eval_neural_rescorer.py

# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""
This script would evaluate a neural language model (Transformer) trained with
`examples/nlp/language_modeling/transformer_lm.py' as a rescorer for ASR systems.
Given a trained TransformerLMModel `.nemo` file, this script can be used to re-score the beams obtained from a beam
search decoder of an ASR model.

USAGE:
1. Obtain `.tsv` file with beams and their corresponding scores. Scores can be from a regular beam search decoder or
   in fusion with an N-gram LM scores. For a given beam size `beam_size` and a number of examples
   for evaluation `num_eval_examples`, it should contain (`beam_size` x `num_eval_examples`) lines of
   form `beam_candidate_text \t score`. This file can be generated by `scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py`.

2. Rescore the candidates:
    python eval_neural_rescorer.py
        --lm_model=[path to .nemo file of the LM]
        --beams_file=[path to beams .tsv file]
        --beam_size=[size of the beams]
        --eval_manifest=[path to eval manifest .json file]
        --batch_size=[batch size used for inference on the LM model]
        --alpha=[the value for the parameter rescorer_alpha]
        --beta=[the value for the parameter rescorer_beta]

You may find more info on how to use this script at:
https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html

"""

import contextlib
import inspect
import json
from argparse import ArgumentParser

import editdistance
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import torch
import tqdm
from transformers import AutoModelForCausalLM

from nemo.collections.nlp.models.language_modeling import TransformerLMModel
from nemo.collections.nlp.modules.common.tokenizer_utils import get_tokenizer
from nemo.utils import logging


class BeamScoresDataset(torch.utils.data.Dataset):
    """
    Dataset to read the score file containing the beams and their score

    Args:
        data_path: path to the beams file
        tokenizer: tokenizer of the LM model
        manifest_path: manifest `.json` file which contains the ground truths transcripts
        beam_size: the number of beams per sample
        max_seq_length: the maximum length of sequences
    """

    def __init__(self, data_path, tokenizer, manifest_path, beam_size=128, max_seq_length=256):
        self.data = pd.read_csv(data_path, delimiter="\t", header=None)
        self.tokenizer = tokenizer
        self.ground_truths = []
        with open(manifest_path, 'r', encoding='utf-8') as f_orig:
            for line in f_orig:
                item = json.loads(line)
                self.ground_truths.append(item['text'])
        self.beam_size = beam_size
        self.max_seq_length = max_seq_length

        if self.tokenizer.pad_id is not None:
            self.pad_id = self.tokenizer.pad_id
        elif self.tokenizer.eos_id is not None:
            self.pad_id = self.tokenizer.eos_id
        else:
            logging.warning(f"Using 0 as pad_id as the tokenizer has no pad_id or eos_id.")
            self.pad_id = 0

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        text = str(self.data[0][idx])
        tokens = self.tokenizer.text_to_ids(text)
        if self.tokenizer.bos_id is not None:
            tokens = [self.tokenizer.bos_id] + tokens
        if self.tokenizer.eos_id is not None:
            tokens = tokens + [self.tokenizer.eos_id]
        input_ids = [self.pad_id] * self.max_seq_length
        input_ids[: len(tokens)] = tokens
        input_ids = np.array(input_ids)
        input_mask = np.zeros(self.max_seq_length)
        input_mask[: len(tokens)] = 1
        acoustic_score = self.data[1][idx]
        dist = editdistance.eval(text.split(), self.ground_truths[idx // self.beam_size].split())
        ref_len = len(self.ground_truths[idx // self.beam_size].split())
        len_in_chars = len(str(self.data[0][idx]))
        return input_ids, input_mask, acoustic_score, dist, ref_len, len_in_chars, idx


def linear_search_wer(
    dists, scores1, scores2, total_len, coef_range=[0, 10], coef_steps=10000, param_name='parameter'
):
    """
    performs linear search to find the best coefficient when two set of scores are getting linearly fused.

