update

examples/conv_tasnet/step_2_train_model.py

@@ -10,6 +10,9 @@ https://github.com/kaituoxu/Conv-TasNet/tree/master/src
 高要求场景（如医疗助听、语音识别）：
 需 SI-SNR ≥ 14 dB，并配合 PESQ ≥ 3.0 和 STOI ≥ 0.851812。
 
+DeepFilterNet2 模型在 DNS4 数据集，超过500小时的音频上训练了 100 个 epoch。
+https://arxiv.org/abs/2205.05474
+
 """
 import argparse
 import json

examples/conv_tasnet/yaml/config.yaml

@@ -19,10 +19,10 @@ mask_nonlinear: "relu"
 min_snr_db: -10
 max_snr_db: 20
 
-lr: 0.001
+lr: 0.005
 lr_scheduler: "CosineAnnealingLR"
 lr_scheduler_kwargs:
   T_max: 250000
-  eta_min: 0.00001
+  eta_min: 0.00005
 
 eval_steps: 25000

examples/conv_tasnet_gan/run.sh

@@ -0,0 +1,156 @@
+#!/usr/bin/env bash
+
+: <<'END'
+
+
+sh run.sh --stage 2 --stop_stage 2 --system_version centos --file_folder_name file_dir --final_model_name conv-tasnet-dns3-20250319 \
+--noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise/dns3-noise" \
+--speech_dir "/data/tianxing/HuggingDatasets/nx_noise/data/speech/dns3-speech" \
+--max_epochs 400
+
+
+END
+
+
+# params
+system_version="windows";
+verbose=true;
+stage=0 # start from 0 if you need to start from data preparation
+stop_stage=9
+
+work_dir="$(pwd)"
+file_folder_name=file_folder_name
+final_model_name=final_model_name
+config_file="yaml/config.yaml"
+discriminator_config_file="yaml/discriminator_config.yaml"
+limit=10
+
+noise_dir=/data/tianxing/HuggingDatasets/nx_noise/data/noise
+speech_dir=/data/tianxing/HuggingDatasets/aishell/data_aishell/wav/train
+
+max_count=10000000
+
+nohup_name=nohup.out
+
+# model params
+batch_size=64
+max_epochs=200
+save_top_k=10
+patience=5
+
+
+# parse options
+while true; do
+  [ -z "${1:-}" ] && break;  # break if there are no arguments
+  case "$1" in
+    --*) name=$(echo "$1" | sed s/^--// | sed s/-/_/g);
+      eval '[ -z "${'"$name"'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
+      old_value="(eval echo \\$$name)";
+      if [ "${old_value}" == "true" ] || [ "${old_value}" == "false" ]; then
+        was_bool=true;
+      else
+        was_bool=false;
+      fi
+
+      # Set the variable to the right value-- the escaped quotes make it work if
+      # the option had spaces, like --cmd "queue.pl -sync y"
+      eval "${name}=\"$2\"";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
+        exit 1;
+      fi
+      shift 2;
+      ;;
+
+    *) break;
+  esac
+done
+
+file_dir="${work_dir}/${file_folder_name}"
+final_model_dir="${work_dir}/../../trained_models/${final_model_name}";
+evaluation_audio_dir="${file_dir}/evaluation_audio"
+
+train_dataset="${file_dir}/train.jsonl"
+valid_dataset="${file_dir}/valid.jsonl"
+
+$verbose && echo "system_version: ${system_version}"
+$verbose && echo "file_folder_name: ${file_folder_name}"
+
+if [ $system_version == "windows" ]; then
+  alias python3='D:/Users/tianx/PycharmProjects/virtualenv/nx_denoise/Scripts/python.exe'
+elif [ $system_version == "centos" ] || [ $system_version == "ubuntu" ]; then
+  #source /data/local/bin/nx_denoise/bin/activate
+  alias python3='/data/local/bin/nx_denoise/bin/python3'
+fi
+
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+  $verbose && echo "stage 1: prepare data"
+  cd "${work_dir}" || exit 1
+  python3 step_1_prepare_data.py \
+  --file_dir "${file_dir}" \
+  --noise_dir "${noise_dir}" \
+  --speech_dir "${speech_dir}" \
+  --train_dataset "${train_dataset}" \
+  --valid_dataset "${valid_dataset}" \
+  --max_count "${max_count}" \
+
+fi
+
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+  $verbose && echo "stage 2: train model"
+  cd "${work_dir}" || exit 1
+  python3 step_2_train_model.py \
+  --train_dataset "${train_dataset}" \
+  --valid_dataset "${valid_dataset}" \
+  --serialization_dir "${file_dir}" \
+  --config_file "${config_file}" \
+  --discriminator_config_file "${discriminator_config_file}" \
+
+fi
+
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+  $verbose && echo "stage 3: test model"
+  cd "${work_dir}" || exit 1
+  python3 step_3_evaluation.py \
+  --valid_dataset "${valid_dataset}" \
+  --model_dir "${file_dir}/best" \
+  --evaluation_audio_dir "${evaluation_audio_dir}" \
+  --limit "${limit}" \
+
+fi
+
+
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+  $verbose && echo "stage 4: collect files"
+  cd "${work_dir}" || exit 1
+
+  mkdir -p ${final_model_dir}
+
+  cp "${file_dir}/best"/* "${final_model_dir}"
+  cp -r "${file_dir}/evaluation_audio" "${final_model_dir}"
+
+  cd "${final_model_dir}/.." || exit 1;
+
+  if [ -e "${final_model_name}.zip" ]; then
+    rm -rf "${final_model_name}_backup.zip"
+    mv "${final_model_name}.zip" "${final_model_name}_backup.zip"
+  fi
+
+  zip -r "${final_model_name}.zip" "${final_model_name}"
+  rm -rf "${final_model_name}"
+
+fi
+
+
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+  $verbose && echo "stage 5: clear file_dir"
+  cd "${work_dir}" || exit 1
+
+  rm -rf "${file_dir}";
+
+fi

