prediction.py

# coding: utf-8
"""
This module holds methods for generating predictions from a model.
"""
import math
import sys
import time
from itertools import zip_longest
from pathlib import Path
from typing import Dict, List, Optional, Tuple

import numpy as np
import torch
from torch.nn import DataParallel as DP
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import Dataset
from tqdm import tqdm

from joeynmt.config import (
    BaseConfig,
    TestConfig,
    parse_global_args,
    set_validation_args,
)
from joeynmt.data import load_data
from joeynmt.datasets import StreamDataset
from joeynmt.helpers import (
    expand_reverse_index,
    load_checkpoint,
    resolve_ckpt_path,
    save_hypothese,
    set_seed,
    store_attention_plots,
    write_list_to_file,
)
from joeynmt.helpers_for_ddp import (
    ddp_merge,
    ddp_reduce,
    ddp_synchronize,
    get_logger,
    use_ddp,
)
from joeynmt.metrics import bleu, chrf, sequence_accuracy, token_accuracy
from joeynmt.model import DataParallelWrapper, Model, build_model
from joeynmt.search import search

logger = get_logger(__name__)


def predict(
    model: Model,
    data: Dataset,
    device: torch.device,
    n_gpu: int,
    rank: int = 0,
    compute_loss: bool = False,
    normalization: str = "batch",
    num_workers: int = 0,
    args: TestConfig = None,
    autocast: Dict = None,
) -> Tuple[Dict[str, float], Optional[List[str]], Optional[List[str]], List[List[str]],
           List[np.ndarray], List[np.ndarray]]:
    """
    Generates translations for the given data.
    If `compute_loss` is True and references are given, also computes the loss.

    :param model: model module
    :param data: dataset for validation
    :param device: torch device
    :param n_gpu: number of GPUs
    :param rank: ddp rank
    :param compute_loss: whether to compute a scalar loss for given inputs and targets
    :param normalization: one of {`batch`, `tokens`, `none`}
    :param num_workers: number of workers for `collate_fn()` in data iterator
    :param args: configuration args
    :param autocast: autocast context
    :return:
        - valid_scores: (dict) current validation scores,
        - valid_ref: (list of str) post-processed validation references,
        - valid_hyp: (list of str) post-processed validation hypotheses,
        - decoded_valid: (list of list of str) token-level validation hypotheses,
        - valid_seq_scores: (list of np.array) log probabilities (hyp or ref)
        - valid_attn_scores: (list of np.array) attention scores (hyp or ref)
    """
    # pylint: disable=too-many-branches,too-many-statements

    if use_ddp():
        if args.batch_type == "token":
            logger.warning(
                "Token-based batch sampling is not supported in distributed "
                "learning. fall back to sentence-based batch sampling."
            )
            args = args._replace(batch_type="sentence", batch_size=64)
        if args.beam_size > 1:
            logger.warning(
                "Beam search is not supported in distributed learning. "
                "fall back to greedy decoding."
            )
            args = set_validation_args(args)
    else:
        # DataParallel distributes batch sequences over devices
        assert args.batch_size >= n_gpu, "`batch_size` must be bigger than `n_gpu`."
        # **CAUTION:** a batch will be expanded to batch.nseqs * beam_size, and it might
        # cause an out-of-memory error.
        # if batch_size > beam_size:
        #     batch_size //= beam_size

    valid_iter = data.make_iter(
        batch_size=args.batch_size,
        batch_type=args.batch_type,
        shuffle=False,
        seed=data.seed,  # for subsampling (not for shuffling)
        num_workers=num_workers,
        eos_index=model.eos_index,
        pad_index=model.pad_index,
        device=device,
    )
    num_samples = valid_iter.batch_sampler.num_samples