    Args:
        dists: Tesnor of the distances between the ground truth and the candidates with shape of [number of samples, beam size]
        scores1: Tensor of the first set of scores with shape of [number of samples, beam size]
        scores2: Tensor of the second set of scores with shape of [number of samples, beam size]
        total_len: The total length of all samples
        coef_range: the search range for the coefficient
        coef_steps: the number of steps that the search range would get divided into
        param_name: the name of the parameter to be used in the figure

    Output:
        (best coefficient found, best WER achieved)
    """
    scale = scores1.mean().abs().item() / scores2.mean().abs().item()
    left = coef_range[0] * scale
    right = coef_range[1] * scale
    coefs = np.linspace(left, right, coef_steps)

    best_wer = 10000
    best_coef = left
    wers = []
    for coef in coefs:
        scores = scores1 + coef * scores2
        wer = compute_wer(dists, scores, total_len)
        wers.append(wer)
        if wer < best_wer:
            best_wer = wer
            best_coef = coef

    plt.plot(coefs, wers)
    plt.title(f'WER% after rescoring with different values of {param_name}')
    plt.ylabel('WER%')
    plt.xlabel(param_name)
    plt.show()
    return best_coef, best_wer


def compute_wer(dists, scores, total_len):
    """
    Sorts the candidates based on the scores and calculates the WER with the new top candidates.

    Args:
        dists: Tensor of the distances between the ground truth and the candidates with shape of [number of samples, beam size]
        scores: Tensor of the scores for candidates with shape of [number of samples, beam size]
        total_len: The total length of all samples

    Output:
        WER with the new scores
    """
    indices = scores.max(dim=1, keepdim=True)[1]
    wer = dists.gather(dim=1, index=indices).sum() / total_len
    wer = wer.item()
    return wer


def main():
    parser = ArgumentParser()
    parser.add_argument(
        "--lm_model_file",
        type=str,
        required=True,
        help="path to LM model .nemo file or the name of a HuggingFace pretrained models like 'transfo-xl-wt103' or 'gpt2'",
    )
    parser.add_argument("--beams_file", type=str, required=True, help="path to beams .tsv file")
    parser.add_argument(
        "--eval_manifest", type=str, required=True, help="path to the evaluation `.json` manifest file"
    )
    parser.add_argument("--beam_size", type=int, required=True, help="number of beams per candidate")
    parser.add_argument("--batch_size", type=int, default=256, help="inference batch size")
    parser.add_argument("--alpha", type=float, default=None, help="parameter alpha of the fusion")
    parser.add_argument("--beta", type=float, default=None, help="parameter beta of the fusion")
    parser.add_argument("--max_seq_length", default=512, help="Maximum sequence length (in tokens) for the input")
    parser.add_argument(
        "--scores_output_file", default=None, type=str, help="The optional path to store the rescored beams"
    )
    parser.add_argument(
        "--device", default="cuda", type=str, help="The device to load the model onto to calculate the scores"
    )
    parser.add_argument(
        "--use_amp", action="store_true", help="Whether to use AMP if available to calculate the scores"
    )
    args = parser.parse_args()

    device = args.device
    if device.startswith("cuda") and not torch.cuda.is_available():
        logging.info(f"cuda is not available! switched to cpu.")
        device = "cpu"

    if args.lm_model_file.endswith(".nemo"):
        nemo_model = True
        logging.info("Attempting to initialize from .nemo file...")
        model = TransformerLMModel.restore_from(
            restore_path=args.lm_model_file, map_location=torch.device(device)
        ).eval()
        model_tokenizer = model.tokenizer
    else:
        nemo_model = False
        logging.info("Attempting to initialize from a pretrained model from HuggingFace...")
        model = (
            AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path=args.lm_model_file, is_decoder=True)
            .to(device)
            .eval()
        )
        model_tokenizer = get_tokenizer(tokenizer_name=args.lm_model_file)

    max_seq_length = args.max_seq_length
    dataset = BeamScoresDataset(args.beams_file, model_tokenizer, args.eval_manifest, args.beam_size, max_seq_length)
    data_loader = torch.utils.data.DataLoader(dataset=dataset, batch_size=args.batch_size)

    if args.use_amp:
        if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'):
            logging.info("AMP is enabled!\n")
            autocast = torch.cuda.amp.autocast
    else:

        @contextlib.contextmanager
        def autocast():
            yield

    if "attention_mask" in inspect.getfullargspec(model.forward).args:
        support_att_mask = True
    else:
        support_att_mask = False
    logging.info(f"Rescoring with beam_size: {args.beam_size}")
    logging.info("Calculating the scores...")
    with autocast():
        with torch.no_grad():
            am_scores, lm_scores, dists, ref_lens, lens_in_chars = [], [], [], [], []
            for batch in tqdm.tqdm(data_loader):
                input_ids, input_mask, acoustic_score, dist, ref_len, len_in_chars, idx = batch

                max_len_in_batch = input_mask.sum(dim=0).argmin().item()
                input_ids, input_mask = input_ids[:, :max_len_in_batch], input_mask[:, :max_len_in_batch]
                if torch.cuda.is_available():
                    input_ids, input_mask = input_ids.to(device), input_mask.to(device)
                    dist, acoustic_score, len_in_chars = (
                        dist.to(device),
                        acoustic_score.to(device),
                        len_in_chars.to(device),
                    )
                # some models like Transformer-XL don't need attention_mask as input
                if support_att_mask:
                    log_probs = model(input_ids=input_ids, attention_mask=input_mask)
                else:
                    log_probs = model(input_ids=input_ids)

                if not nemo_model:
                    log_probs = torch.nn.functional.log_softmax(log_probs.logits, dim=-1)

                target_log_probs = log_probs[:, :-1].gather(2, input_ids[:, 1:].unsqueeze(2)).squeeze(2)
                neural_lm_score = torch.sum(target_log_probs * input_mask[:, 1:], dim=-1)

                am_scores.append(acoustic_score)
                lm_scores.append(neural_lm_score)
                dists.append(dist)
                ref_lens.append(ref_len)
                lens_in_chars.append(len_in_chars)

    am_scores = torch.cat(am_scores).view(-1, args.beam_size)
    lm_scores = torch.cat(lm_scores).view(-1, args.beam_size)
    dists = torch.cat(dists).view(-1, args.beam_size)
    ref_lens = torch.cat(ref_lens).view(-1, args.beam_size)
    lens_in_chars = torch.cat(lens_in_chars).view(-1, args.beam_size).to(am_scores.dtype)

    total_len = ref_lens[:, 0].sum()
    model_wer = dists[:, 0].sum() / total_len
    ideal_wer = dists.min(dim=1)[0].sum() / total_len

    if args.alpha is None:
        logging.info("Linear search for alpha...")
        coef1, _ = linear_search_wer(
            dists=dists, scores1=am_scores, scores2=lm_scores, total_len=total_len, param_name='alpha'
        )
        coef1 = np.round(coef1, 3)
        logging.info(f"alpha={coef1} achieved the best WER.")
        logging.info(f"------------------------------------------------")
    else:
        coef1 = args.alpha

    scores = am_scores + coef1 * lm_scores

    if args.beta is None:
        logging.info("Linear search for beta...")
        coef2, _ = linear_search_wer(
            dists=dists, scores1=scores, scores2=lens_in_chars, total_len=total_len, param_name='beta'
        )
        coef2 = np.round(coef2, 3)
        logging.info(f"beta={coef2} achieved the best WER.")
        logging.info(f"------------------------------------------------")
    else:
        coef2 = args.beta

    new_scores = am_scores + coef1 * lm_scores + coef2 * lens_in_chars
    rescored_wer = compute_wer(dists, new_scores, total_len)

    logging.info(f"Input beams WER: {np.round(model_wer.item() * 100, 2)}%")
    logging.info(f"------------------------------------------------")
    logging.info(f"  +LM rescoring WER: {np.round(rescored_wer * 100, 2)}%")
    logging.info(f"  with alpha={coef1}, beta={coef2}")
    logging.info(f"------------------------------------------------")
    logging.info(f"Oracle WER: {np.round(ideal_wer.item() * 100, 2)}%")
    logging.info(f"------------------------------------------------")

    new_scores_flatten = new_scores.flatten()
    if args.scores_output_file is not None:
        logging.info(f'Saving the candidates with their new scores at `{args.scores_output_file}`...')
        with open(args.scores_output_file, "w", encoding='utf-8') as fout:
            for sample_id in range(len(dataset)):
                fout.write(f"{dataset.data[0][sample_id]}\t{new_scores_flatten[sample_id]}\n")


if __name__ == '__main__':
    main()