examples/conv_tasnet_gan/step_1_prepare_data.py

@@ -0,0 +1,162 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+import argparse
+import json
+import os
+from pathlib import Path
+import random
+import sys
+
+pwd = os.path.abspath(os.path.dirname(__file__))
+sys.path.append(os.path.join(pwd, "../../"))
+
+import librosa
+import numpy as np
+from tqdm import tqdm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--file_dir", default="./", type=str)
+
+    parser.add_argument(
+        "--noise_dir",
+        default=r"E:\Users\tianx\HuggingDatasets\nx_noise\data\noise",
+        type=str
+    )
+    parser.add_argument(
+        "--speech_dir",
+        default=r"E:\programmer\asr_datasets\aishell\data_aishell\wav\train",
+        type=str
+    )
+
+    parser.add_argument("--train_dataset", default="train.jsonl", type=str)
+    parser.add_argument("--valid_dataset", default="valid.jsonl", type=str)
+
+    parser.add_argument("--duration", default=4.0, type=float)
+    parser.add_argument("--min_snr_db", default=-10, type=float)
+    parser.add_argument("--max_snr_db", default=20, type=float)
+
+    parser.add_argument("--target_sample_rate", default=8000, type=int)
+
+    parser.add_argument("--max_count", default=10000, type=int)
+
+    args = parser.parse_args()
+    return args
+
+
+def filename_generator(data_dir: str):
+    data_dir = Path(data_dir)
+    for filename in data_dir.glob("**/*.wav"):
+        yield filename.as_posix()
+
+
+def target_second_signal_generator(data_dir: str, duration: int = 2, sample_rate: int = 8000, max_epoch: int = 20000):
+    data_dir = Path(data_dir)
+    for epoch_idx in range(max_epoch):
+        for filename in data_dir.glob("**/*.wav"):
+            signal, _ = librosa.load(filename.as_posix(), sr=sample_rate)
+            raw_duration = librosa.get_duration(y=signal, sr=sample_rate)
+
+            if raw_duration < duration:
+                # print(f"duration less than {duration} s. skip filename: {filename.as_posix()}")
+                continue
+            if signal.ndim != 1:
+                raise AssertionError(f"expected ndim 1, instead of {signal.ndim}")
+
+            signal_length = len(signal)
+            win_size = int(duration * sample_rate)
+            for begin in range(0, signal_length - win_size, win_size):
+                if np.sum(signal[begin: begin+win_size]) == 0:
+                    continue
+                row = {
+                    "epoch_idx": epoch_idx,
+                    "filename": filename.as_posix(),
+                    "raw_duration": round(raw_duration, 4),
+                    "offset": round(begin / sample_rate, 4),
+                    "duration": round(duration, 4),
+                }
+                yield row
+
+
+def main():
+    args = get_args()
+
+    file_dir = Path(args.file_dir)
+    file_dir.mkdir(exist_ok=True)
+
+    noise_dir = Path(args.noise_dir)
+    speech_dir = Path(args.speech_dir)
+
+    noise_generator = target_second_signal_generator(
+        noise_dir.as_posix(),
+        duration=args.duration,
+        sample_rate=args.target_sample_rate,
+        max_epoch=100000,
+    )
+    speech_generator = target_second_signal_generator(
+        speech_dir.as_posix(),
+        duration=args.duration,
+        sample_rate=args.target_sample_rate,
+        max_epoch=1,
+    )
+
+    dataset = list()
+
+    count = 0
+    process_bar = tqdm(desc="build dataset excel")
+    with open(args.train_dataset, "w", encoding="utf-8") as ftrain, open(args.valid_dataset, "w", encoding="utf-8") as fvalid:
+        for noise, speech in zip(noise_generator, speech_generator):
+            if count >= args.max_count:
+                break
+
+            noise_filename = noise["filename"]
+            noise_raw_duration = noise["raw_duration"]
+            noise_offset = noise["offset"]
+            noise_duration = noise["duration"]
+
+            speech_filename = speech["filename"]
+            speech_raw_duration = speech["raw_duration"]
+            speech_offset = speech["offset"]
+            speech_duration = speech["duration"]
+
+            random1 = random.random()
+            random2 = random.random()
+
+            row = {
+                "noise_filename": noise_filename,
+                "noise_raw_duration": noise_raw_duration,
+                "noise_offset": noise_offset,
+                "noise_duration": noise_duration,
+
+                "speech_filename": speech_filename,
+                "speech_raw_duration": speech_raw_duration,
+                "speech_offset": speech_offset,
+                "speech_duration": speech_duration,
+
+                "snr_db": random.uniform(args.min_snr_db, args.max_snr_db),
+
+                "random1": random1,
+            }
+            row = json.dumps(row, ensure_ascii=False)
+            if random2 < (1 / 300 / 1):
+                fvalid.write(f"{row}\n")
+            else:
+                ftrain.write(f"{row}\n")
+
+            count += 1
+            duration_seconds = count * args.duration
+            duration_hours = duration_seconds / 3600
+
+            process_bar.update(n=1)
+            process_bar.set_postfix({
+                # "duration_seconds": round(duration_seconds, 4),
+                "duration_hours": round(duration_hours, 4),
+
+            })
+
+    return
+
+
+if __name__ == "__main__":
+    main()

