# Copyright (c) Meta, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # First author is Simon Rouard. """ This code contains the spectrogram and Hybrid version of Demucs. """ import math from .filtering import wiener import torch from torch import nn from torch.nn import functional as F from fractions import Fraction from einops import rearrange from .transformer import CrossTransformerEncoder from .demucs import rescale_module from .states import capture_init from .spec import spectro, ispectro from .hdemucs import pad1d, ScaledEmbedding, HEncLayer, MultiWrap, HDecLayer class HTDemucs(nn.Module): """ Spectrogram and hybrid Demucs model. The spectrogram model has the same structure as Demucs, except the first few layers are over the frequency axis, until there is only 1 frequency, and then it moves to time convolutions. Frequency layers can still access information across time steps thanks to the DConv residual. Hybrid model have a parallel time branch. At some layer, the time branch has the same stride as the frequency branch and then the two are combined. The opposite happens in the decoder. Models can either use naive iSTFT from masking, Wiener filtering ([Ulhih et al. 2017]), or complex as channels (CaC) [Choi et al. 2020]. Wiener filtering is based on Open Unmix implementation [Stoter et al. 2019]. The loss is always on the temporal domain, by backpropagating through the above output methods and iSTFT. This allows to define hybrid models nicely. However, this breaks a bit Wiener filtering, as doing more iteration at test time will change the spectrogram contribution, without changing the one from the waveform, which will lead to worse performance. I tried using the residual option in OpenUnmix Wiener implementation, but it didn't improve. CaC on the other hand provides similar performance for hybrid, and works naturally with hybrid models. This model also uses frequency embeddings are used to improve efficiency on convolutions over the freq. axis, following [Isik et al. 2020] (https://arxiv.org/pdf/2008.04470.pdf). Unlike classic Demucs, there is no resampling here, and normalization is always applied. """ @capture_init def __init__( self, sources, # Channels audio_channels=2, channels=48, channels_time=None, growth=2, # STFT nfft=4096, wiener_iters=0, end_iters=0, wiener_residual=False, cac=True, # Main structure depth=4, rewrite=True, # Frequency branch multi_freqs=None, multi_freqs_depth=3, freq_emb=0.2, emb_scale=10, emb_smooth=True, # Convolutions kernel_size=8, time_stride=2, stride=4, context=1, context_enc=0, # Normalization norm_starts=4, norm_groups=4, # DConv residual branch dconv_mode=1, dconv_depth=2, dconv_comp=8, dconv_init=1e-3, # Before the Transformer bottom_channels=0, # Transformer t_layers=5, t_emb="sin", t_hidden_scale=4.0, t_heads=8, t_dropout=0.0, t_max_positions=10000, t_norm_in=True, t_norm_in_group=False, t_group_norm=False, t_norm_first=True, t_norm_out=True, t_max_period=10000.0, t_weight_decay=0.0, t_lr=None, t_layer_scale=True, t_gelu=True, t_weight_pos_embed=1.0, t_sin_random_shift=0, t_cape_mean_normalize=True, t_cape_augment=True, t_cape_glob_loc_scale=[5000.0, 1.0, 1.4], t_sparse_self_attn=False, t_sparse_cross_attn=False, t_mask_type="diag", t_mask_random_seed=42, t_sparse_attn_window=500, t_global_window=100, t_sparsity=0.95, t_auto_sparsity=False, # ------ Particuliar parameters t_cross_first=False, # Weight init rescale=0.1, # Metadata samplerate=44100, segment=10, use_train_segment=True, ): """ Args: sources (list[str]): list of source names. audio_channels (int): input/output audio channels. channels (int): initial number of hidden channels. channels_time: if not None, use a different `channels` value for the time branch. growth: increase the number of hidden channels by this factor at each layer. nfft: number of fft bins. Note that changing this require careful computation of various shape parameters and will not work out of the box for hybrid models. wiener_iters: when using Wiener filtering, number of iterations at test time. end_iters: same but at train time. For a hybrid model, must be equal to `wiener_iters`. wiener_residual: add residual source before wiener filtering. cac: uses complex as channels, i.e. complex numbers are 2 channels each in input and output. no further processing is done before ISTFT. depth (int): number of layers in the encoder and in the decoder. rewrite (bool): add 1x1 convolution to each layer. multi_freqs: list of