Upload 41 files

lib_v5/mdxnet.py

@@ -0,0 +1,140 @@
+from abc import ABCMeta
+
+import torch
+import torch.nn as nn
+from pytorch_lightning import LightningModule
+from .modules import TFC_TDF
+
+dim_s = 4
+
+class AbstractMDXNet(LightningModule):
+    __metaclass__ = ABCMeta
+
+    def __init__(self, target_name, lr, optimizer, dim_c, dim_f, dim_t, n_fft, hop_length, overlap):
+        super().__init__()
+        self.target_name = target_name
+        self.lr = lr
+        self.optimizer = optimizer
+        self.dim_c = dim_c
+        self.dim_f = dim_f
+        self.dim_t = dim_t
+        self.n_fft = n_fft
+        self.n_bins = n_fft // 2 + 1
+        self.hop_length = hop_length
+        self.window = nn.Parameter(torch.hann_window(window_length=self.n_fft, periodic=True), requires_grad=False)
+        self.freq_pad = nn.Parameter(torch.zeros([1, dim_c, self.n_bins - self.dim_f, self.dim_t]), requires_grad=False)
+
+    def configure_optimizers(self):
+        if self.optimizer == 'rmsprop':
+            return torch.optim.RMSprop(self.parameters(), self.lr)
+        
+        if self.optimizer == 'adamw':
+            return torch.optim.AdamW(self.parameters(), self.lr)
+
+class ConvTDFNet(AbstractMDXNet):
+    def __init__(self, target_name, lr, optimizer, dim_c, dim_f, dim_t, n_fft, hop_length,
+                 num_blocks, l, g, k, bn, bias, overlap):
+
+        super(ConvTDFNet, self).__init__(
+            target_name, lr, optimizer, dim_c, dim_f, dim_t, n_fft, hop_length, overlap)
+        self.save_hyperparameters()
+
+        self.num_blocks = num_blocks
+        self.l = l
+        self.g = g
+        self.k = k
+        self.bn = bn
+        self.bias = bias
+
+        if optimizer == 'rmsprop':
+            norm = nn.BatchNorm2d
+            
+        if optimizer == 'adamw':
+            norm = lambda input:nn.GroupNorm(2, input)
+            
+        self.n = num_blocks // 2
+        scale = (2, 2)
+
+        self.first_conv = nn.Sequential(
+            nn.Conv2d(in_channels=self.dim_c, out_channels=g, kernel_size=(1, 1)),
+            norm(g),
+            nn.ReLU(),
+        )
+
+        f = self.dim_f
+        c = g
+        self.encoding_blocks = nn.ModuleList()
+        self.ds = nn.ModuleList()
+        for i in range(self.n):
+            self.encoding_blocks.append(TFC_TDF(c, l, f, k, bn, bias=bias, norm=norm))
+            self.ds.append(
+                nn.Sequential(
+                    nn.Conv2d(in_channels=c, out_channels=c + g, kernel_size=scale, stride=scale),
+                    norm(c + g),
+                    nn.ReLU()
+                )
+            )
+            f = f // 2
+            c += g
+
+        self.bottleneck_block = TFC_TDF(c, l, f, k, bn, bias=bias, norm=norm)
+
+        self.decoding_blocks = nn.ModuleList()
+        self.us = nn.ModuleList()
+        for i in range(self.n):
+            self.us.append(
+                nn.Sequential(
+                    nn.ConvTranspose2d(in_channels=c, out_channels=c - g, kernel_size=scale, stride=scale),
+                    norm(c - g),
+                    nn.ReLU()
+                )
+            )
+            f = f * 2
+            c -= g
+
+            self.decoding_blocks.append(TFC_TDF(c, l, f, k, bn, bias=bias, norm=norm))
+
+        self.final_conv = nn.Sequential(
+            nn.Conv2d(in_channels=c, out_channels=self.dim_c, kernel_size=(1, 1)),
+        )
+
+    def forward(self, x):
+
+        x = self.first_conv(x)
+
+        x = x.transpose(-1, -2)
+
+        ds_outputs = []
+        for i in range(self.n):
+            x = self.encoding_blocks[i](x)
+            ds_outputs.append(x)
+            x = self.ds[i](x)
+
+        x = self.bottleneck_block(x)
+
+        for i in range(self.n):
+            x = self.us[i](x)
+            x *= ds_outputs[-i - 1]
+            x = self.decoding_blocks[i](x)
+
+        x = x.transpose(-1, -2)
+
+        x = self.final_conv(x)
+
+        return x
+    
+class Mixer(nn.Module):
+    def __init__(self, device, mixer_path):
+        
+        super(Mixer, self).__init__()
+        
+        self.linear = nn.Linear((dim_s+1)*2, dim_s*2, bias=False)
+        
+        self.load_state_dict(
+            torch.load(mixer_path, map_location=device)
+        )
+
+    def forward(self, x):
+        x = x.reshape(1,(dim_s+1)*2,-1).transpose(-1,-2)
+        x = self.linear(x)
+        return x.transpose(-1,-2).reshape(dim_s,2,-1)

lib_v5/mixer.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:946e03c789160ae4631f7037e54b5de90c32fe2c302fc2e5022696bde6902300
+size 129

lib_v5/modules.py

@@ -0,0 +1,74 @@
+import torch
+import torch.nn as nn
+
+
+class TFC(nn.Module):
+    def __init__(self, c, l, k, norm):
+        super(TFC, self).__init__()
+
+        self.H = nn.ModuleList()
+        for i in range(l):
+            self.H.append(
+                nn.Sequential(
+                    nn.Conv2d(in_channels=c, out_channels=c, kernel_size=k, stride=1, padding=k // 2),
+                    norm(c),
+                    nn.ReLU(),
+                )
+            )
+
+    def forward(self, x):
+        for h in self.H:
+            x = h(x)
+        return x
+
+
+class DenseTFC(nn.Module):
+    def __init__(self, c, l, k, norm):
+        super(DenseTFC, self).__init__()
+
+        self.conv = nn.ModuleList()
+        for i in range(l):
+            self.conv.append(
+                nn.Sequential(
+                    nn.Conv2d(in_channels=c, out_channels=c, kernel_size=k, stride=1, padding=k // 2),
+                    norm(c),
+                    nn.ReLU(),
+                )
+            )
+
+    def forward(self, x):
+        for layer in self.conv[:-1]:
+            x = torch.cat([layer(x), x], 1)
+        return self.conv[-1](x)
+
+
+class TFC_TDF(nn.Module):
+    def __init__(self, c, l, f, k, bn, dense=False, bias=True, norm=nn.BatchNorm2d):
+
+        super(TFC_TDF, self).__init__()
+
+        self.use_tdf = bn is not None
+
+        self.tfc = DenseTFC(c, l, k, norm) if dense else TFC(c, l, k, norm)
+
+        if self.use_tdf:
+            if bn == 0:
+                self.tdf = nn.Sequential(
+                    nn.Linear(f, f, bias=bias),
+                    norm(c),
+                    nn.ReLU()
+                )
+            else:
+                self.tdf = nn.Sequential(
+                    nn.Linear(f, f // bn, bias=bias),
+                    norm(c),
+                    nn.ReLU(),
+                    nn.Linear(f // bn, f, bias=bias),
+                    norm(c),
+                    nn.ReLU()
+                )
+
+    def forward(self, x):
+        x = self.tfc(x)
+        return x + self.tdf(x) if self.use_tdf else x
+