examples/conv_tasnet_gan/step_2_train_model.py

@@ -0,0 +1,579 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+"""
+https://github.com/kaituoxu/Conv-TasNet/tree/master/src
+
+一般场景：
+
+目标 SI-SNR ≥ 10 dB，适用于电话通信、基础语音助手等。
+
+高要求场景（如医疗助听、语音识别）：
+需 SI-SNR ≥ 14 dB，并配合 PESQ ≥ 3.0 和 STOI ≥ 0.851812。
+
+DeepFilterNet2 模型在 DNS4 数据集，超过500小时的音频上训练了 100 个 epoch。
+https://arxiv.org/abs/2205.05474
+
+"""
+import argparse
+import json
+import logging
+from logging.handlers import TimedRotatingFileHandler
+import os
+import platform
+from pathlib import Path
+import random
+import sys
+import shutil
+from typing import List
+
+pwd = os.path.abspath(os.path.dirname(__file__))
+sys.path.append(os.path.join(pwd, "../../"))
+
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from torch.utils.data.dataloader import DataLoader
+from tqdm import tqdm
+
+from toolbox.torch.utils.data.dataset.denoise_jsonl_dataset import DenoiseJsonlDataset
+from toolbox.torchaudio.models.conv_tasnet.configuration_conv_tasnet import ConvTasNetConfig
+from toolbox.torchaudio.models.conv_tasnet.modeling_conv_tasnet import ConvTasNet, ConvTasNetPretrainedModel
+from toolbox.torchaudio.models.discriminators.conv_tasnet_discriminator.modeling_conv_tasnet_discriminator import ConvTasNetDiscriminatorPretrainedModel
+from toolbox.torchaudio.models.discriminators.conv_tasnet_discriminator.configuration_conv_tasnet_discriminator import ConvTasNetDiscriminatorConfig
+from toolbox.torchaudio.models.nx_clean_unet.metrics import run_batch_pesq, run_pesq_score
+from toolbox.torchaudio.losses.snr import NegativeSISNRLoss
+from toolbox.torchaudio.losses.spectral import LSDLoss, MultiResolutionSTFTLoss
+from toolbox.torchaudio.losses.perceptual import NegSTOILoss, PesqLoss
+from toolbox.torchaudio.metrics.pesq import run_pesq_score
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--train_dataset", default="train.xlsx", type=str)
+    parser.add_argument("--valid_dataset", default="valid.xlsx", type=str)
+
+    parser.add_argument("--max_epochs", default=200, type=int)
+
+    parser.add_argument("--batch_size", default=8, type=int)
+    parser.add_argument("--num_serialized_models_to_keep", default=10, type=int)
+    parser.add_argument("--patience", default=5, type=int)
+    parser.add_argument("--serialization_dir", default="serialization_dir", type=str)
+    parser.add_argument("--seed", default=1234, type=int)
+
+    parser.add_argument("--config_file", default="config.yaml", type=str)
+    parser.add_argument("--discriminator_config_file", default="discriminator_config.yaml", type=str)
+
+    args = parser.parse_args()
+    return args
+
+
+def logging_config(file_dir: str):
+    fmt = "%(asctime)s - %(name)s - %(levelname)s  %(filename)s:%(lineno)d >  %(message)s"
+
+    logging.basicConfig(format=fmt,
+                        datefmt="%m/%d/%Y %H:%M:%S",
+                        level=logging.INFO)
+    file_handler = TimedRotatingFileHandler(
+        filename=os.path.join(file_dir, "main.log"),
+        encoding="utf-8",
+        when="D",
+        interval=1,
+        backupCount=7
+    )
+    file_handler.setLevel(logging.INFO)
+    file_handler.setFormatter(logging.Formatter(fmt))
+    logger = logging.getLogger(__name__)
+    logger.addHandler(file_handler)
+
+    return logger
+
+
+class CollateFunction(object):
+    def __init__(self):
+        pass
+
+    def __call__(self, batch: List[dict]):
+        clean_audios = list()
+        noisy_audios = list()
+
+        for sample in batch:
+            # noise_wave: torch.Tensor = sample["noise_wave"]
+            clean_audio: torch.Tensor = sample["speech_wave"]
+            noisy_audio: torch.Tensor = sample["mix_wave"]
+            # snr_db: float = sample["snr_db"]
+
+            clean_audios.append(clean_audio)
+            noisy_audios.append(noisy_audio)
+
+        clean_audios = torch.stack(clean_audios)
+        noisy_audios = torch.stack(noisy_audios)
+
+        # assert
+        if torch.any(torch.isnan(clean_audios)) or torch.any(torch.isinf(clean_audios)):
+            raise AssertionError("nan or inf in clean_audios")
+        if torch.any(torch.isnan(noisy_audios)) or torch.any(torch.isinf(noisy_audios)):
+            raise AssertionError("nan or inf in noisy_audios")
+        return clean_audios, noisy_audios
+
+
+collate_fn = CollateFunction()
+
+
+def main():
+    args = get_args()
+
+    config = ConvTasNetConfig.from_pretrained(
+        pretrained_model_name_or_path=args.config_file,
+    )
+    discriminator_config = ConvTasNetDiscriminatorConfig.from_pretrained(
+        pretrained_model_name_or_path=args.discriminator_config_file,
+    )
+
+    serialization_dir = Path(args.serialization_dir)
+    serialization_dir.mkdir(parents=True, exist_ok=True)
+
+    logger = logging_config(serialization_dir)
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    logger.info(f"set seed: {args.seed}")
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    n_gpu = torch.cuda.device_count()
+    logger.info(f"GPU available count: {n_gpu}; device: {device}")
+
+    # datasets
+    train_dataset = DenoiseJsonlDataset(
+        jsonl_file=args.train_dataset,
+        expected_sample_rate=config.sample_rate,
+        max_wave_value=32768.0,
+        min_snr_db=config.min_snr_db,
+        max_snr_db=config.max_snr_db,
+        # skip=825000,
+    )
+    valid_dataset = DenoiseJsonlDataset(
+        jsonl_file=args.valid_dataset,
+        expected_sample_rate=config.sample_rate,
+        max_wave_value=32768.0,
+        min_snr_db=config.min_snr_db,
+        max_snr_db=config.max_snr_db,
+    )
+    train_data_loader = DataLoader(
+        dataset=train_dataset,
+        batch_size=args.batch_size,
+        # shuffle=True,
+        sampler=None,
+        # Linux 系统中可以使用多个子进程加载数据, 而在 Windows 系统中不能.
+        num_workers=0 if platform.system() == "Windows" else os.cpu_count() // 2,
+        collate_fn=collate_fn,
+        pin_memory=False,
+        prefetch_factor=2,
+    )
+    valid_data_loader = DataLoader(
+        dataset=valid_dataset,
+        batch_size=args.batch_size,
+        # shuffle=True,
+        sampler=None,
