Python tasks.setup_task方法代碼示例

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
def load_models_and_criterions(filenames, arg_overrides=None, task=None):
    models = []
    criterions = []
    for filename in filenames:
        if not os.path.exists(filename):
            raise IOError("Model file not found: {}".format(filename))
        state = checkpoint_utils.load_checkpoint_to_cpu(filename, arg_overrides)

        args = state["args"]
        if task is None:
            task = tasks.setup_task(args)

        # build model for ensemble
        model = task.build_model(args)
        model.load_state_dict(state["model"], strict=True)
        models.append(model)

        criterion = task.build_criterion(args)
        if "criterion" in state:
            criterion.load_state_dict(state["criterion"], strict=True)
        criterions.append(criterion)
    return models, criterions, args 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
def load_model(self, args):
        args.user_dir = os.path.join(os.path.dirname(__file__), '..', '..')
        utils.import_user_module(args)
        filename = args.model_path
        if not os.path.exists(filename):
            raise IOError("Model file not found: {}".format(filename))

        state = checkpoint_utils.load_checkpoint_to_cpu(filename, json.loads(args.model_overrides))

        saved_args = state["args"]
        saved_args.data = args.data_bin

        task = tasks.setup_task(saved_args)

        # build model for ensemble
        self.model = task.build_model(saved_args)
        self.model.load_state_dict(state["model"], strict=True)

        # Set dictionary
        self.load_dictionary(task) 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
def load_model_ensemble_and_task(filenames, arg_overrides=None, task=None, strict=True, suffix=''):
    from fairseq import tasks

    ensemble = []
    for filename in filenames:
        filename = filename.replace(".pt", suffix + ".pt")
        if not PathManager.exists(filename):
            raise IOError("Model file not found: {}".format(filename))
        state = load_checkpoint_to_cpu(filename, arg_overrides)

        args = state["args"]
        if task is None:
            task = tasks.setup_task(args)

        # build model for ensemble
        model = task.build_model(args)
        model.load_state_dict(state["model"], strict=strict, args=args)
        ensemble.append(model)
    return ensemble, args, task 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
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() 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
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 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
def create_task_and_model(args):
    task = tasks.setup_task(args)
    model = task.build_model(args)
    return task, model 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
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) 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
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))) 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
def load_diverse_ensemble_for_inference(
    filenames: List[str], task: Optional[tasks.FairseqTask] = None
):
    """Load an ensemble of diverse models for inference.

    This method is similar to fairseq.utils.load_ensemble_for_inference
    but allows to load diverse models with non-uniform args.

    Args:
        filenames: List of file names to checkpoints
        task: Optional[FairseqTask]. If this isn't provided, we setup the task
            using the first checkpoint's model args loaded from the saved state.

    Return:
        models, args: Tuple of lists. models contains the loaded models, args
            the corresponding configurations.
        task: Either the input task or the task created within this function
            using args
    """

    # load model architectures and weights
    checkpoints_data = []
    for filename in filenames:
        if not PathManager.exists(filename):
            raise IOError("Model file not found: {}".format(filename))
        with PathManager.open(filename, "rb") as f:
            checkpoints_data.append(
                torch.load(
                    f,
                    map_location=lambda s, l: torch.serialization.default_restore_location(
                        s, "cpu"
                    ),
                )
            )
    # build ensemble
    ensemble = []
    if task is None:
        if hasattr(checkpoints_data[0]["args"], "mode"):
            checkpoints_data[0]["args"].mode = "eval"
        task = tasks.setup_task(checkpoints_data[0]["args"])
    for checkpoint_data in checkpoints_data:
        model = task.build_model(checkpoint_data["args"])
        model.load_state_dict(checkpoint_data["model"])
        ensemble.append(model)
    args_list = [s["args"] for s in checkpoints_data]
    return ensemble, args_list, task 

# 需要導入模塊: from fairseq import tasks [as 別名]
# 或者: from fairseq.tasks import setup_task [as 別名]
def __init__(self, model_path, user_dir, lang_pair, n_cpu_threads=-1):
        """Initializes a fairseq predictor.

        Args:
            model_path (string): Path to the fairseq model (*.pt). Like
                                 --path in fairseq-interactive.
            lang_pair (string): Language pair string (e.g. 'en-fr').
            user_dir (string): Path to fairseq user directory.
            n_cpu_threads (int): Number of CPU threads. If negative,
                                 use GPU.
        """
        super(FairseqPredictor, self).__init__()
        _initialize_fairseq(user_dir)
        self.use_cuda = torch.cuda.is_available() and n_cpu_threads < 0

        parser = options.get_generation_parser()
        input_args = ["--path", model_path, os.path.dirname(model_path)]
        if lang_pair:
            src, trg = lang_pair.split("-")
            input_args.extend(["--source-lang", src, "--target-lang", trg])
        args = options.parse_args_and_arch(parser, input_args)

        # Setup task, e.g., translation
        task = tasks.setup_task(args)
        self.src_vocab_size = len(task.source_dictionary)
        self.trg_vocab_size = len(task.target_dictionary)
        self.pad_id = task.source_dictionary.pad()

        # Load ensemble
        logging.info('Loading fairseq model(s) from {}'.format(model_path))
        self.models, _ = checkpoint_utils.load_model_ensemble(
            model_path.split(':'),
            task=task,
        )

        # Optimize ensemble for generation
        for model in self.models:
            model.make_generation_fast_(
                beamable_mm_beam_size=1,
                need_attn=False,
            )
            if self.use_cuda:
                model.cuda()
        self.model = EnsembleModel(self.models)
        self.model.eval()