lib_v5/pyrb.py

@@ -0,0 +1,92 @@
+import os
+import subprocess
+import tempfile
+import six
+import numpy as np
+import soundfile as sf
+import sys
+
+if getattr(sys, 'frozen', False):
+    BASE_PATH_RUB = sys._MEIPASS
+else:
+    BASE_PATH_RUB = os.path.dirname(os.path.abspath(__file__))
+
+__all__ = ['time_stretch', 'pitch_shift']
+
+__RUBBERBAND_UTIL = os.path.join(BASE_PATH_RUB, 'rubberband')
+
+if six.PY2:
+    DEVNULL = open(os.devnull, 'w')
+else:
+    DEVNULL = subprocess.DEVNULL
+
+def __rubberband(y, sr, **kwargs):
+
+    assert sr > 0
+
+    # Get the input and output tempfile
+    fd, infile = tempfile.mkstemp(suffix='.wav')
+    os.close(fd)
+    fd, outfile = tempfile.mkstemp(suffix='.wav')
+    os.close(fd)
+
+    # dump the audio
+    sf.write(infile, y, sr)
+
+    try:
+        # Execute rubberband
+        arguments = [__RUBBERBAND_UTIL, '-q']
+
+        for key, value in six.iteritems(kwargs):
+            arguments.append(str(key))
+            arguments.append(str(value))
+
+        arguments.extend([infile, outfile])
+
+        subprocess.check_call(arguments, stdout=DEVNULL, stderr=DEVNULL)
+
+        # Load the processed audio.
+        y_out, _ = sf.read(outfile, always_2d=True)
+
+        # make sure that output dimensions matches input
+        if y.ndim == 1:
+            y_out = np.squeeze(y_out)
+
+    except OSError as exc:
+        six.raise_from(RuntimeError('Failed to execute rubberband. '
+                                    'Please verify that rubberband-cli '
+                                    'is installed.'),
+                       exc)
+
+    finally:
+        # Remove temp files
+        os.unlink(infile)
+        os.unlink(outfile)
+
+    return y_out
+
+def time_stretch(y, sr, rate, rbargs=None):
+    if rate <= 0:
+        raise ValueError('rate must be strictly positive')
+
+    if rate == 1.0:
+        return y
+
+    if rbargs is None:
+        rbargs = dict()
+
+    rbargs.setdefault('--tempo', rate)
+
+    return __rubberband(y, sr, **rbargs)
+
+def pitch_shift(y, sr, n_steps, rbargs=None):
+
+    if n_steps == 0:
+        return y
+
+    if rbargs is None:
+        rbargs = dict()
+
+    rbargs.setdefault('--pitch', n_steps)
+
+    return __rubberband(y, sr, **rbargs)