+        # Linux 系统中可以使用多个子进程加载数据, 而在 Windows 系统中不能.
+        num_workers=0 if platform.system() == "Windows" else os.cpu_count() // 2,
+        collate_fn=collate_fn,
+        pin_memory=False,
+        prefetch_factor=2,
+    )
+
+    # models
+    logger.info(f"prepare models. config_file: {args.config_file}")
+    model = ConvTasNetPretrainedModel(config).to(device)
+    model.to(device)
+    model.train()
+
+    discriminator = ConvTasNetDiscriminatorPretrainedModel(discriminator_config).to(device)
+    discriminator.to(device)
+    discriminator.train()
+
+    # optimizer
+    logger.info("prepare optimizer, lr_scheduler, loss_fn, categorical_accuracy")
+    optimizer = torch.optim.AdamW(model.parameters(), config.lr)
+    discriminator_optimizer = torch.optim.AdamW(discriminator.parameters(), config.lr, betas=[config.adam_b1, config.adam_b2])
+
+    # resume training
+    last_step_idx = -1
+    last_epoch = -1
+    for step_idx_str in serialization_dir.glob("steps-*"):
+        step_idx_str = Path(step_idx_str)
+        step_idx = step_idx_str.stem.split("-")[1]
+        step_idx = int(step_idx)
+        if step_idx > last_step_idx:
+            last_step_idx = step_idx
+
+    if last_step_idx != -1:
+        logger.info(f"resume from steps-{last_step_idx}.")
+        model_pt = serialization_dir / f"steps-{last_step_idx}/model.pt"
+        discriminator_pt = serialization_dir / f"steps-{last_step_idx}/discriminator.pt"
+
+        optimizer_pth = serialization_dir / f"steps-{last_step_idx}/optimizer.pth"
+        discriminator_optimizer_pth = serialization_dir / f"steps-{last_step_idx}/discriminator_optimizer.pth"
+
+        logger.info(f"load state dict for model.")
+        with open(model_pt.as_posix(), "rb") as f:
+            state_dict = torch.load(f, map_location="cpu", weights_only=True)
+        model.load_state_dict(state_dict, strict=True)
+
+        logger.info(f"load state dict for optimizer.")
+        with open(optimizer_pth.as_posix(), "rb") as f:
+            state_dict = torch.load(f, map_location="cpu", weights_only=True)
+        optimizer.load_state_dict(state_dict)
+
+        if discriminator_pt.exists():
+            logger.info(f"load state dict for discriminator.")
+            with open(model_pt.as_posix(), "rb") as f:
+                state_dict = torch.load(f, map_location="cpu", weights_only=True)
+            discriminator.load_state_dict(state_dict, strict=True)
+
+        if discriminator_optimizer_pth.exists():
+            logger.info(f"load state dict for discriminator_optimizer.")
+            with open(optimizer_pth.as_posix(), "rb") as f:
+                state_dict = torch.load(f, map_location="cpu", weights_only=True)
+            discriminator_optimizer.load_state_dict(state_dict)
+
+    if config.lr_scheduler == "CosineAnnealingLR":
+        lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
+            optimizer,
+            last_epoch=last_epoch,
+            # T_max=10 * config.eval_steps,
+            # eta_min=0.01 * config.lr,
+            **config.lr_scheduler_kwargs,
+        )
+        discriminator_lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
+            discriminator_optimizer,
+            last_epoch=last_epoch,
+            # T_max=10 * config.eval_steps,
+            # eta_min=0.01 * config.lr,
+            **config.lr_scheduler_kwargs,
+        )
+    elif config.lr_scheduler == "MultiStepLR":
+        lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
+            optimizer,
+            last_epoch=last_epoch,
+            milestones=[10000, 20000, 30000, 40000, 50000], gamma=0.5
+        )
+        discriminator_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
+            discriminator_optimizer,
+            last_epoch=last_epoch,
+            milestones=[10000, 20000, 30000, 40000, 50000], gamma=0.5
+        )
+    else:
+        raise AssertionError(f"invalid lr_scheduler: {config.lr_scheduler}")
+
+    ae_loss_fn = nn.L1Loss(reduction="mean").to(device)
+    neg_si_snr_loss_fn = NegativeSISNRLoss(reduction="mean").to(device)
+    neg_stoi_loss_fn = NegSTOILoss(sample_rate=config.sample_rate, reduction="mean").to(device)
+    mr_stft_loss_fn = MultiResolutionSTFTLoss(
+        fft_size_list=[256, 512, 1024],
+        win_size_list=[120, 240, 480],
+        hop_size_list=[25, 50, 100],
+        factor_sc=1.5,
+        factor_mag=1.0,
+        reduction="mean"
+    ).to(device)
+    pesq_loss_fn = PesqLoss(0.5, sample_rate=config.sample_rate).to(device)
+
+    # training loop
+
+    # state
+    average_pesq_score = 1000000000
+    average_loss = 1000000000
+    average_ae_loss = 1000000000
+    average_neg_si_snr_loss = 1000000000
+    average_neg_stoi_loss = 1000000000
+    average_mr_stft_loss = 1000000000
+    average_pesq_loss = 1000000000
+    average_discriminator_g_loss = 1000000000
+    average_discriminator_d_loss = 1000000000
+
+    model_list = list()
+    best_epoch_idx = None
+    best_step_idx = None
+    best_metric = None
+    patience_count = 0
+
+    step_idx = 0 if last_step_idx == -1 else last_step_idx
+
+    logger.info("training")
+    for epoch_idx in range(max(0, last_epoch+1), args.max_epochs):
+        # train
+        model.train()
+
+        total_pesq_score = 0.
+        total_loss = 0.
+        total_ae_loss = 0.
+        total_neg_si_snr_loss = 0.
+        total_neg_stoi_loss = 0.
+        total_mr_stft_loss = 0.
+        total_pesq_loss = 0.
+        total_discriminator_g_loss = 0.
+        total_discriminator_d_loss = 0.
+        total_batches = 0.
+
+        progress_bar_train = tqdm(
+            initial=step_idx,
+            desc="Training; epoch-{}".format(epoch_idx),
+        )
+        for train_batch in train_data_loader:
+            clean_audios, noisy_audios = train_batch
+            clean_audios: torch.Tensor = clean_audios.to(device)
+            noisy_audios: torch.Tensor = noisy_audios.to(device)
+            one_labels = torch.ones(clean_audios.shape[0]).to(device)
+
+            denoise_audios = model.forward(noisy_audios)
+            denoise_audios = torch.squeeze(denoise_audios, dim=1)