    # log decoding configs
    if args.return_prob == "ref":  # no decoding needed
        decoding_description = ""
    else:
        decoding_description = (  # write the decoding strategy in the log
            " (Greedy decoding with " if args.beam_size < 2 else
            f" (Beam search with beam_size={args.beam_size}, "
            f"beam_alpha={args.beam_alpha}, n_best={args.n_best}, ")
        decoding_description += (
            f"min_output_length={args.min_output_length}, "
            f"max_output_length={args.max_output_length}, "
            f"return_prob='{args.return_prob}', generate_unk={args.generate_unk}, "
            f"repetition_penalty={args.repetition_penalty}, "
            f"no_repeat_ngram_size={args.no_repeat_ngram_size})"
        )
    logger.info("Predicting %d example(s)...%s", num_samples, decoding_description)

    # disable dropout
    model.eval()

    # placeholders for scores
    # Note: access these variables only when rank == 0
    valid_scores = {"loss": float("nan"), "acc": float("nan"), "ppl": float("nan")}
    all_outputs, all_indices, valid_attn_scores, valid_seq_scores = [], [], [], []
    total_loss, total_nseqs, total_ntokens, total_n_correct = 0, 0, 0, 0
    output, ref_scores, hyp_scores, attention_scores = None, None, None, None

    ddp_synchronize()  # ensure that all processes are ready

    gen_start_time = time.time()

    disable_tqdm = isinstance(data, StreamDataset) or rank != 0
    with tqdm(total=num_samples, disable=disable_tqdm, desc="Predicting...") as pbar:
        for batch in valid_iter:
            # number of sentences in the current batch
            # = batch.nseqs * n_gpu if use_ddp()
            batch_nseqs = ddp_reduce(batch.nseqs, device, torch.long).item()

            # sort batch now by src length and keep track of order
            reverse_index = batch.sort_by_src_length()
            sort_reverse_index = expand_reverse_index(reverse_index, args.n_best)
            batch_size = len(sort_reverse_index)  # = batch.nseqs * args.n_best

            # run as during training to get validation loss (e.g. xent)
            if compute_loss and batch.has_trg:
                assert model.loss_function is not None

                # don't track gradients during validation
                with torch.autocast(**autocast):
                    with torch.no_grad():
                        batch_loss, log_probs, attn, n_correct = model(
                            return_type="loss",
                            return_prob=args.return_prob,
                            return_attention=args.return_attention,
                            **vars(batch)
                        )

                # gather
                batch_loss = ddp_reduce(batch_loss)
                n_correct = ddp_reduce(n_correct)
                batch_ntokens = ddp_reduce(batch.ntokens, device, torch.long)
                # sum over multiple gpus
                batch_loss = batch.normalize(batch_loss, "sum", n_gpu=n_gpu)
                n_correct = batch.normalize(n_correct, "sum", n_gpu=n_gpu)

                if args.return_prob == "ref":
                    # gather
                    log_probs = ddp_merge(log_probs, 0.0)
                    attn = ddp_merge(attn, 0.0)
                    batch_trg = ddp_merge(batch.trg, model.pad_index)

                    ref_scores = batch.score(log_probs, batch_trg, model.pad_index)
                    attention_scores = attn.detach().cpu().numpy()
                    output = batch_trg.detach().cpu().numpy()

                if rank == 0:
                    total_loss += batch_loss.item()
                    total_n_correct += n_correct.item()
                    total_ntokens += batch_ntokens.item()

            # if return_prob == "ref", then no search needed.
            # (just look up the prob of the ground truth.)
            if args.return_prob != "ref":
                # run search as during inference to produce translations
                output, hyp_scores, attention_scores = search(
                    model=model,
                    batch=batch,
                    beam_size=args.beam_size,
                    beam_alpha=args.beam_alpha,
                    max_output_length=args.max_output_length,
                    n_best=args.n_best,
                    return_attention=args.return_attention,
                    return_prob=args.return_prob,
                    generate_unk=args.generate_unk,
                    repetition_penalty=args.repetition_penalty,
                    no_repeat_ngram_size=args.no_repeat_ngram_size,
                    autocast=autocast,
                )