lib_v5/spec_utils.py

@@ -0,0 +1,692 @@
+import librosa
+import numpy as np
+import soundfile as sf
+import math
+import random
+import math
+import platform
+import traceback
+from . import pyrb
+#cur
+OPERATING_SYSTEM = platform.system()
+SYSTEM_ARCH = platform.platform()
+SYSTEM_PROC = platform.processor()
+ARM = 'arm'
+
+if OPERATING_SYSTEM == 'Windows':
+    from pyrubberband import pyrb
+else:
+    from . import pyrb
+
+if OPERATING_SYSTEM == 'Darwin':
+    wav_resolution = "polyphase" if SYSTEM_PROC == ARM or ARM in SYSTEM_ARCH else "sinc_fastest" 
+else:
+    wav_resolution = "sinc_fastest"
+
+MAX_SPEC = 'Max Spec'
+MIN_SPEC = 'Min Spec'
+AVERAGE = 'Average'
+
+def crop_center(h1, h2):
+    h1_shape = h1.size()
+    h2_shape = h2.size()
+
+    if h1_shape[3] == h2_shape[3]:
+        return h1
+    elif h1_shape[3] < h2_shape[3]:
+        raise ValueError('h1_shape[3] must be greater than h2_shape[3]')
+
+    s_time = (h1_shape[3] - h2_shape[3]) // 2
+    e_time = s_time + h2_shape[3]
+    h1 = h1[:, :, :, s_time:e_time]
+
+    return h1
+
+def preprocess(X_spec):
+    X_mag = np.abs(X_spec)
+    X_phase = np.angle(X_spec)
+
+    return X_mag, X_phase
+
+def make_padding(width, cropsize, offset):
+    left = offset
+    roi_size = cropsize - offset * 2
+    if roi_size == 0:
+        roi_size = cropsize
+    right = roi_size - (width % roi_size) + left
+
+    return left, right, roi_size
+
+def wave_to_spectrogram(wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False):
+    if reverse:
+        wave_left = np.flip(np.asfortranarray(wave[0]))
+        wave_right = np.flip(np.asfortranarray(wave[1]))
+    elif mid_side:
+        wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
+    elif mid_side_b2:
+        wave_left = np.asfortranarray(np.add(wave[1], wave[0] * .5))
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * .5))
+    else:
+        wave_left = np.asfortranarray(wave[0])
+        wave_right = np.asfortranarray(wave[1])
+
+    spec_left = librosa.stft(wave_left, n_fft, hop_length=hop_length)
+    spec_right = librosa.stft(wave_right, n_fft, hop_length=hop_length)
+    
+    spec = np.asfortranarray([spec_left, spec_right])
+
+    return spec
+   
+def wave_to_spectrogram_mt(wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False):
+    import threading
+
+    if reverse:
+        wave_left = np.flip(np.asfortranarray(wave[0]))
+        wave_right = np.flip(np.asfortranarray(wave[1]))
+    elif mid_side:
+        wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
+    elif mid_side_b2:
+        wave_left = np.asfortranarray(np.add(wave[1], wave[0] * .5))
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * .5))
+    else:
+        wave_left = np.asfortranarray(wave[0])
+        wave_right = np.asfortranarray(wave[1])
+   
+    def run_thread(**kwargs):
+        global spec_left
+        spec_left = librosa.stft(**kwargs)
+
+    thread = threading.Thread(target=run_thread, kwargs={'y': wave_left, 'n_fft': n_fft, 'hop_length': hop_length})
+    thread.start()
+    spec_right = librosa.stft(wave_right, n_fft, hop_length=hop_length)
+    thread.join()   
+    
+    spec = np.asfortranarray([spec_left, spec_right])
+
+    return spec
+    
+def normalize(wave, is_normalize=False):
+    """Save output music files"""
+    maxv = np.abs(wave).max()
+    if maxv > 1.0:
+        print(f"\nNormalization Set {is_normalize}: Input above threshold for clipping. Max:{maxv}")
+        if is_normalize:
+            print(f"The result was normalized.")
+            wave /= maxv
+        else:
+            print(f"The result was not normalized.")
+    else:
+        print(f"\nNormalization Set {is_normalize}: Input not above threshold for clipping. Max:{maxv}")
+    
+    return wave
+    
+def normalize_two_stem(wave, mix, is_normalize=False):
+    """Save output music files"""
+    
+    maxv = np.abs(wave).max()
+    max_mix = np.abs(mix).max()
+    
+    if maxv > 1.0:
+        print(f"\nNormalization Set {is_normalize}: Primary source above threshold for clipping. Max:{maxv}")
+        print(f"\nNormalization Set {is_normalize}: Mixture above threshold for clipping. Max:{max_mix}")
+        if is_normalize:
+            print(f"The result was normalized.")
+            wave /= maxv
+            mix /= maxv
+        else:
+            print(f"The result was not normalized.")
+    else:
+        print(f"\nNormalization Set {is_normalize}: Input not above threshold for clipping. Max:{maxv}")
+    
+    
+    print(f"\nNormalization Set {is_normalize}: Primary source - Max:{np.abs(wave).max()}")
+    print(f"\nNormalization Set {is_normalize}: Mixture - Max:{np.abs(mix).max()}")
+    
+    return wave, mix    
+
+def combine_spectrograms(specs, mp):
+    l = min([specs[i].shape[2] for i in specs])    
+    spec_c = np.zeros(shape=(2, mp.param['bins'] + 1, l), dtype=np.complex64)
+    offset = 0
+    bands_n = len(mp.param['band'])
+    
+    for d in range(1, bands_n + 1):
+        h = mp.param['band'][d]['crop_stop'] - mp.param['band'][d]['crop_start']
+        spec_c[:, offset:offset+h, :l] = specs[d][:, mp.param['band'][d]['crop_start']:mp.param['band'][d]['crop_stop'], :l]
+        offset += h
+        
+    if offset > mp.param['bins']:
+        raise ValueError('Too much bins')
+        
+    # lowpass fiter
+    if mp.param['pre_filter_start'] > 0: # and mp.param['band'][bands_n]['res_type'] in ['scipy', 'polyphase']:   
+        if bands_n == 1:
+            spec_c = fft_lp_filter(spec_c, mp.param['pre_filter_start'], mp.param['pre_filter_stop'])
+        else:
+            gp = 1        
+            for b in range(mp.param['pre_filter_start'] + 1, mp.param['pre_filter_stop']):
+                g = math.pow(10, -(b - mp.param['pre_filter_start']) * (3.5 - gp) / 20.0)
+                gp = g
+                spec_c[:, b, :] *= g
+                
+    return np.asfortranarray(spec_c)
+    
+def spectrogram_to_image(spec, mode='magnitude'):
+    if mode == 'magnitude':
+        if np.iscomplexobj(spec):
+            y = np.abs(spec)
+        else:
+            y = spec
+        y = np.log10(y ** 2 + 1e-8)
+    elif mode == 'phase':
+        if np.iscomplexobj(spec):
+            y = np.angle(spec)
+        else:
+            y = spec
+
+    y -= y.min()
+    y *= 255 / y.max()
+    img = np.uint8(y)
+
+    if y.ndim == 3:
+        img = img.transpose(1, 2, 0)
+        img = np.concatenate([
+            np.max(img, axis=2, keepdims=True), img
+        ], axis=2)
+
+    return img
+
+def reduce_vocal_aggressively(X, y, softmask):
+    v = X - y
+    y_mag_tmp = np.abs(y)
+    v_mag_tmp = np.abs(v)
+
+    v_mask = v_mag_tmp > y_mag_tmp
+    y_mag = np.clip(y_mag_tmp - v_mag_tmp * v_mask * softmask, 0, np.inf)
+
+    return y_mag * np.exp(1.j * np.angle(y))
+
+def merge_artifacts(y_mask, thres=0.01, min_range=64, fade_size=32):
+    mask = y_mask
+    
+    try:
+        if min_range < fade_size * 2:
+            raise ValueError('min_range must be >= fade_size * 2')
+
+        idx = np.where(y_mask.min(axis=(0, 1)) > thres)[0]
+        start_idx = np.insert(idx[np.where(np.diff(idx) != 1)[0] + 1], 0, idx[0])
+        end_idx = np.append(idx[np.where(np.diff(idx) != 1)[0]], idx[-1])
+        artifact_idx = np.where(end_idx - start_idx > min_range)[0]
+        weight = np.zeros_like(y_mask)
+        if len(artifact_idx) > 0:
+            start_idx = start_idx[artifact_idx]
+            end_idx = end_idx[artifact_idx]
+            old_e = None
+            for s, e in zip(start_idx, end_idx):
+                if old_e is not None and s - old_e < fade_size:
+                    s = old_e - fade_size * 2
+
+                if s != 0:
+                    weight[:, :, s:s + fade_size] = np.linspace(0, 1, fade_size)
+                else:
+                    s -= fade_size
+
+                if e != y_mask.shape[2]:
+                    weight[:, :, e - fade_size:e] = np.linspace(1, 0, fade_size)
+                else:
+                    e += fade_size
+
+                weight[:, :, s + fade_size:e - fade_size] = 1