+
+            if torch.any(torch.isnan(denoise_audios)) or torch.any(torch.isinf(denoise_audios)):
+                raise AssertionError("nan or inf in denoise_audios")
+
+            # Discriminator
+            clean_audio_list = torch.split(clean_audios, 1, dim=0)
+            enhanced_audio_list = torch.split(denoise_audios, 1, dim=0)
+            clean_audio_list = [t.squeeze().detach().cpu().numpy() for t in clean_audio_list]
+            enhanced_audio_list = [t.squeeze().detach().cpu().numpy() for t in enhanced_audio_list]
+
+            pesq_score_list: List[float] = run_batch_pesq(clean_audio_list, enhanced_audio_list, sample_rate=config.sample_rate, mode="nb")
+
+            metric_r = discriminator.forward(clean_audios, clean_audios)
+            metric_g = discriminator.forward(clean_audios, denoise_audios.detach())
+            loss_disc_r = F.mse_loss(one_labels, metric_r.flatten())
+
+            if -1 in pesq_score_list:
+                # print("-1 in batch_pesq_score!")
+                loss_disc_g = 0
+            else:
+                pesq_score_list: torch.FloatTensor = torch.tensor([(score - 1) / 3.5 for score in pesq_score_list], dtype=torch.float32)
+                loss_disc_g = F.mse_loss(pesq_score_list.to(device), metric_g.flatten())
+
+            discriminator_d_loss = loss_disc_r + loss_disc_g
+            discriminator_optimizer.zero_grad()
+            discriminator_d_loss.backward()
+            discriminator_optimizer.step()
+            discriminator_lr_scheduler.step()
+
+            # Generator
+            ae_loss = ae_loss_fn.forward(denoise_audios, clean_audios)
+            neg_si_snr_loss = neg_si_snr_loss_fn.forward(denoise_audios, clean_audios)
+            neg_stoi_loss = neg_stoi_loss_fn.forward(denoise_audios, clean_audios)
+            mr_stft_loss = mr_stft_loss_fn.forward(denoise_audios, clean_audios)
+            pesq_loss = pesq_loss_fn.forward(clean_audios, denoise_audios)
+
+            metric_g = discriminator.forward(denoise_audios, clean_audios)
+            discriminator_g_loss = F.mse_loss(metric_g.flatten(), one_labels)
+
+            loss = 1.0 * ae_loss + 0.8 * neg_si_snr_loss + 0.7 * mr_stft_loss + 0.5 * neg_stoi_loss + 0.5 * pesq_loss + 0.2 * discriminator_g_loss
+            if torch.any(torch.isnan(loss)) or torch.any(torch.isinf(loss)):
+                logger.info(f"find nan or inf in loss.")
+                continue
+
+            denoise_audios_list_r = list(denoise_audios.detach().cpu().numpy())
+            clean_audios_list_r = list(clean_audios.detach().cpu().numpy())
+            pesq_score = run_pesq_score(clean_audios_list_r, denoise_audios_list_r, sample_rate=config.sample_rate, mode="nb")
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+            lr_scheduler.step()
+
+            total_pesq_score += pesq_score
+            total_loss += loss.item()
+            total_ae_loss += ae_loss.item()
+            total_neg_si_snr_loss += neg_si_snr_loss.item()
+            total_neg_stoi_loss += neg_stoi_loss.item()
+            total_mr_stft_loss += mr_stft_loss.item()
+            total_pesq_loss += pesq_loss.item()
+            total_discriminator_g_loss += discriminator_g_loss.item()
+            total_discriminator_d_loss += discriminator_d_loss.item()
+            total_batches += 1
+
+            average_pesq_score = round(total_pesq_score / total_batches, 4)
+            average_loss = round(total_loss / total_batches, 4)
+            average_ae_loss = round(total_ae_loss / total_batches, 4)
+            average_neg_si_snr_loss = round(total_neg_si_snr_loss / total_batches, 4)
+            average_neg_stoi_loss = round(total_neg_stoi_loss / total_batches, 4)
+            average_mr_stft_loss = round(total_mr_stft_loss / total_batches, 4)
+            average_pesq_loss = round(total_pesq_loss / total_batches, 4)
+            average_discriminator_g_loss = round(total_discriminator_g_loss / total_batches, 4)
+            average_discriminator_d_loss = round(total_discriminator_d_loss / total_batches, 4)
+
+            progress_bar_train.update(1)
+            progress_bar_train.set_postfix({
+                "lr": lr_scheduler.get_last_lr()[0],
+                "pesq_score": average_pesq_score,
+                "loss": average_loss,
+                "ae_loss": average_ae_loss,
+                "neg_si_snr_loss": average_neg_si_snr_loss,
+                "neg_stoi_loss": average_neg_stoi_loss,
+                "mr_stft_loss": average_mr_stft_loss,
+                "pesq_loss": average_pesq_loss,
+                "disc_g_loss": average_discriminator_g_loss,
+                "disc_d_loss": average_discriminator_d_loss,
+
+            })
+
+            # evaluation
+            step_idx += 1
+            if step_idx % config.eval_steps == 0:
+                with torch.no_grad():
+                    torch.cuda.empty_cache()
+
+                    total_pesq_score = 0.
+                    total_loss = 0.
+                    total_ae_loss = 0.
+                    total_neg_si_snr_loss = 0.
+                    total_neg_stoi_loss = 0.
+                    total_mr_stft_loss = 0.
+                    total_pesq_loss = 0.
+                    total_batches = 0.
+
+                    progress_bar_train.close()
+                    progress_bar_eval = tqdm(
+                        desc="Evaluation; steps-{}k".format(int(step_idx/1000)),
+                    )
+                    for eval_batch in valid_data_loader:
+                        clean_audios, noisy_audios = eval_batch
+                        clean_audios = clean_audios.to(device)
+                        noisy_audios = noisy_audios.to(device)
+
+                        denoise_audios = model.forward(noisy_audios)
+                        denoise_audios = torch.squeeze(denoise_audios, dim=1)
+
+                        # Generator
+                        ae_loss = ae_loss_fn.forward(denoise_audios, clean_audios)
+                        neg_si_snr_loss = neg_si_snr_loss_fn.forward(denoise_audios, clean_audios)
+                        neg_stoi_loss = neg_stoi_loss_fn.forward(denoise_audios, clean_audios)