frequency ratios for splitting frequency bands with `MultiWrap`. multi_freqs_depth: how many layers to wrap with `MultiWrap`. Only the outermost layers will be wrapped. freq_emb: add frequency embedding after the first frequency layer if > 0, the actual value controls the weight of the embedding. emb_scale: equivalent to scaling the embedding learning rate emb_smooth: initialize the embedding with a smooth one (with respect to frequencies). kernel_size: kernel_size for encoder and decoder layers. stride: stride for encoder and decoder layers. time_stride: stride for the final time layer, after the merge. context: context for 1x1 conv in the decoder. context_enc: context for 1x1 conv in the encoder. norm_starts: layer at which group norm starts being used. decoder layers are numbered in reverse order. norm_groups: number of groups for group norm. dconv_mode: if 1: dconv in encoder only, 2: decoder only, 3: both. dconv_depth: depth of residual DConv branch. dconv_comp: compression of DConv branch. dconv_attn: adds attention layers in DConv branch starting at this layer. dconv_lstm: adds a LSTM layer in DConv branch starting at this layer. dconv_init: initial scale for the DConv branch LayerScale. bottom_channels: if >0 it adds a linear layer (1x1 Conv) before and after the transformer in order to change the number of channels t_layers: number of layers in each branch (waveform and spec) of the transformer t_emb: "sin", "cape" or "scaled" t_hidden_scale: the hidden scale of the Feedforward parts of the transformer for instance if C = 384 (the number of channels in the transformer) and t_hidden_scale = 4.0 then the intermediate layer of the FFN has dimension 384 * 4 = 1536 t_heads: number of heads for the transformer t_dropout: dropout in the transformer t_max_positions: max_positions for the "scaled" positional embedding, only useful if t_emb="scaled" t_norm_in: (bool) norm before addinf positional embedding and getting into the transformer layers t_norm_in_group: (bool) if True while t_norm_in=True, the norm is on all the timesteps (GroupNorm with group=1) t_group_norm: (bool) if True, the norms of the Encoder Layers are on all the timesteps (GroupNorm with group=1) t_norm_first: (bool) if True the norm is before the attention and before the FFN t_norm_out: (bool) if True, there is a GroupNorm (group=1) at the end of each layer t_max_period: (float) denominator in the sinusoidal embedding expression t_weight_decay: (float) weight decay for the transformer t_lr: (float) specific learning rate for the transformer t_layer_scale: (bool) Layer Scale for the transformer t_gelu: (bool) activations of the transformer are GeLU if True, ReLU else t_weight_pos_embed: (float) weighting of the positional embedding t_cape_mean_normalize: (bool) if t_emb="cape", normalisation of positional embeddings see: https://arxiv.org/abs/2106.03143 t_cape_augment: (bool) if t_emb="cape", must be True during training and False during the inference, see: https://arxiv.org/abs/2106.03143 t_cape_glob_loc_scale: (list of 3 floats) if t_emb="cape", CAPE parameters see: https://arxiv.org/abs/2106.03143 t_sparse_self_attn: (bool) if True, the self attentions are sparse t_sparse_cross_attn: (bool) if True, the cross-attentions are sparse (don't use it unless you designed really specific masks) t_mask_type: (str) can be "diag", "jmask", "random", "global" or any combination with '_' between: i.e. "diag_jmask_random" (note that this is permutation invariant i.e. "diag_jmask_random" is equivalent to "jmask_random_diag") t_mask_random_seed: (int) if "random" is in t_mask_type, controls the seed that generated the random part of the mask t_sparse_attn_window: (int) if "diag" is in t_mask_type, for a query (i), and a key (j), the mask is True id |i-j|<=t_sparse_attn_window t_global_window: (int) if "global" is in t_mask_type, mask[:t_global_window, :] and mask[:, :t_global_window] will be True t_sparsity: (float) if "random" is in t_mask_type, t_sparsity is the sparsity level of the random part of the mask. t_cross_first: (bool) if True cross attention is the first layer of the transformer (False seems to be better) rescale: weight rescaling trick use_train_segment: (bool) if True, the actual size that is used during the training is used during inference. """ super().