        model.make_generation_fast_(
            beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
        )

    # Initialize generator
    translator = SequenceGenerator(
        models, beam_size=args.beam, stop_early=(not args.no_early_stop),
        normalize_scores=(not args.unnormalized), len_penalty=args.lenpen,
        unk_penalty=args.unkpen)
    if use_cuda:
        translator.cuda()

    # Load alignment dictionary for unknown word replacement
    # (None if no unknown word replacement, empty if no path to align dictionary)
    align_dict = utils.load_align_dict(args.replace_unk)

    print('| Type the input sentence and press return:')
    for src_str in sys.stdin:
        src_str = src_str.strip()
        src_tokens = tokenizer.Tokenizer.tokenize(src_str, src_dict, add_if_not_exist=False).long()
        if use_cuda:
            src_tokens = src_tokens.cuda()
        src_lengths = src_tokens.new([src_tokens.numel()])
        translations = translator.generate(
            Variable(src_tokens.view(1, -1)),
            Variable(src_lengths.view(-1)),
        )
        hypos = translations[0]
        print('O\t{}'.format(src_str))

        # Process top predictions
        for hypo in hypos[:min(len(hypos), args.nbest)]:
            hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
                hypo_tokens=hypo['tokens'].int().cpu(),
                src_str=src_str,
                alignment=hypo['alignment'].int().cpu(),
                align_dict=align_dict,
                dst_dict=dst_dict,
                remove_bpe=args.remove_bpe,
            )
            print('H\t{}\t{}'.format(hypo['score'], hypo_str))
            print('A\t{}'.format(' '.join(map(str, alignment))))