+                        mr_stft_loss = mr_stft_loss_fn.forward(denoise_audios, clean_audios)
+                        pesq_loss = pesq_loss_fn.forward(clean_audios, denoise_audios)
+
+                        loss = 1.0 * ae_loss + 0.8 * neg_si_snr_loss + 0.7 * mr_stft_loss + 0.5 * neg_stoi_loss + 0.5 * pesq_loss
+                        if torch.any(torch.isnan(loss)) or torch.any(torch.isinf(loss)):
+                            logger.info(f"find nan or inf in loss.")
+                            continue
+
+                        denoise_audios_list_r = list(denoise_audios.detach().cpu().numpy())
+                        clean_audios_list_r = list(clean_audios.detach().cpu().numpy())
+                        pesq_score = run_pesq_score(clean_audios_list_r, denoise_audios_list_r, sample_rate=config.sample_rate, mode="nb")
+
+                        total_pesq_score += pesq_score
+                        total_loss += loss.item()
+                        total_ae_loss += ae_loss.item()
+                        total_neg_si_snr_loss += neg_si_snr_loss.item()
+                        total_neg_stoi_loss += neg_stoi_loss.item()
+                        total_mr_stft_loss += mr_stft_loss.item()
+                        total_pesq_loss += pesq_loss.item()
+                        total_batches += 1
+
+                        average_pesq_score = round(total_pesq_score / total_batches, 4)
+                        average_loss = round(total_loss / total_batches, 4)
+                        average_ae_loss = round(total_ae_loss / total_batches, 4)
+                        average_neg_si_snr_loss = round(total_neg_si_snr_loss / total_batches, 4)
+                        average_neg_stoi_loss = round(total_neg_stoi_loss / total_batches, 4)
+                        average_mr_stft_loss = round(total_mr_stft_loss / total_batches, 4)
+                        average_pesq_loss = round(total_pesq_loss / total_batches, 4)
+
+                        progress_bar_eval.update(1)
+                        progress_bar_eval.set_postfix({
+                            "lr": lr_scheduler.get_last_lr()[0],
+                            "pesq_score": average_pesq_score,
+                            "loss": average_loss,
+                            "ae_loss": average_ae_loss,
+                            "neg_si_snr_loss": average_neg_si_snr_loss,
+                            "neg_stoi_loss": average_neg_stoi_loss,
+                            "mr_stft_loss": average_mr_stft_loss,
+                            "pesq_loss": average_pesq_loss,
+                        })
+
+                total_pesq_score = 0.
+                total_loss = 0.
+                total_ae_loss = 0.
+                total_neg_si_snr_loss = 0.
+                total_neg_stoi_loss = 0.
+                total_mr_stft_loss = 0.
+                total_pesq_loss = 0.
+                total_batches = 0.
+
+                progress_bar_eval.close()
+                progress_bar_train = tqdm(
+                    initial=progress_bar_train.n,
+                    postfix=progress_bar_train.postfix,
+                    desc=progress_bar_train.desc,
+                )
+
+                # save path
+                save_dir = serialization_dir / "steps-{}".format(step_idx)
+                save_dir.mkdir(parents=True, exist_ok=False)
+
+                # save models
+                model.save_pretrained(save_dir.as_posix())
+                discriminator.save_pretrained(save_dir.as_posix())
+
+                # save optim
+                torch.save(optimizer.state_dict(), (save_dir / "optimizer.pth").as_posix())
+                torch.save(discriminator_optimizer.state_dict(), (save_dir / "discriminator_optimizer.pth").as_posix())
+
+                model_list.append(save_dir)
+                if len(model_list) >= args.num_serialized_models_to_keep:
+                    model_to_delete: Path = model_list.pop(0)
+                    shutil.rmtree(model_to_delete.as_posix())
+
+                # save metric
+                if best_metric is None:
+                    best_epoch_idx = epoch_idx
+                    best_step_idx = step_idx
+                    best_metric = average_pesq_score
+                elif average_pesq_score > best_metric:
+                    # great is better.
+                    best_epoch_idx = epoch_idx
+                    best_step_idx = step_idx
+                    best_metric = average_pesq_score
+                else:
+                    pass
+
+                metrics = {
+                    "epoch_idx": epoch_idx,
+                    "best_epoch_idx": best_epoch_idx,
+                    "best_step_idx": best_step_idx,
+                    "pesq_score": average_pesq_score,
+                    "loss": average_loss,
+                    "ae_loss": average_ae_loss,
+                    "neg_si_snr_loss": average_neg_si_snr_loss,
+                    "neg_stoi_loss": average_neg_stoi_loss,
+                    "mr_stft_loss": average_mr_stft_loss,
+                    "pesq_loss": average_pesq_loss,
+                }
+                metrics_filename = save_dir / "metrics_epoch.json"
+                with open(metrics_filename, "w", encoding="utf-8") as f:
+                    json.dump(metrics, f, indent=4, ensure_ascii=False)
+
+                # save best
+                best_dir = serialization_dir / "best"
+                if best_epoch_idx == epoch_idx and best_step_idx == step_idx:
+                    if best_dir.exists():
+                        shutil.rmtree(best_dir)
+                    shutil.copytree(save_dir, best_dir)
+
+                # early stop
+                early_stop_flag = False
+                if best_epoch_idx == epoch_idx and best_step_idx == step_idx:
+                    patience_count = 0
+                else:
+                    patience_count += 1
+                if patience_count >= args.patience:
+                    early_stop_flag = True
+
+                # early stop
+                if early_stop_flag:
+                    break
+
+    return
+
+
+if __name__ == "__main__":
+    main()