+                old_e = e
+
+        v_mask = 1 - y_mask
+        y_mask += weight * v_mask
+        
+        mask = y_mask
+    except Exception as e:
+        error_name = f'{type(e).__name__}'
+        traceback_text = ''.join(traceback.format_tb(e.__traceback__))
+        message = f'{error_name}: "{e}"\n{traceback_text}"'
+        print('Post Process Failed: ', message)
+        
+
+    return mask
+
+def align_wave_head_and_tail(a, b):
+    l = min([a[0].size, b[0].size])  
+    
+    return a[:l,:l], b[:l,:l]
+    
+def spectrogram_to_wave(spec, hop_length, mid_side, mid_side_b2, reverse, clamp=False):
+    spec_left = np.asfortranarray(spec[0])
+    spec_right = np.asfortranarray(spec[1])
+
+    wave_left = librosa.istft(spec_left, hop_length=hop_length)
+    wave_right = librosa.istft(spec_right, hop_length=hop_length)
+
+    if reverse:
+        return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
+    elif mid_side:
+        return np.asfortranarray([np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)])
+    elif mid_side_b2:
+        return np.asfortranarray([np.add(wave_right / 1.25, .4 * wave_left), np.subtract(wave_left / 1.25, .4 * wave_right)])
+    else:
+        return np.asfortranarray([wave_left, wave_right])
+    
+def spectrogram_to_wave_mt(spec, hop_length, mid_side, reverse, mid_side_b2):
+    import threading
+
+    spec_left = np.asfortranarray(spec[0])
+    spec_right = np.asfortranarray(spec[1])
+    
+    def run_thread(**kwargs):
+        global wave_left
+        wave_left = librosa.istft(**kwargs)
+        
+    thread = threading.Thread(target=run_thread, kwargs={'stft_matrix': spec_left, 'hop_length': hop_length})
+    thread.start()
+    wave_right = librosa.istft(spec_right, hop_length=hop_length)
+    thread.join()   
+    
+    if reverse:
+        return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
+    elif mid_side:
+        return np.asfortranarray([np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)])
+    elif mid_side_b2:
+        return np.asfortranarray([np.add(wave_right / 1.25, .4 * wave_left), np.subtract(wave_left / 1.25, .4 * wave_right)])
+    else:
+        return np.asfortranarray([wave_left, wave_right])
+    
+def cmb_spectrogram_to_wave(spec_m, mp, extra_bins_h=None, extra_bins=None):
+    bands_n = len(mp.param['band'])    
+    offset = 0
+
+    for d in range(1, bands_n + 1):
+        bp = mp.param['band'][d]
+        spec_s = np.ndarray(shape=(2, bp['n_fft'] // 2 + 1, spec_m.shape[2]), dtype=complex)
+        h = bp['crop_stop'] - bp['crop_start']
+        spec_s[:, bp['crop_start']:bp['crop_stop'], :] = spec_m[:, offset:offset+h, :]
+        
+        offset += h
+        if d == bands_n: # higher
+            if extra_bins_h: # if --high_end_process bypass
+                max_bin = bp['n_fft'] // 2
+                spec_s[:, max_bin-extra_bins_h:max_bin, :] = extra_bins[:, :extra_bins_h, :]
+            if bp['hpf_start'] > 0:
+                spec_s = fft_hp_filter(spec_s, bp['hpf_start'], bp['hpf_stop'] - 1)
+            if bands_n == 1:
+                wave = spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse'])
+            else:
+                wave = np.add(wave, spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse']))
+        else:
+            sr = mp.param['band'][d+1]['sr']
+            if d == 1: # lower
+                spec_s = fft_lp_filter(spec_s, bp['lpf_start'], bp['lpf_stop'])
+                wave = librosa.resample(spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse']), bp['sr'], sr, res_type=wav_resolution)
+            else: # mid
+                spec_s = fft_hp_filter(spec_s, bp['hpf_start'], bp['hpf_stop'] - 1)
+                spec_s = fft_lp_filter(spec_s, bp['lpf_start'], bp['lpf_stop'])
+                wave2 = np.add(wave, spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse']))
+                wave = librosa.resample(wave2, bp['sr'], sr, res_type=wav_resolution)
+        
+    return wave
+
+def fft_lp_filter(spec, bin_start, bin_stop):
+    g = 1.0
+    for b in range(bin_start, bin_stop):
+        g -= 1 / (bin_stop - bin_start)
+        spec[:, b, :] = g * spec[:, b, :]
+        
+    spec[:, bin_stop:, :] *= 0
+
+    return spec
+
+def fft_hp_filter(spec, bin_start, bin_stop):
+    g = 1.0
+    for b in range(bin_start, bin_stop, -1):
+        g -= 1 / (bin_start - bin_stop)
+        spec[:, b, :] = g * spec[:, b, :]
+    
+    spec[:, 0:bin_stop+1, :] *= 0
+
+    return spec
+
+def mirroring(a, spec_m, input_high_end, mp):
+    if 'mirroring' == a:
+        mirror = np.flip(np.abs(spec_m[:, mp.param['pre_filter_start']-10-input_high_end.shape[1]:mp.param['pre_filter_start']-10, :]), 1)
+        mirror = mirror * np.exp(1.j * np.angle(input_high_end))
+        
+        return np.where(np.abs(input_high_end) <= np.abs(mirror), input_high_end, mirror)
+        
+    if 'mirroring2' == a:
+        mirror = np.flip(np.abs(spec_m[:, mp.param['pre_filter_start']-10-input_high_end.shape[1]:mp.param['pre_filter_start']-10, :]), 1)
+        mi = np.multiply(mirror, input_high_end * 1.7)
+        
+        return np.where(np.abs(input_high_end) <= np.abs(mi), input_high_end, mi)
+
+def adjust_aggr(mask, is_non_accom_stem, aggressiveness):
+    aggr = aggressiveness['value']
+
+    if aggr != 0:
+        if is_non_accom_stem:
+            aggr = 1 - aggr
+    
+        aggr = [aggr, aggr]
+    
+        if aggressiveness['aggr_correction'] is not None:
+            aggr[0] += aggressiveness['aggr_correction']['left']
+            aggr[1] += aggressiveness['aggr_correction']['right']
+
+        for ch in range(2):
+            mask[ch, :aggressiveness['split_bin']] = np.power(mask[ch, :aggressiveness['split_bin']], 1 + aggr[ch] / 3)
+            mask[ch, aggressiveness['split_bin']:] = np.power(mask[ch, aggressiveness['split_bin']:], 1 + aggr[ch])
+
+        # if is_non_accom_stem:
+        #     mask = (1.0 - mask)
+        
+    return mask
+
+def stft(wave, nfft, hl):
+    wave_left = np.asfortranarray(wave[0])
+    wave_right = np.asfortranarray(wave[1])
+    spec_left = librosa.stft(wave_left, nfft, hop_length=hl)
+    spec_right = librosa.stft(wave_right, nfft, hop_length=hl)
+    spec = np.asfortranarray([spec_left, spec_right])
+
+    return spec
+
+def istft(spec, hl):
+    spec_left = np.asfortranarray(spec[0])
+    spec_right = np.asfortranarray(spec[1])
+    wave_left = librosa.istft(spec_left, hop_length=hl)
+    wave_right = librosa.istft(spec_right, hop_length=hl)
+    wave = np.asfortranarray([wave_left, wave_right])
+
+    return wave
+
+def spec_effects(wave, algorithm='Default', value=None):
+    spec = [stft(wave[0],2048,1024), stft(wave[1],2048,1024)]
+    if algorithm == 'Min_Mag':
+        v_spec_m = np.where(np.abs(spec[1]) <= np.abs(spec[0]), spec[1], spec[0])
+        wave = istft(v_spec_m,1024)
+    elif algorithm == 'Max_Mag':
+        v_spec_m = np.where(np.abs(spec[1]) >= np.abs(spec[0]), spec[1], spec[0])
+        wave = istft(v_spec_m,1024)
+    elif algorithm == 'Default':
+        wave = (wave[1] * value) + (wave[0] * (1-value))
+    elif algorithm == 'Invert_p':
+        X_mag = np.abs(spec[0])
+        y_mag = np.abs(spec[1])            
+        max_mag = np.where(X_mag >= y_mag, X_mag, y_mag)  
+        v_spec = spec[1] - max_mag * np.exp(1.j * np.angle(spec[0]))
+        wave = istft(v_spec,1024)
+            
+    return wave      
+
+def spectrogram_to_wave_no_mp(spec, n_fft=2048, hop_length=1024):
+    wave = librosa.istft(spec, n_fft=n_fft, hop_length=hop_length)
+    
+    if wave.ndim == 1:
+        wave = np.asfortranarray([wave,wave])
+
+    return wave
+
+def wave_to_spectrogram_no_mp(wave):
+    
+    spec = librosa.stft(wave, n_fft=2048, hop_length=1024)
+    
+    if spec.ndim == 1:
+        spec = np.asfortranarray([spec,spec])
+
+    return spec
+
+def invert_audio(specs, invert_p=True):
+    
+    ln = min([specs[0].shape[2], specs[1].shape[2]])
+    specs[0] = specs[0][:,:,:ln]
+    specs[1] = specs[1][:,:,:ln]