            if use_ddp():
                # we don't know the order of merged outputs.
                # sort them back by indices after the batch loop has ended.
                batch_indices = ddp_merge(batch.indices.unsqueeze(1), -1).squeeze()
                assert torch.all(batch_indices >= 0).item(), batch_indices
                assert len(batch_indices) == len(output) == batch_nseqs
                if rank == 0:
                    all_outputs.extend(output)
                    all_indices.extend(batch_indices.detach().cpu().numpy())
            else:
                # sort outputs back to original order
                all_outputs.extend(output[sort_reverse_index])  # either hyp or ref
                valid_attn_scores.extend(
                    attention_scores[sort_reverse_index]
                    if attention_scores is not None else []
                )
                valid_seq_scores.extend(
                    ref_scores[sort_reverse_index] \
                    if ref_scores is not None and ref_scores.shape[0] == batch_size
                    else hyp_scores[sort_reverse_index] \
                    if hyp_scores is not None and hyp_scores.shape[0] == batch_size
                    else [])

            if rank == 0:
                total_nseqs += batch_nseqs
                pbar.update(batch_nseqs)

    gen_duration = time.time() - gen_start_time
    logger.info("Generation took %.4f[sec].", gen_duration)

    if rank == 0:
        if use_ddp():
            num_samples = valid_iter.batch_sampler.num_samples
            # sort back to the original order according to `all_indices`
            _all_outputs = [None] * len(data)  # not len(all_outputs)!
            for i, row in zip(all_indices, all_outputs):
                _all_outputs[i] = row
            all_outputs = [out for out in _all_outputs if out is not None]

        assert total_nseqs == num_samples, (total_nseqs, num_samples)
        assert len(all_outputs
                   ) == num_samples * args.n_best, (len(all_outputs), num_samples)

        if compute_loss:
            if normalization == "batch":
                normalizer = total_nseqs
            elif normalization == "tokens":
                normalizer = total_ntokens
            elif normalization == "none":
                normalizer = 1

            # avoid zero division
            assert normalizer > 0, normalizer
            assert total_ntokens > 0, total_ntokens

            # normalized loss
            valid_scores["loss"] = total_loss / normalizer
            # accuracy before decoding
            valid_scores["acc"] = total_n_correct / total_ntokens
            # exponent of token-level negative log likelihood
            valid_scores["ppl"] = math.exp(total_loss / total_ntokens)

        # decode ids back to str symbols (cut-off AFTER eos; eos itself is included.)
        decoded_valid = model.trg_vocab.arrays_to_sentences(
            arrays=all_outputs, cut_at_eos=True
        )
        # TODO: `valid_seq_scores` should have the same seq length as `decoded_valid`
        #     -> needed to be cut-off at eos/sep synchronously

        if args.return_prob == "ref":  # no evaluation needed
            logger.info(
                "Evaluation result (scoring) %s.", ", ".join([
                    f"{eval_metric}: {valid_scores[eval_metric]:6.2f}"
                    for eval_metric in ["loss", "ppl", "acc"]
                ])
            )
            return (
                valid_scores, None, None, decoded_valid, valid_seq_scores,
                valid_attn_scores
            )

        # retrieve detokenized hypotheses
        valid_hyp = []
        for i, sentence in enumerate(decoded_valid):
            try:
                sentence = data.tokenizer[data.trg_lang].post_process(
                    sentence, generate_unk=args.generate_unk, cut_at_sep=True
                )
            except AssertionError as e:
                logger.error("empty hypothesis at %d: %r (%r)", i, sentence, e)
                # pylint: disable=protected-access
                sentence = model.trg_vocab._itos[model.unk_index]
            valid_hyp.append(sentence)
        assert len(valid_hyp) == len(all_outputs), (len(valid_hyp), len(all_outputs))

        # if references are given, compute evaluation metrics
        if data.has_trg:
            valid_scores, valid_ref = evaluate(valid_scores, valid_hyp, data, args)
        else:
            valid_ref = None  # no references

    return (
        valid_scores,
        valid_ref,
        valid_hyp,
        decoded_valid,
        valid_seq_scores,
        valid_attn_scores,
    ) if rank == 0 else None


def evaluate(valid_scores: Dict, valid_hyp: List, data: Dataset,
             args: TestConfig) -> Tuple[Dict[str, float], List[str]]:
    """
    Compute evaluateion metrics