examples/conv_tasnet_gan/yaml/config.yaml

@@ -0,0 +1,31 @@
+model_name: "conv_tasnet_gan"
+
+sample_rate: 8000
+segment_size: 4
+
+win_size: 20
+freq_bins: 256
+bottleneck_channels: 256
+num_speakers: 1
+num_blocks: 4
+num_sub_blocks: 8
+sub_blocks_channels: 512
+sub_blocks_kernel_size: 3
+
+norm_type: "gLN"
+causal: false
+mask_nonlinear: "relu"
+
+min_snr_db: -10
+max_snr_db: 20
+
+lr: 0.005
+adam_b1: 0.8
+adam_b2: 0.99
+
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.00005
+
+eval_steps: 25000

examples/conv_tasnet_gan/yaml/discriminator_config.yaml

@@ -0,0 +1,10 @@
+model_name: "conv_tasnet_gan"
+
+sample_rate: 8000
+segment_size: 16000
+n_fft: 512
+win_size: 200
+hop_size: 80
+
+discriminator_dim: 32
+discriminator_in_channel: 2

examples/nx_clean_unet/step_2_train_model.py

@@ -285,7 +285,8 @@ def main():
             # Time Loss
             loss_time = F.l1_loss(clean_audios, audio_g)
             # Metric Loss
-            metric_g = discriminator.forward(clean_audios, audio_g.detach())
+            metric_g = discriminator.forward(clean_audios, audio_g)
+            # metric_g = discriminator.forward(clean_audios, audio_g.detach())
             loss_metric = F.mse_loss(metric_g.flatten(), one_labels)
 
             # loss_gen_all = loss_mag * 0.9 + loss_pha * 0.3  + loss_com * 0.1 + loss_metric * 0.05 + loss_time * 0.2

toolbox/torchaudio/configuration_utils.py

@@ -8,6 +8,7 @@ import yaml
 
 
 CONFIG_FILE = "config.yaml"
+DISCRIMINATOR_CONFIG_FILE = "discriminator_config.yaml"
 
 
 class PretrainedConfig(object):

toolbox/torchaudio/models/conv_tasnet/configuration_conv_tasnet.py

@@ -31,6 +31,9 @@ class ConvTasNetConfig(PretrainedConfig):
                  max_snr_db: float = 20,
 
                  lr: float = 1e-3,
+                 adam_b1: float = 0.8,
+                 adam_b2: float = 0.99,
+
                  lr_scheduler: str = "CosineAnnealingLR",
                  lr_scheduler_kwargs: dict = None,
 
@@ -60,6 +63,9 @@ class ConvTasNetConfig(PretrainedConfig):
         self.max_snr_db = max_snr_db
 
         self.lr = lr
+        self.adam_b1 = adam_b1
+        self.adam_b2 = adam_b2
+
         self.lr_scheduler = lr_scheduler
         self.lr_scheduler_kwargs = lr_scheduler_kwargs or dict()
 

toolbox/torchaudio/models/conv_tasnet/inference_conv_tasnet.py

@@ -83,7 +83,7 @@ class InferenceConvTasNet(object):
 
 
 def main():
-    model_zip_file = project_path / "trained_models/conv-tasnet-dns3-575k-steps.zip"
+    model_zip_file = project_path / "trained_models/conv-tasnet-dns3-1025k-steps.zip"
     infer_conv_tasnet = InferenceConvTasNet(model_zip_file)
 
     sample_rate = 8000

toolbox/torchaudio/models/discriminators/__init__.py

@@ -0,0 +1,6 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+
+
+if __name__ == "__main__":
+    pass

toolbox/torchaudio/models/discriminators/conv_tasnet_discriminator/__init__.py

@@ -0,0 +1,6 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+
+
+if __name__ == "__main__":
+    pass

toolbox/torchaudio/models/discriminators/conv_tasnet_discriminator/configuration_conv_tasnet_discriminator.py

