o oi6@spddlZddlZddlZddlZddlZddlmZmZmZm Z ddl Z ddl m Z ddl m Z mZddlmZddlmZmZddlmZddlmZdd lmZmZmZdd lmZdd l m!Z!dd l"m#Z#dd l$m%Z%m&Z&m'Z'm(Z(m)Z)ddl*m+Z+ddl,m-Z-m.Z.m/Z/m0Z0dZ1dZ2GdddeZ3ddZ4dee5de5fddZ6dddddde2dfdee5d e7d!e5d"ee5d#e7d$e e5e8dfd%e5d&e7deej9e-e5ffd'd(Z:dAd)e d*e-d+e8dee e e ffd,d-Z;e < dBd.ej9d/e-d)e d0ee=fd1d2Z>e2fd3e5de5fd4d5Z?d6e5de e8e5ffd7d8Z@Gd9d:d:ejAZB dCd;e5dejCfdd?ZEeFd@kr6eEdSdS)DN)ListOptionalTupleUnion)logger)Tensornn) functional) DataLoaderDataset) load_model)config) load_audioresample save_audio) init_logger) ModelParams) get_device) as_complexas_real download_file get_cache_dirget_norm_alpha)version)DFerberb_norm unit_norm) DeepFilterNetDeepFilterNet2DeepFilterNet3r csNeZdZdeededdffdd ZdeeeeffddZ d d Z Z S) AudioDatasetfilessrreturnNcsNtg|_|D]}tj|std|d|j|q ||_ dS)NzFile not found: z . Skipping...) super__init__r"ospathisfilerwarningappendr#)selfr"r#file __class__H/home/ubuntu/SoloSpeech/.venv/lib/python3.10/site-packages/df/enhance.pyr&s   zAudioDataset.__init__cCs(|j|}t||jd\}}|||jfS)Ncpu)r"rr# sample_rate)r,indexfnaudiometar0r0r1 __getitem__&s  zAudioDataset.__getitem__cCs t|jSN)lenr")r,r0r0r1__len__+s zAudioDataset.__len__) __name__ __module__ __qualname__rstrintr&rrr8r; __classcell__r0r0r.r1r!s r!c Cst|j|j|jd|j|jd\}}}|jr|nd}|jdur"d|_n tj |js/t |jt j }|jdurQt|jdkrHtdtdt|jd}nt|jdks\Jd|j}t||}t|d dd }t|}t|D]s\} \} } } | d} | d} | d|d } t}t||| |j|jd } t}| jd |}||}||}tj | }|dkr| ddnd}t|d|d|dd|ddt | !d|| } t"| | | |j|ddqsdS)NT) post_filter log_levelconfig_allow_defaultsepoch mask_only.rzGOnly one of `noisy_audio_files` or `noisy_dir` arguments are supported.z/*zNo audio files provided) num_workers pin_memoryd)pad atten_lim_dbz2.0fz% | zEnhanced noisy audio file 'z' in z.2fzs (RT factor: z.3f)r2F)r# output_dirsuffixlog)#init_dfmodel_base_dirpfrCrE no_df_stagerSrRr'r(isdirmkdirrr# noisy_dirr:noisy_audio_filesrerrorexitglobr!r enumeratesqueezetimeenhancecompensate_delay atten_limshapebasenameinfortor)argsmodeldf_staterSdf_sr input_filesdsloader n_samplesir-r6audio_srprogresst0t1t_audiotrtfr5p_strr0r0r1main/sT        &r{mr$cCs$|durt}|tv}|rt|S|Sr9) DEFAULT_MODELPRETRAINED_MODELSmaybe_download_model)r|is_default_modelr0r0r1get_model_basedir\s rFINFOz enhance.logTbestrVrBrClog_filerDrE default_modelrFcCs<zddlm}m} |jdd| Wn tyYnw|dup$|tv} t|p)|}tj |s8t d ||durCtj ||nd}t |||d| rYtd|d |tjtj |d d|dd |rtd dttjztd ttj} d| d} Wn tyd} Ynwtd| t} t| j| j| j| j| jd} tj |d}|duot|t o|!dk }|sd}|ptdtdddd}t"|| ||d\}}|dus|dkr|rt#dt$dt%d||&t'}tj(tj)|}|r |d7}td t'td|| |fS) aInitializes and loads config, model and deep filtering state. Args: model_base_dir (str): Path to the model directory containing checkpoint and config. If None, load the pretrained DeepFilterNet2 model. post_filter (bool): Enable post filter for some minor, extra noise reduction. log_level (str): Control amount of logging. Defaults to `INFO`. log_file (str): Optional log file name. None disables it. Defaults to `enhance.log`. config_allow_defaults (bool): Whether to allow initializing new config values with defaults. epoch (str): Checkpoint epoch to load. Options are `best`, `latest`, ``, and `none`. `none` disables checkpoint loading. Defaults to `best`. Returns: model (nn.Modules): Intialized model, moved to GPU if available. df_state (DF): Deep filtering state for stft/istft/erb suffix (str): Suffix based on the model name. This can