    :param valid_scores: scores dict
    :param valid_hyp: decoded hypotheses
    :param data: eval Dataset
    :param args: configuration args
    :return:
        - valid_scores: evaluation scores
        - valid_ref: postprocessed references
    """

    # references are not length-filtered, not duplicated for n_best > 1
    valid_ref = [data.tokenizer[data.trg_lang].post_process(t) for t in data.trg]

    # metrics are computed on the 1-best hypotheses
    valid_hyp_1best = [valid_hyp[i] for i in range(0, len(valid_hyp), args.n_best)] \
        if args.n_best > 1 else valid_hyp
    assert len(valid_hyp_1best) == len(valid_ref), (valid_hyp_1best, valid_ref)

    eval_start_time = time.time()

    # evaluate with metrics on dev dataset
    for eval_metric in args.eval_metrics:
        if eval_metric == "bleu":
            valid_scores[eval_metric] = bleu(
                valid_hyp_1best, valid_ref, **args.sacrebleu_cfg
            )
        elif eval_metric == "chrf":
            valid_scores[eval_metric] = chrf(
                valid_hyp_1best, valid_ref, **args.sacrebleu_cfg
            )
        elif eval_metric == "token_accuracy":
            valid_scores[eval_metric] = token_accuracy(
                valid_hyp_1best, valid_ref, data.tokenizer[data.trg_lang]
            )
        elif eval_metric == "sequence_accuracy":
            valid_scores[eval_metric] = sequence_accuracy(valid_hyp_1best, valid_ref)

    eval_duration = time.time() - eval_start_time

    score_str = ", ".join([
        f"{eval_metric}: {valid_scores[eval_metric]:6.2f}"
        for eval_metric in args.eval_metrics + ["loss", "ppl", "acc"]
        if not math.isnan(valid_scores[eval_metric])
    ])
    logger.info(
        "Evaluation result (%s): %s, %.4f[sec]",
        "beam search" if args.beam_size > 1 else "greedy",
        score_str,
        eval_duration,
    )

    return valid_scores, valid_ref


def prepare(args: BaseConfig, rank: int,
            mode: str) -> Tuple[Model, Dataset, Dataset, Dataset]:
    """
    Helper function for model and data loading.

    :param args: config args
    :param rank: ddp rank
    :param mode: execution mode
    """
    # load the data
    if mode == "train":
        datasets = ["train", "dev", "test"]
    if mode == "test":
        datasets = ["dev", "test"]
    if mode == "translate":
        datasets = ["stream"]

    if mode != "train":
        if "voc_file" not in args.data["src"] or not args.data["src"]["voc_file"]:
            args.data["src"]["voc_file"] = (args.model_dir / "src_vocab.txt").as_posix()
        if "voc_file" not in args.data["trg"] or not args.data["trg"]["voc_file"]:
            args.data["trg"]["voc_file"] = (args.model_dir / "trg_vocab.txt").as_posix()

    src_vocab, trg_vocab, train_data, dev_data, test_data = load_data(
        cfg=args.data, datasets=datasets
    )

    if mode == "train" and rank == 0:
        # store the vocabs and tokenizers
        src_vocab.to_file(args.model_dir / "src_vocab.txt")
        if hasattr(train_data.tokenizer[train_data.src_lang], "copy_cfg_file"):
            train_data.tokenizer[train_data.src_lang].copy_cfg_file(args.model_dir)
        trg_vocab.to_file(args.model_dir / "trg_vocab.txt")
        if hasattr(train_data.tokenizer[train_data.trg_lang], "copy_cfg_file"):
            train_data.tokenizer[train_data.trg_lang].copy_cfg_file(args.model_dir)

    # build an encoder-decoder model
    model = build_model(args.model, src_vocab=src_vocab, trg_vocab=trg_vocab)
    model.log_parameters_list()
    # need to instantiate loss func after `build_model()`
    model.loss_function = (args.train.loss, args.train.label_smoothing)

    if mode != "train":
        # infer ckpt path for testing
        ckpt = resolve_ckpt_path(args.test.load_model, args.model_dir)