@@ -0,0 +1,34 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+from toolbox.torchaudio.configuration_utils import PretrainedConfig
+
+
+class ConvTasNetDiscriminatorConfig(PretrainedConfig):
+    """
+    https://github.com/yxlu-0102/MP-SENet/blob/main/config.json
+    """
+    def __init__(self,
+                 sample_rate: int = 8000,
+                 segment_size: int = 16000,
+                 n_fft: int = 512,
+                 win_length: int = 200,
+                 hop_length: int = 80,
+
+                 discriminator_dim: int = 16,
+                 discriminator_in_channel: int = 2,
+
+                 **kwargs
+                 ):
+        super(ConvTasNetDiscriminatorConfig, self).__init__(**kwargs)
+        self.sample_rate = sample_rate
+        self.segment_size = segment_size
+        self.n_fft = n_fft
+        self.win_length = win_length
+        self.hop_length = hop_length
+
+        self.discriminator_dim = discriminator_dim
+        self.discriminator_in_channel = discriminator_in_channel
+
+
+if __name__ == "__main__":
+    pass

toolbox/torchaudio/models/discriminators/conv_tasnet_discriminator/modeling_conv_tasnet_discriminator.py

@@ -0,0 +1,145 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+import os
+from typing import Optional, Union
+
+import torch
+import torch.nn as nn
+import torchaudio
+
+from toolbox.torchaudio.configuration_utils import DISCRIMINATOR_CONFIG_FILE
+from toolbox.torchaudio.models.discriminators.conv_tasnet_discriminator.configuration_conv_tasnet_discriminator import ConvTasNetDiscriminatorConfig
+
+
+class LearnableSigmoid1d(nn.Module):
+    def __init__(self, in_features, beta=1):
+        super().__init__()
+        self.beta = beta
+        self.slope = nn.Parameter(torch.ones(in_features))
+        self.slope.requiresGrad = True
+
+    def forward(self, x):
+        # x shape: [batch_size, time_steps, spec_bins]
+        return self.beta * torch.sigmoid(self.slope * x)
+
+
+class ConvTasNetDiscriminator(nn.Module):
+    def __init__(self, config: ConvTasNetDiscriminatorConfig):
+        super(ConvTasNetDiscriminator, self).__init__()
+        dim = config.discriminator_dim
+        self.in_channel = config.discriminator_in_channel
+
+        self.n_fft = config.n_fft
+        self.win_length = config.win_length
+        self.hop_length = config.hop_length
+
+        self.transform = torchaudio.transforms.Spectrogram(
+            n_fft=self.n_fft,
+            win_length=self.win_length,
+            hop_length=self.hop_length,
+            power=1.0,
+            window_fn=torch.hann_window,
+            # window_fn=torch.hamming_window if window_fn == "hamming" else torch.hann_window,
+        )
+
+        self.layers = nn.Sequential(
+            nn.utils.spectral_norm(nn.Conv2d(self.in_channel, dim, (4,4), (2,2), (1,1), bias=False)),
+            nn.InstanceNorm2d(dim, affine=True),
+            nn.PReLU(dim),
+            nn.utils.spectral_norm(nn.Conv2d(dim, dim*2, (4,4), (2,2), (1,1), bias=False)),
+            nn.InstanceNorm2d(dim*2, affine=True),
+            nn.PReLU(dim*2),
+            nn.utils.spectral_norm(nn.Conv2d(dim*2, dim*4, (4,4), (2,2), (1,1), bias=False)),
+            nn.InstanceNorm2d(dim*4, affine=True),
+            nn.PReLU(dim*4),
+            nn.utils.spectral_norm(nn.Conv2d(dim*4, dim*8, (4,4), (2,2), (1,1), bias=False)),
+            nn.InstanceNorm2d(dim*8, affine=True),
+            nn.PReLU(dim*8),
+            nn.AdaptiveMaxPool2d(1),
+            nn.Flatten(),
+            nn.utils.spectral_norm(nn.Linear(dim*8, dim*4)),
+            nn.Dropout(0.3),
+            nn.PReLU(dim*4),
+            nn.utils.spectral_norm(nn.Linear(dim*4, 1)),
+            LearnableSigmoid1d(1)
+        )
+
+    def forward(self, denoise_audios, clean_audios):
+        x = denoise_audios
+        y = clean_audios
+        x = self.transform.forward(x)
+        y = self.transform.forward(y)
+
+        xy = torch.stack((x, y), dim=1)
+        return self.layers(xy)
+
+
+MODEL_FILE = "discriminator.pt"
+
+
+class ConvTasNetDiscriminatorPretrainedModel(ConvTasNetDiscriminator):
+    def __init__(self,
+                 config: ConvTasNetDiscriminatorConfig,
+                 ):
+        super(ConvTasNetDiscriminatorPretrainedModel, self).__init__(
+            config=config,
+        )
+        self.config = config
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        config = ConvTasNetDiscriminatorConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+
+        model = cls(config)
+
+        if os.path.isdir(pretrained_model_name_or_path):
+            ckpt_file = os.path.join(pretrained_model_name_or_path, MODEL_FILE)
+        else:
+            ckpt_file = pretrained_model_name_or_path
+
+        with open(ckpt_file, "rb") as f:
+            state_dict = torch.load(f, map_location="cpu", weights_only=True)
+        model.load_state_dict(state_dict, strict=True)
+        return model
+
+    def save_pretrained(self,
+                        save_directory: Union[str, os.PathLike],
+                        state_dict: Optional[dict] = None,
+                        ):
+
+        model = self
+
+        if state_dict is None:
+            state_dict = model.state_dict()
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        # save state dict
+        model_file = os.path.join(save_directory, MODEL_FILE)
+        torch.save(state_dict, model_file)
+
+        # save config
+        config_file = os.path.join(save_directory, CONFIG_FILE)
+        self.config.to_yaml_file(config_file)
+        return save_directory
+
+
+def main():
+    config = ConvTasNetDiscriminatorConfig()
+    discriminator = ConvTasNetDiscriminator(config=config)
+
+    # shape: [batch_size, num_samples]
+    # x = torch.ones([4, int(4.5 * 16000)])
+    # y = torch.ones([4, int(4.5 * 16000)])
+    x = torch.ones([4, 16000])
+    y = torch.ones([4, 16000])
+
+    output = discriminator.forward(x, y)
+    print(output.shape)
+    print(output)
+
+    return
+
+
+if __name__ == "__main__":
+    main()