be used for saving the enhanced audio. r)icinstallT)includeContextNzBase directory not found at {})r-levelrkzUsing z model at config.ini)config_must_existallow_defaults allow_reloadmask_pfpf_betaz(beta: rQrPzRunning with post-filter )r#fft_sizehop_sizenb_bandsmin_nb_erb_freqs checkpointsnonerFtrainF)castsectiondefaultsave)rErFzCould not find a checkpointrHzLoaded checkpoint from epoch _pfzRunning on device {}z Model loaded)*icecreamrrconfigureOutput ImportErrorr~rr'r(rYNotADirectoryErrorformatjoinrrrhr loadsetboolrrgetfloatKeyErrorrr#rrnb_erb min_nb_freqs isinstancer?lower load_model_cpr]r^debugrirrgabspath)rVrBrCrrDrErrFrruse_default_modelbetaprlcheckpoint_dirload_cprkrSr0r0r1rUesr             rUr6dfnb_dfc Cs||}td}|}ttdtt t t ||| d}Wdn1s2wYt t t|dd|f| d}t t | d}|durg||}||}||}|||fS)NFignorerH.)analysisnumpyr erb_widthswarningscatch_warnings simplefilter UserWarningtorch as_tensorrr unsqueezerrri) r6rrdevicespecaerb_fberb_feat spec_featr0r0r1 df_featuress  &    rrkrlrNcCsN||jd}t|dr|j|td|jd}d\}}|r2||}}t|d|f}t |dt |dt j } t ||| td\} } } || | | d} t| d } |d urt|dkrd t| d }t| d || d |} t|| }|r||dksJ||}|d d |||f}|S) a Enhance a single audio given a preloaded model and DF state. Args: model (nn.Module): A DeepFilterNet model. df_state (DF): DF state for STFT/ISTFT and feature calculation. audio (Tensor): Time domain audio of shape [C, T]. Sampling rate needs to match to `model` and `df_state`. pad (bool): Pad the audio to compensate for delay due to STFT/ISTFT. atten_lim_db (float): An optional noise attenuation limit in dB. E.g. an attenuation limit of 12 dB only suppresses 12 dB and keeps the remaining noise in the resulting audio. Returns: enhanced audio (Tensor): If `pad` was `False` of shape [C, T'] where T'r&rrAr0r0r.r1rs  rdefault_log_levelcCs|durt}|jddtddd|jdddd|jd d tdd d|jd t|d d|jdddddd|jdddtdd|jddtd|S)Nz--model-base-dirz-mzModel directory containing checkpoints and config. To load a pretrained model, you may just provide the model name, e.g. `DeepFilterNet`. By default, the pretrained DeepFilterNet2 model is loaded.typerrz--pfz>Post-filter that slightly over-attenuates very noisy sections. store_true)ractionz --output-dirz-oz;Directory in which the enhanced audio files will be stored.z --log-levelz:Logger verbosity. Can be one of (debug, info, error, none)z--debugz-d store_constDEBUGrC)rconstrz--epochz-erzFEpoch for checkpoint loading. Can be one of ['best', 'latest', ].)rrrz-vz --versionr)argparseArgumentParser add_argumentr?rr)rparserr0r0r1setup_df_argument_parser*sJ rcCst}|jddddd|jddtddd |jd td d d |jddtddd |jddddd|jddd|}t|dS)Nz--no-delay-compensationrd store_falsezeDont't add some paddig to compensate the delay introduced by the real-time STFT/ISTFT implementation.)rrrz --atten-limz-azLAttenuation limit in dB by mixing the enhanced signal with the noisy signal.rr\*z6List of noise files to mix with the clean speech file.)rrrz --noisy-dirz-izQInput directory containing noisy audio files. Use instead of `noisy_audio_files`.z --no-suffixrSz6Don't add the model suffix to the enhanced audio files)rrrz --no-df-stagerr)rrr@r? parse_argsr{)rrjr0r0r1runVsH r__main__r9)TN)rN)Grr_r'rbrtypingrrrrrlogururrrtorch.nnr rtorch.utils.datar r df.checkpointr r df.configr df.iorrr df.loggerrdf.modelr df.modulesrdf.utilsrrrrr df.versionrlibdfrrrrr~r}r!r{r?rrr@ModulerUrno_gradrrcrrActionrrrrr<r0r0r0r1s        -   &X . , '