+        
+    if invert_p:
+        X_mag = np.abs(specs[0])
+        y_mag = np.abs(specs[1])            
+        max_mag = np.where(X_mag >= y_mag, X_mag, y_mag)  
+        v_spec = specs[1] - max_mag * np.exp(1.j * np.angle(specs[0]))
+    else:
+        specs[1] = reduce_vocal_aggressively(specs[0], specs[1], 0.2)
+        v_spec = specs[0] - specs[1]
+
+    return v_spec
+
+def invert_stem(mixture, stem):
+    
+    mixture = wave_to_spectrogram_no_mp(mixture)
+    stem = wave_to_spectrogram_no_mp(stem)
+    output = spectrogram_to_wave_no_mp(invert_audio([mixture, stem]))
+
+    return -output.T
+
+def ensembling(a, specs):   
+    for i in range(1, len(specs)):
+        if i == 1:
+            spec = specs[0]
+
+        ln = min([spec.shape[2], specs[i].shape[2]])
+        spec = spec[:,:,:ln]
+        specs[i] = specs[i][:,:,:ln]
+        
+        if MIN_SPEC == a:
+            spec = np.where(np.abs(specs[i]) <= np.abs(spec), specs[i], spec)
+        if MAX_SPEC == a:
+            spec = np.where(np.abs(specs[i]) >= np.abs(spec), specs[i], spec)  
+        if AVERAGE == a:
+            spec = np.where(np.abs(specs[i]) == np.abs(spec), specs[i], spec)  
+
+    return spec
+
+def ensemble_inputs(audio_input, algorithm, is_normalization, wav_type_set, save_path):
+
+    wavs_ = []
+    
+    if algorithm == AVERAGE:
+        output = average_audio(audio_input)
+        samplerate = 44100
+    else:
+        specs = []
+        
+        for i in range(len(audio_input)):  
+            wave, samplerate = librosa.load(audio_input[i], mono=False, sr=44100)
+            wavs_.append(wave)
+            spec = wave_to_spectrogram_no_mp(wave)
+            specs.append(spec)
+        
+        wave_shapes = [w.shape[1] for w in wavs_]
+        target_shape = wavs_[wave_shapes.index(max(wave_shapes))]
+        
+        output = spectrogram_to_wave_no_mp(ensembling(algorithm, specs))
+        output = to_shape(output, target_shape.shape)
+
+    sf.write(save_path, normalize(output.T, is_normalization), samplerate, subtype=wav_type_set)
+
+def to_shape(x, target_shape):
+    padding_list = []
+    for x_dim, target_dim in zip(x.shape, target_shape):
+        pad_value = (target_dim - x_dim)
+        pad_tuple = ((0, pad_value))
+        padding_list.append(pad_tuple)
+    
+    return np.pad(x, tuple(padding_list), mode='constant')
+
+def to_shape_minimize(x: np.ndarray, target_shape):
+    
+    padding_list = []
+    for x_dim, target_dim in zip(x.shape, target_shape):
+        pad_value = (target_dim - x_dim)
+        pad_tuple = ((0, pad_value))
+        padding_list.append(pad_tuple)
+    
+    return np.pad(x, tuple(padding_list), mode='constant')
+
+def augment_audio(export_path, audio_file, rate, is_normalization, wav_type_set, save_format=None, is_pitch=False):
+
+    wav, sr = librosa.load(audio_file, sr=44100, mono=False)
+
+    if wav.ndim == 1:
+        wav = np.asfortranarray([wav,wav])
+
+    if is_pitch:
+        wav_1 = pyrb.pitch_shift(wav[0], sr, rate, rbargs=None)
+        wav_2 = pyrb.pitch_shift(wav[1], sr, rate, rbargs=None)
+    else:
+        wav_1 = pyrb.time_stretch(wav[0], sr, rate, rbargs=None)
+        wav_2 = pyrb.time_stretch(wav[1], sr, rate, rbargs=None)
+
+    if wav_1.shape > wav_2.shape:
+        wav_2 = to_shape(wav_2, wav_1.shape)
+    if wav_1.shape < wav_2.shape:
+        wav_1 = to_shape(wav_1, wav_2.shape)
+        
+    wav_mix = np.asfortranarray([wav_1, wav_2])
+    
+    sf.write(export_path, normalize(wav_mix.T, is_normalization), sr, subtype=wav_type_set)
+    save_format(export_path)
+    
+def average_audio(audio):
+    
+    waves = []
+    wave_shapes = []
+    final_waves = []
+
+    for i in range(len(audio)):
+        wave = librosa.load(audio[i], sr=44100, mono=False)
+        waves.append(wave[0])
+        wave_shapes.append(wave[0].shape[1])
+
+    wave_shapes_index = wave_shapes.index(max(wave_shapes))
+    target_shape = waves[wave_shapes_index]
+    waves.pop(wave_shapes_index)
+    final_waves.append(target_shape)
+
+    for n_array in waves:
+        wav_target = to_shape(n_array, target_shape.shape)
+        final_waves.append(wav_target)
+
+    waves = sum(final_waves)
+    waves = waves/len(audio)
+
+    return waves
+    
+def average_dual_sources(wav_1, wav_2, value):
+    
+    if wav_1.shape > wav_2.shape:
+        wav_2 = to_shape(wav_2, wav_1.shape)
+    if wav_1.shape < wav_2.shape:
+        wav_1 = to_shape(wav_1, wav_2.shape)
+
+    wave = (wav_1 * value) + (wav_2 * (1-value))
+
+    return wave
+    
+def reshape_sources(wav_1: np.ndarray, wav_2: np.ndarray):
+    
+    if wav_1.shape > wav_2.shape:
+        wav_2 = to_shape(wav_2, wav_1.shape)
+    if wav_1.shape < wav_2.shape:
+        ln = min([wav_1.shape[1], wav_2.shape[1]])
+        wav_2 = wav_2[:,:ln]
+
+    ln = min([wav_1.shape[1], wav_2.shape[1]])
+    wav_1 = wav_1[:,:ln]
+    wav_2 = wav_2[:,:ln]
+
+    return wav_2
+    
+def align_audio(file1, file2, file2_aligned, file_subtracted, wav_type_set, is_normalization, command_Text, progress_bar_main_var, save_format):
+    def get_diff(a, b):
+        corr = np.correlate(a, b, "full")
+        diff = corr.argmax() - (b.shape[0] - 1)
+        return diff
+  
+    progress_bar_main_var.set(10)
+    
+    # read tracks
+    wav1, sr1 = librosa.load(file1, sr=44100, mono=False)
+    wav2, sr2 = librosa.load(file2, sr=44100, mono=False)
+    wav1 = wav1.transpose()
+    wav2 = wav2.transpose()
+
+    command_Text(f"Audio file shapes: {wav1.shape} / {wav2.shape}\n")
+    
+    wav2_org = wav2.copy()
+    progress_bar_main_var.set(20)
+    
+    command_Text("Processing files... \n")
+    
+  # pick random position and get diff
+    
+    counts = {}       # counting up for each diff value
+    progress = 20
+    
+    check_range = 64
+
+    base = (64 / check_range)
+
+    for i in range(check_range):
+        index = int(random.uniform(44100 * 2, min(wav1.shape[0], wav2.shape[0]) - 44100 * 2))
+        shift = int(random.uniform(-22050,+22050))
+        samp1 = wav1[index      :index      +44100, 0]          # currently use left channel
+        samp2 = wav2[index+shift:index+shift+44100, 0]
+        progress += 1 * base
+        progress_bar_main_var.set(progress)
+        diff = get_diff(samp1, samp2)
+        diff -= shift
+        
+    if abs(diff) < 22050:
+        if not diff in counts:
+            counts[diff] = 0
+        counts[diff] += 1
+  
+  # use max counted diff value
+    max_count = 0
+    est_diff  = 0
+    for diff in counts.keys():
+        if counts[diff] > max_count:
+            max_count = counts[diff]
+            est_diff = diff
+    
+    command_Text(f"Estimated difference is {est_diff} (count: {max_count})\n")
+
+    progress_bar_main_var.set(90)
+    
+    audio_files = []
+
+    def save_aligned_audio(wav2_aligned):
+        command_Text(f"Aligned File 2 with File 1.\n")
+        command_Text(f"Saving files... ")
+        sf.write(file2_aligned, normalize(wav2_aligned, is_normalization), sr2, subtype=wav_type_set)
+        save_format(file2_aligned)
+        min_len = min(wav1.shape[0], wav2_aligned.shape[0])
+        wav_sub = wav1[:min_len] - wav2_aligned[:min_len]
+        audio_files.append(file2_aligned)
+        return min_len, wav_sub
+    
+  # make aligned track 2
+    if est_diff > 0:
+        wav2_aligned = np.append(np.zeros((est_diff, 2)), wav2_org, axis=0)
+        min_len, wav_sub = save_aligned_audio(wav2_aligned)
+    elif est_diff < 0:
+        wav2_aligned = wav2_org[-est_diff:]
+        min_len, wav_sub = save_aligned_audio(wav2_aligned)
+    else:
+        command_Text(f"Audio files already aligned.\n")
+        command_Text(f"Saving inverted track... ")
+        min_len = min(wav1.shape[0], wav2.shape[0])
+        wav_sub = wav1[:min_len] - wav2[:min_len]
+
+    wav_sub = np.clip(wav_sub, -1, +1)
+  
+    sf.write(file_subtracted, normalize(wav_sub, is_normalization), sr1, subtype=wav_type_set)
+    save_format(file_subtracted)
+  
+    progress_bar_main_var.set(95)