        # load model checkpoint
        logger.info("Loading model from %s", ckpt)
        map_location = {"cuda:0": f"cuda:{rank}"} if use_ddp() else args.device
        model_checkpoint = load_checkpoint(ckpt, map_location=map_location)

        # restore model and optimizer parameters
        model.load_state_dict(model_checkpoint["model_state"])

    # move to gpu
    if args.device.type == "cuda":
        model.to(args.device)

    # multi-gpu training
    if args.n_gpu > 1:
        if use_ddp():
            model = DataParallelWrapper(
                DDP(model, device_ids=[rank], output_device=rank)
            )
        else:
            model = DataParallelWrapper(DP(model))
    logger.info(model)

    # set the random seed
    set_seed(seed=args.seed)

    return model, train_data, dev_data, test_data


def test(
    cfg: Dict,
    output_path: str = None,
    prepared: Dict = None,
    save_attention: bool = False,
    save_scores: bool = False,
) -> None:
    """
    Main test function. Handles loading a model from checkpoint, generating
    translations, storing them, and plotting attention.

    :param cfg: configuration dict
    :param output_path: path to output
    :param prepared: model and datasets passed from training
    :param save_attention: whether to save attention visualizations
    :param save_scores: whether to save scores
    """
    # parse args
    args = parse_global_args(cfg, rank=0, mode="test")

    # load the model and data
    if prepared is None:
        model, _, dev_data, test_data = prepare(args, rank=0, mode="test")
        data_to_predict = {"dev": dev_data, "test": test_data}

    else:  # avoid to load model and data again
        model = prepared["model"]
        data_to_predict = {"dev": prepared["dev"], "test": prepared["test"]}

    # check options
    if save_attention:
        if cfg["model"]["decoder"]["type"] == "transformer":
            assert cfg["testing"].get("beam_size", 1) == 1, (
                "Attention plots can be saved with greedy decoding only. Please set "
                "`beam_size: 1` in the config."
            )
        args = args._replace(test=args.test._replace(return_attention=True))
    if save_scores:
        assert output_path, "Please specify --output_path for saving scores."
        if args.test.return_prob == "none":
            logger.warning(
                "Please specify prob type: {`ref` or `hyp`} in the config. "
                "Scores will not be saved."
            )
            save_scores = False
        elif args.test.return_prob == "ref":
            assert cfg["testing"].get("beam_size", 1) == 1, (
                "Scores of given references can be computed with greedy decoding only. "
                "Please set `beam_size: 1` in the config."
            )

    # prediction loop over datasets
    for data_set_name, data_set in data_to_predict.items():
        if data_set is not None:
            data_set.reset_indices(random_subset=-1)  # no subsampling in evaluation

            logger.info(
                "%s on %s set... (device: %s, n_gpu: %s, use_ddp: %r, fp16: %r)",
                "Scoring" if args.test.return_prob == "ref" else "Decoding",
                data_set_name, args.device.type, args.n_gpu, use_ddp(),
                args.autocast["enabled"]
            )
            _, _, hypotheses, hypotheses_raw, seq_scores, att_scores, = predict(
                model=model,
                data=data_set,
                compute_loss=args.test.return_prob == "ref",
                device=args.device,
                rank=0,
                n_gpu=args.n_gpu,
                num_workers=args.num_workers,
                normalization=args.train.normalization,
                args=args.test,
                autocast=args.autocast,
            )

            if save_attention:
                if att_scores:
                    attention_file_name = f"{output_path}.{data_set_name}.att"
                    logger.info("Saving attention plots. This might take a while..")
                    store_attention_plots(
                        attentions=att_scores,
                        targets=hypotheses_raw,
                        sources=data_set.get_list(
                            lang=data_set.src_lang, tokenized=True
                        ),
                        indices=range(len(hypotheses)),
                        output_prefix=attention_file_name,
                    )
                    logger.info("Attention plots saved to: %s", attention_file_name)
                else:
                    logger.warning(
                        "Attention scores could not be saved. Note that attention "
                        "scores are not available when using beam search. "
                        "Set beam_size to 1 for greedy decoding."
                    )