lib_v5/vr_network/__init__.py

@@ -0,0 +1 @@
+# VR init.

lib_v5/vr_network/layers.py

@@ -0,0 +1,143 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from lib_v5 import spec_utils
+
+class Conv2DBNActiv(nn.Module):
+
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin, nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False),
+            nn.BatchNorm2d(nout),
+            activ()
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+class SeperableConv2DBNActiv(nn.Module):
+
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin, nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False),
+            nn.Conv2d(
+                nin, nout,
+                kernel_size=1,
+                bias=False),
+            nn.BatchNorm2d(nout),
+            activ()
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+
+    def __init__(self, nn_architecture, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        )
+        
+        self.nn_architecture = nn_architecture
+        self.six_layer = [129605]
+        self.seven_layer = [537238, 537227, 33966]
+        
+        extra_conv = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
+        
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ)
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ)
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
+        
+        if self.nn_architecture in self.six_layer:
+            self.conv6 = extra_conv
+            nin_x = 6
+        elif self.nn_architecture in self.seven_layer:
+            self.conv6 = extra_conv
+            self.conv7 = extra_conv
+            nin_x = 7
+        else:
+            nin_x = 5
+            
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * nin_x, nout, 1, 1, 0, activ=activ),
+            nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        
+        if self.nn_architecture in self.six_layer:
+            feat6 = self.conv6(x)
+            out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6), dim=1)
+        elif self.nn_architecture in self.seven_layer:
+            feat6 = self.conv6(x)
+            feat7 = self.conv7(x)
+            out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+        else:
+            out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+            
+        bottle = self.bottleneck(out)
+        return bottle

lib_v5/vr_network/layers_new.py

@@ -0,0 +1,126 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from lib_v5 import spec_utils
+
+class Conv2DBNActiv(nn.Module):
+
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin, nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False),
+            nn.BatchNorm2d(nout),
+            activ()
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+class Encoder(nn.Module):
+
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, stride, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+
+    def __call__(self, x):
+        h = self.conv1(x)
+        h = self.conv2(h)
+
+        return h
+
+
+class Decoder(nn.Module):
+
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
+        super(Decoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        # self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
+
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+
+        h = self.conv1(x)
+        # h = self.conv2(h)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+
+    def __init__(self, nin, nout, dilations=(4, 8, 12), activ=nn.ReLU, dropout=False):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ)
+        )
+        self.conv2 = Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ)
+        self.conv3 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = Conv2DBNActiv(nout * 5, nout, 1, 1, 0, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        out = self.bottleneck(out)
+
+        if self.dropout is not None:
+            out = self.dropout(out)
+
+        return out
+
+
+class LSTMModule(nn.Module):
+
+    def __init__(self, nin_conv, nin_lstm, nout_lstm):
+        super(LSTMModule, self).__init__()
+        self.conv = Conv2DBNActiv(nin_conv, 1, 1, 1, 0)
+        self.lstm = nn.LSTM(
+            input_size=nin_lstm,
+            hidden_size=nout_lstm // 2,
+            bidirectional=True
+        )
+        self.dense = nn.Sequential(
+            nn.Linear(nout_lstm, nin_lstm),
+            nn.BatchNorm1d(nin_lstm),
+            nn.ReLU()
+        )
+
+    def forward(self, x):
+        N, _, nbins, nframes = x.size()
+        h = self.conv(x)[:, 0]  # N, nbins, nframes
+        h = h.permute(2, 0, 1)  # nframes, N, nbins
+        h, _ = self.lstm(h)
+        h = self.dense(h.reshape(-1, h.size()[-1]))  # nframes * N, nbins
+        h = h.reshape(nframes, N, 1, nbins)
+        h = h.permute(1, 2, 3, 0)
+
+        return h

lib_v5/vr_network/model_param_init.py

@@ -0,0 +1,59 @@
+import json
+import pathlib
+
+default_param = {}
+default_param['bins'] = 768
+default_param['unstable_bins'] = 9 # training only
+default_param['reduction_bins'] = 762 # training only
+default_param['sr'] = 44100
+default_param['pre_filter_start'] = 757
+default_param['pre_filter_stop'] = 768
+default_param['band'] = {}
+
+
+default_param['band'][1] = {
+    'sr': 11025,
+    'hl': 128,
+    'n_fft': 960,
+    'crop_start': 0,
+    'crop_stop': 245,
+    'lpf_start': 61, # inference only
+    'res_type': 'polyphase'
+}
+
+default_param['band'][2] = {
+    'sr': 44100,
+    'hl': 512,
+    'n_fft': 1536,
+    'crop_start': 24,
+    'crop_stop': 547,
+    'hpf_start': 81, # inference only
+    'res_type': 'sinc_best'
+}
+
+
+def int_keys(d):
+    r = {}
+    for k, v in d:
+        if k.isdigit():
+            k = int(k)
+        r[k] = v
+    return r
+    
+
+class ModelParameters(object):
+    def __init__(self, config_path=''):
+        if '.pth' == pathlib.Path(config_path).suffix:
+            import zipfile
+            
+            with zipfile.ZipFile(config_path, 'r') as zip:
+                self.param = json.loads(zip.read('param.json'), object_pairs_hook=int_keys)
+        elif '.json' == pathlib.Path(config_path).suffix:
+            with open(config_path, 'r') as f:
+                self.param = json.loads(f.read(), object_pairs_hook=int_keys)
+        else:
+            self.param = default_param
+            
+        for k in ['mid_side', 'mid_side_b', 'mid_side_b2', 'stereo_w', 'stereo_n', 'reverse']:
+            if not k in self.param:
+                self.param[k] = False

lib_v5/vr_network/modelparams/1band_sr16000_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 16000,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 16000,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

lib_v5/vr_network/modelparams/1band_sr32000_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 32000,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 32000,
+	"pre_filter_start": 1000,
+	"pre_filter_stop": 1021
+}

lib_v5/vr_network/modelparams/1band_sr33075_hl384.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 33075,
+			"hl": 384,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 33075,
+	"pre_filter_start": 1000,
+	"pre_filter_stop": 1021
+}

lib_v5/vr_network/modelparams/1band_sr44100_hl1024.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 1024,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

lib_v5/vr_network/modelparams/1band_sr44100_hl256.json

@@ -0,0 +1,19 @@
+{
+	"bins": 256,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 256,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 256,
+	"pre_filter_stop": 256
+}

lib_v5/vr_network/modelparams/1band_sr44100_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

lib_v5/vr_network/modelparams/1band_sr44100_hl512_cut.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 700,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 700
+}

lib_v5/vr_network/modelparams/1band_sr44100_hl512_nf1024.json

@@ -0,0 +1,19 @@
+{
+	"bins": 512,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 512,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 511,
+	"pre_filter_stop": 512
+}

lib_v5/vr_network/modelparams/2band_32000.json

@@ -0,0 +1,30 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 705,
+	"band": {
+		"1": {
+			"sr": 6000,
+			"hl": 66,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 240,
+			"lpf_start": 60,
+			"lpf_stop": 118,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 32000,
+			"hl": 352,
+			"n_fft": 1024,
+			"crop_start": 22,
+			"crop_stop": 505,
+			"hpf_start": 44,
+			"hpf_stop": 23,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 32000,
+	"pre_filter_start": 710,
+	"pre_filter_stop": 731
+}

lib_v5/vr_network/modelparams/2band_44100_lofi.json

@@ -0,0 +1,30 @@
+{
+	"bins": 512,
+	"unstable_bins": 7,
+	"reduction_bins": 510,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 160,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 192,
+			"lpf_start": 41,
+			"lpf_stop": 139,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 44100,
+			"hl": 640,
+			"n_fft": 1024,
+			"crop_start": 10,
+			"crop_stop": 320,
+			"hpf_start": 47,
+			"hpf_stop": 15,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 510,
+	"pre_filter_stop": 512
+}

lib_v5/vr_network/modelparams/2band_48000.json

@@ -0,0 +1,30 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 705,
+	"band": {
+		"1": {
+			"sr": 6000,
+			"hl": 66,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 240,
+			"lpf_start": 60,
+			"lpf_stop": 240,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 48000,
+			"hl": 528,
+			"n_fft": 1536,
+			"crop_start": 22,
+			"crop_stop": 505,
+			"hpf_start": 82,
+			"hpf_stop": 22,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 48000,
+	"pre_filter_start": 710,
+	"pre_filter_stop": 731
+}

lib_v5/vr_network/modelparams/3band_44100.json

@@ -0,0 +1,42 @@
+{
+	"bins": 768,
+	"unstable_bins": 5,
+	"reduction_bins": 733,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 278,
+			"lpf_start": 28,
+			"lpf_stop": 140,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 768,
+			"crop_start": 14,
+			"crop_stop": 322,
+			"hpf_start": 70,
+			"hpf_stop": 14,
+			"lpf_start": 283,
+			"lpf_stop": 314,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 131,
+			"crop_stop": 313,
+			"hpf_start": 154,
+			"hpf_stop": 141,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 757,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/3band_44100_mid.json

@@ -0,0 +1,43 @@
+{
+	"mid_side": true,
+	"bins": 768,
+	"unstable_bins": 5,
+	"reduction_bins": 733,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 278,
+			"lpf_start": 28,
+			"lpf_stop": 140,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 768,
+			"crop_start": 14,
+			"crop_stop": 322,
+			"hpf_start": 70,
+			"hpf_stop": 14,
+			"lpf_start": 283,
+			"lpf_stop": 314,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 131,
+			"crop_stop": 313,
+			"hpf_start": 154,
+			"hpf_stop": 141,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 757,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/3band_44100_msb2.json

@@ -0,0 +1,43 @@
+{
+	"mid_side_b2": true,
+	"bins": 640,
+	"unstable_bins": 7,
+	"reduction_bins": 565,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 108,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 187,
+			"lpf_start": 92,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 216,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 212,
+			"hpf_start": 68,
+			"hpf_stop": 34,
+			"lpf_start": 174,
+			"lpf_stop": 209,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 432,
+			"n_fft": 640,
+			"crop_start": 66,
+			"crop_stop": 307,
+			"hpf_start": 86,
+			"hpf_stop": 72,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 639,
+	"pre_filter_stop": 640
+}

lib_v5/vr_network/modelparams/4band_44100.json

@@ -0,0 +1,54 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/4band_44100_mid.json

@@ -0,0 +1,55 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"mid_side": true,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/4band_44100_msb.json

@@ -0,0 +1,55 @@
+{
+	"mid_side_b": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/4band_44100_msb2.json

@@ -0,0 +1,55 @@
+{
+	"mid_side_b": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/4band_44100_reverse.json

@@ -0,0 +1,55 @@
+{
+	"reverse": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/4band_44100_sw.json

@@ -0,0 +1,55 @@
+{
+	"stereo_w": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

lib_v5/vr_network/modelparams/4band_v2.json

@@ -0,0 +1,54 @@
+{
+	"bins": 672,
+	"unstable_bins": 8,
+	"reduction_bins": 637,
+	"band": {
+		"1": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 640,
+			"crop_start": 0,
+			"crop_stop": 85,
+			"lpf_start": 25,
+			"lpf_stop": 53,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 320,
+			"crop_start": 4,
+			"crop_stop": 87,
+			"hpf_start": 25,
+			"hpf_stop": 12,
+			"lpf_start": 31,
+			"lpf_stop": 62,
+			"res_type": "polyphase"
+		},		
+		"3": {
+			"sr": 14700,
+			"hl": 160,
+			"n_fft": 512,
+			"crop_start": 17,
+			"crop_stop": 216,
+			"hpf_start": 48,
+			"hpf_stop": 24,
+			"lpf_start": 139,
+			"lpf_stop": 210,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 480,
+			"n_fft": 960,
+			"crop_start": 78,
+			"crop_stop": 383,
+			"hpf_start": 130,
+			"hpf_stop": 86,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 668,
+	"pre_filter_stop": 672
+}

lib_v5/vr_network/modelparams/4band_v2_sn.json

@@ -0,0 +1,55 @@
+{
+	"bins": 672,
+	"unstable_bins": 8,
+	"reduction_bins": 637,
+	"band": {
+		"1": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 640,
+			"crop_start": 0,
+			"crop_stop": 85,
+			"lpf_start": 25,
+			"lpf_stop": 53,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 320,
+			"crop_start": 4,
+			"crop_stop": 87,
+			"hpf_start": 25,
+			"hpf_stop": 12,
+			"lpf_start": 31,
+			"lpf_stop": 62,
+			"res_type": "polyphase"
+		},		
+		"3": {
+			"sr": 14700,
+			"hl": 160,
+			"n_fft": 512,
+			"crop_start": 17,
+			"crop_stop": 216,
+			"hpf_start": 48,
+			"hpf_stop": 24,
+			"lpf_start": 139,
+			"lpf_stop": 210,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 480,
+			"n_fft": 960,
+			"crop_start": 78,
+			"crop_stop": 383,
+			"hpf_start": 130,
+			"hpf_stop": 86,
+			"convert_channels": "stereo_n",
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 668,
+	"pre_filter_stop": 672
+}

lib_v5/vr_network/modelparams/4band_v3.json

@@ -0,0 +1,54 @@
+{
+	"bins": 672,
+	"unstable_bins": 8,
+	"reduction_bins": 530,
+	"band": {
+		"1": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 640,
+			"crop_start": 0,
+			"crop_stop": 85,
+			"lpf_start": 25,
+			"lpf_stop": 53,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 320,
+			"crop_start": 4,
+			"crop_stop": 87,
+			"hpf_start": 25,
+			"hpf_stop": 12,
+			"lpf_start": 31,
+			"lpf_stop": 62,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 14700,
+			"hl": 160,
+			"n_fft": 512,
+			"crop_start": 17,
+			"crop_stop": 216,
+			"hpf_start": 48,
+			"hpf_stop": 24,
+			"lpf_start": 139,
+			"lpf_stop": 210,
+			"res_type": "polyphase"
+		},
+		"4": {
+			"sr": 44100,
+			"hl": 480,
+			"n_fft": 960,
+			"crop_start": 78,
+			"crop_stop": 383,
+			"hpf_start": 130,
+			"hpf_stop": 86,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 668,
+	"pre_filter_stop": 672
+}

lib_v5/vr_network/modelparams/ensemble.json

@@ -0,0 +1,43 @@
+{
+	"mid_side_b2": true,
+	"bins": 1280,
+	"unstable_bins": 7,
+	"reduction_bins": 565,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 108,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 374,
+			"lpf_start": 92,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 216,
+			"n_fft": 1536,
+			"crop_start": 0,
+			"crop_stop": 424,
+			"hpf_start": 68,
+			"hpf_stop": 34,
+			"lpf_start": 348,
+			"lpf_stop": 418,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 432,
+			"n_fft": 1280,
+			"crop_start": 132,
+			"crop_stop": 614,
+			"hpf_start": 172,
+			"hpf_stop": 144,
+			"res_type": "polyphase"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1280,
+	"pre_filter_stop": 1280
+}