            if output_path is not None:
                if save_scores and seq_scores is not None:
                    # save scores
                    output_path_scores = Path(f"{output_path}.{data_set_name}.scores")
                    write_list_to_file(output_path_scores, seq_scores)
                    # save tokens
                    output_path_tokens = Path(f"{output_path}.{data_set_name}.tokens")
                    write_list_to_file(output_path_tokens, hypotheses_raw)
                    logger.info(
                        "Scores and corresponding tokens saved to: %s.{scores|tokens}",
                        f"{output_path}.{data_set_name}",
                    )
                if hypotheses is not None:
                    # save translations
                    output_path_set = Path(f"{output_path}.{data_set_name}")
                    save_hypothese(output_path_set, hypotheses, args.test.n_best)
                    logger.info("Translations saved to: %s.", output_path_set)


def translate(cfg: Dict, output_path: str = None) -> None:
    """
    Interactive translation function.
    Loads model from checkpoint and translates either the stdin input or asks for
    input to translate interactively. Translations and scores are printed to stdout.
    Note: The input sentences don't have to be pre-tokenized.

    :param cfg: configuration dict
    :param output_path: path to output file
    """

    # parse args
    args = parse_global_args(cfg, rank=0, mode="translate")

    # load the model
    model, _, _, test_data = prepare(args, rank=0, mode="translate")
    assert isinstance(test_data, StreamDataset)

    logger.info(
        "Ready to decode. (device: %s, n_gpu: %s, use_ddp: %r, fp16: %r)",
        args.device.type, args.n_gpu, use_ddp(), args.autocast["enabled"]
    )

    def _translate_data(test_data: Dataset, args: BaseConfig):
        """Translates given dataset, using parameters from outer scope."""
        _, _, hypotheses, trg_tokens, trg_scores, _ = predict(
            model=model,
            data=test_data,
            compute_loss=False,
            device=args.device,
            rank=0,
            n_gpu=args.n_gpu,
            normalization="none",
            num_workers=args.num_workers,
            args=args.test,
            autocast=args.autocast,
        )
        return hypotheses, trg_tokens, trg_scores

    if not sys.stdin.isatty():  # pylint: disable=too-many-nested-blocks
        # input stream given
        for i, line in enumerate(sys.stdin.readlines()):
            if not line.strip():
                # skip empty lines and print warning
                logger.warning("The sentence in line %d is empty. Skip to load.", i)
                continue
            test_data.set_item(line.rstrip())
        all_hypotheses, tokens, scores = _translate_data(test_data, args)
        assert len(all_hypotheses) == len(test_data) * args.test.n_best

        if output_path is not None:
            # write to outputfile if given
            out_file = Path(output_path).expanduser()
            save_hypothese(out_file, all_hypotheses, args.n_best)
            logger.info("Translations saved to: %s.", out_file)

        else:
            # print to stdout
            for hyp in all_hypotheses:
                print(hyp)

    else:
        # CAUTION: this will raise an error if n_gpus > 1
        args = args._replace(
            test=args.test._replace(batch_size=1, batch_type="sentence")
        )
        # enter interactive mode
        np.set_printoptions(linewidth=sys.maxsize)  # for printing scores in stdout
        while True:
            try:
                src_input = input("\nPlease enter a source sentence:\n")
                if not src_input.strip():
                    break

                # every line has to be made into dataset
                test_data.set_item(src_input.rstrip())
                hypotheses, tokens, scores = _translate_data(test_data, args)

                print("JoeyNMT:")
                for i, (hyp, token,
                        score) in enumerate(zip_longest(hypotheses, tokens, scores)):
                    assert hyp is not None, (i, hyp, token, score)
                    print(f"#{i + 1}: {hyp}")
                    if args.test.return_prob in ["hyp"]:
                        if args.test.beam_size > 1:  # beam search: seq-level scores
                            print(f"\ttokens: {token}\n\tsequence score: {score[0]}")
                        else:  # greedy: token-level scores
                            assert len(token) == len(score), (token, score)
                            print(f"\ttokens: {token}\n\tscores: {score}")

                # reset cache
                test_data.reset_cache()

            except (KeyboardInterrupt, EOFError):
                print("\nBye.")
                break