lib_v5/vr_network/nets.py

@@ -0,0 +1,166 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from . import layers
+
+class BaseASPPNet(nn.Module):
+
+    def __init__(self, nn_architecture, nin, ch, dilations=(4, 8, 16)):
+        super(BaseASPPNet, self).__init__()
+        self.nn_architecture = nn_architecture
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+        
+        if self.nn_architecture == 129605:
+            self.enc5 = layers.Encoder(ch * 8, ch * 16, 3, 2, 1)
+            self.aspp = layers.ASPPModule(nn_architecture, ch * 16, ch * 32, dilations)
+            self.dec5 = layers.Decoder(ch * (16 + 32), ch * 16, 3, 1, 1)
+        else:
+            self.aspp = layers.ASPPModule(nn_architecture, ch * 8, ch * 16, dilations)
+            
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+        
+        if self.nn_architecture == 129605:
+            h, e5 = self.enc5(h)
+            h = self.aspp(h)
+            h = self.dec5(h, e5)
+        else:
+            h = self.aspp(h)
+            
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+def determine_model_capacity(n_fft_bins, nn_architecture):
+    
+    sp_model_arch = [31191, 33966, 129605]
+    hp_model_arch = [123821, 123812]
+    hp2_model_arch = [537238, 537227]
+    
+    if nn_architecture in sp_model_arch:
+        model_capacity_data = [
+            (2, 16),
+            (2, 16),
+            (18, 8, 1, 1, 0),
+            (8, 16),
+            (34, 16, 1, 1, 0),
+            (16, 32),
+            (32, 2, 1),
+            (16, 2, 1),
+            (16, 2, 1),
+        ]
+    
+    if nn_architecture in hp_model_arch:
+        model_capacity_data = [
+            (2, 32),
+            (2, 32),
+            (34, 16, 1, 1, 0),
+            (16, 32),
+            (66, 32, 1, 1, 0),
+            (32, 64),
+            (64, 2, 1),
+            (32, 2, 1),
+            (32, 2, 1),
+        ]
+       
+    if nn_architecture in hp2_model_arch: 
+        model_capacity_data = [
+            (2, 64),
+            (2, 64),
+            (66, 32, 1, 1, 0),
+            (32, 64),
+            (130, 64, 1, 1, 0),
+            (64, 128),
+            (128, 2, 1),
+            (64, 2, 1),
+            (64, 2, 1),
+        ]
+
+    cascaded = CascadedASPPNet
+    model = cascaded(n_fft_bins, model_capacity_data, nn_architecture)
+    
+    return model
+
+class CascadedASPPNet(nn.Module):
+
+    def __init__(self, n_fft, model_capacity_data, nn_architecture):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(nn_architecture, *model_capacity_data[0])
+        self.stg1_high_band_net = BaseASPPNet(nn_architecture, *model_capacity_data[1])
+
+        self.stg2_bridge = layers.Conv2DBNActiv(*model_capacity_data[2])
+        self.stg2_full_band_net = BaseASPPNet(nn_architecture, *model_capacity_data[3])
+
+        self.stg3_bridge = layers.Conv2DBNActiv(*model_capacity_data[4])
+        self.stg3_full_band_net = BaseASPPNet(nn_architecture, *model_capacity_data[5])
+
+        self.out = nn.Conv2d(*model_capacity_data[6], bias=False)
+        self.aux1_out = nn.Conv2d(*model_capacity_data[7], bias=False)
+        self.aux2_out = nn.Conv2d(*model_capacity_data[8], bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, :self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat([
+            self.stg1_low_band_net(x[:, :, :bandw]),
+            self.stg1_high_band_net(x[:, :, bandw:])
+        ], dim=2)
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode='replicate')
+ 
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode='replicate')
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode='replicate')
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            return mask# * mix
+
+    def predict_mask(self, x):
+        mask = self.forward(x)
+
+        if self.offset > 0:
+            mask = mask[:, :, :, self.offset:-self.offset]
+
+        return mask

lib_v5/vr_network/nets_new.py

@@ -0,0 +1,125 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from . import layers_new as layers
+
+class BaseNet(nn.Module):
+
+    def __init__(self, nin, nout, nin_lstm, nout_lstm, dilations=((4, 2), (8, 4), (12, 6))):
+        super(BaseNet, self).__init__()
+        self.enc1 = layers.Conv2DBNActiv(nin, nout, 3, 1, 1)
+        self.enc2 = layers.Encoder(nout, nout * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(nout * 2, nout * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(nout * 4, nout * 6, 3, 2, 1)
+        self.enc5 = layers.Encoder(nout * 6, nout * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(nout * 8, nout * 8, dilations, dropout=True)
+
+        self.dec4 = layers.Decoder(nout * (6 + 8), nout * 6, 3, 1, 1)
+        self.dec3 = layers.Decoder(nout * (4 + 6), nout * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(nout * (2 + 4), nout * 2, 3, 1, 1)
+        self.lstm_dec2 = layers.LSTMModule(nout * 2, nin_lstm, nout_lstm)
+        self.dec1 = layers.Decoder(nout * (1 + 2) + 1, nout * 1, 3, 1, 1)
+
+    def __call__(self, x):
+        e1 = self.enc1(x)
+        e2 = self.enc2(e1)
+        e3 = self.enc3(e2)
+        e4 = self.enc4(e3)
+        e5 = self.enc5(e4)
+
+        h = self.aspp(e5)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = torch.cat([h, self.lstm_dec2(h)], dim=1)
+        h = self.dec1(h, e1)
+
+        return h
+
+class CascadedNet(nn.Module):
+
+    def __init__(self, n_fft, nn_arch_size, nout=32, nout_lstm=128):
+        super(CascadedNet, self).__init__()
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+        self.nin_lstm = self.max_bin // 2
+        self.offset = 64
+        nout = 64 if nn_arch_size == 218409 else nout
+
+        self.stg1_low_band_net = nn.Sequential(
+            BaseNet(2, nout // 2, self.nin_lstm // 2, nout_lstm),
+            layers.Conv2DBNActiv(nout // 2, nout // 4, 1, 1, 0)
+            )
+        
+        self.stg1_high_band_net = BaseNet(2, nout // 4, self.nin_lstm // 2, nout_lstm // 2)
+
+        self.stg2_low_band_net = nn.Sequential(
+            BaseNet(nout // 4 + 2, nout, self.nin_lstm // 2, nout_lstm),
+            layers.Conv2DBNActiv(nout, nout // 2, 1, 1, 0)
+            )
+        self.stg2_high_band_net = BaseNet(nout // 4 + 2, nout // 2, self.nin_lstm // 2, nout_lstm // 2)
+
+        self.stg3_full_band_net = BaseNet(3 * nout // 4 + 2, nout, self.nin_lstm, nout_lstm)
+
+        self.out = nn.Conv2d(nout, 2, 1, bias=False)
+        self.aux_out = nn.Conv2d(3 * nout // 4, 2, 1, bias=False)
+
+    def forward(self, x):
+        x = x[:, :, :self.max_bin]
+
+        bandw = x.size()[2] // 2
+        l1_in = x[:, :, :bandw]
+        h1_in = x[:, :, bandw:]
+        l1 = self.stg1_low_band_net(l1_in)
+        h1 = self.stg1_high_band_net(h1_in)
+        aux1 = torch.cat([l1, h1], dim=2)
+
+        l2_in = torch.cat([l1_in, l1], dim=1)
+        h2_in = torch.cat([h1_in, h1], dim=1)
+        l2 = self.stg2_low_band_net(l2_in)
+        h2 = self.stg2_high_band_net(h2_in)
+        aux2 = torch.cat([l2, h2], dim=2)
+
+        f3_in = torch.cat([x, aux1, aux2], dim=1)
+        f3 = self.stg3_full_band_net(f3_in)
+
+        mask = torch.sigmoid(self.out(f3))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode='replicate'
+        )
+
+        if self.training:
+            aux = torch.cat([aux1, aux2], dim=1)
+            aux = torch.sigmoid(self.aux_out(aux))
+            aux = F.pad(
+                input=aux,
+                pad=(0, 0, 0, self.output_bin - aux.size()[2]),
+                mode='replicate'
+            )
+            return mask, aux
+        else:
+            return mask
+
+    def predict_mask(self, x):
+        mask = self.forward(x)
+
+        if self.offset > 0:
+            mask = mask[:, :, :, self.offset:-self.offset]
+            assert mask.size()[3] > 0
+
+        return mask
+
+    def predict(self, x):
+        mask = self.forward(x)
+        pred_mag = x * mask
+
+        if self.offset > 0:
+            pred_mag = pred_mag[:, :, :, self.offset:-self.offset]
+            assert pred_mag.size()[3] > 0
+
+        return pred_mag