__init__() self.cac = cac self.wiener_residual = wiener_residual self.audio_channels = audio_channels self.sources = sources self.kernel_size = kernel_size self.context = context self.stride = stride self.depth = depth self.bottom_channels = bottom_channels self.channels = channels self.samplerate = samplerate self.segment = segment self.use_train_segment = use_train_segment self.nfft = nfft self.hop_length = nfft // 4 self.wiener_iters = wiener_iters self.end_iters = end_iters self.freq_emb = None assert wiener_iters == end_iters self.encoder = nn.ModuleList() self.decoder = nn.ModuleList() self.tencoder = nn.ModuleList() self.tdecoder = nn.ModuleList() chin = audio_channels chin_z = chin # number of channels for the freq branch if self.cac: chin_z *= 2 chout = channels_time or channels chout_z = channels freqs = nfft // 2 for index in range(depth): norm = index >= norm_starts freq = freqs > 1 stri = stride ker = kernel_size if not freq: assert freqs == 1 ker = time_stride * 2 stri = time_stride pad = True last_freq = False if freq and freqs <= kernel_size: ker = freqs pad = False last_freq = True kw = { "kernel_size": ker, "stride": stri, "freq": freq, "pad": pad, "norm": norm, "rewrite": rewrite, "norm_groups": norm_groups, "dconv_kw": { "depth": dconv_depth, "compress": dconv_comp, "init": dconv_init, "gelu": True, }, } kwt = dict(kw) kwt["freq"] = 0 kwt["kernel_size"] = kernel_size kwt["stride"] = stride kwt["pad"] = True kw_dec = dict(kw) multi = False if multi_freqs and index < multi_freqs_depth: multi = True kw_dec["context_freq"] = False if last_freq: chout_z = max(chout, chout_z) chout = chout_z enc = HEncLayer( chin_z, chout_z, dconv=dconv_mode & 1, context=context_enc, **kw ) if freq: tenc = HEncLayer( chin, chout, dconv=dconv_mode & 1, context=context_enc, empty=last_freq, **kwt ) self.tencoder.append(tenc) if multi: enc = MultiWrap(enc, multi_freqs) self.encoder.append(enc) if index == 0: chin = self.audio_channels * len(self.sources) chin_z = chin if self.cac: chin_z *= 2 dec = HDecLayer( chout_z, chin_z, dconv=dconv_mode & 2, last=index == 0, context=context, **kw_dec ) if multi: dec = MultiWrap(dec, multi_freqs) if freq: tdec = HDecLayer( chout, chin, dconv=dconv_mode & 2, empty=last_freq, last=index == 0, context=context, **kwt ) self.tdecoder.insert(0, tdec) self.decoder.insert(0, dec) chin = chout chin_z = chout_z chout = int(growth * chout) chout_z = int(growth * chout_z) if freq: if freqs <= kernel_size: freqs = 1 else: freqs //= stride if index == 0 and freq_emb: self.freq_emb = ScaledEmbedding( freqs, chin_z, smooth=emb_smooth, scale=emb_scale ) self.freq_emb_scale = freq_emb if rescale: rescale_module(self, reference=rescale) transformer_channels = channels * growth ** (depth - 1) if bottom_channels: self.channel_upsampler = nn.Conv1d(transformer_channels, bottom_channels, 1) self.channel_downsampler = nn.Conv1d( bottom_channels, transformer_channels, 1 ) self.channel_upsampler_t = nn.Conv1d( transformer_channels, bottom_channels, 1 ) self.channel_downsampler_t = nn.Conv1d( bottom_channels, transformer_channels, 1 ) transformer_channels = bottom_channels if t_layers > 0: self.crosstransformer = CrossTransformerEncoder( dim=transformer_channels, emb=t_emb, hidden_scale=t_hidden_scale, num_heads=t_heads, num_layers=t_layers, cross_first=t_cross_first, dropout=t_dropout, max_positions=t_max_positions, norm_in=t_norm_in, norm_in_group=t_norm_in_group, group_norm=t_group_norm, norm_first=t_norm_first, norm_out=t_norm_out, max_period=t_max_period, weight_decay=t_weight_decay, lr=t_lr, layer_scale=t_layer_scale, gelu=t_gelu, sin_random_shift=t_sin_random_shift, weight_pos_embed=t_weight_pos_embed, cape_mean_normalize=t_cape_mean_normalize, cape_augment=t_cape_augment, cape_glob_loc_scale=t_cape_glob_loc_scale, sparse_self_attn=t_sparse_self_attn, sparse_cross_attn=t_sparse_cross_attn, mask_type=t_mask_type, mask_random_seed=t_mask_random_seed, sparse_attn_window=t_sparse_attn_window, global_window=t_global_window, sparsity=t_sparsity, auto_sparsity=t_auto_sparsity, ) else: self.crosstransformer = None def _spec(self, x): hl = self.hop_length nfft = self.nfft x0 = x # noqa # We re-pad the signal in order to keep the property # that the size of the output is exactly the size of the input # divided by the stride (here hop_length), when divisible. # This is achieved by padding by 1/4th of the kernel size (here nfft). # which is not supported by torch.stft. # Having all convolution operations follow this convention allow to easily # align the time and frequency branches later on. assert hl == nfft // 4 le = int(math.ceil(x.shape[-1] / hl)) pad = hl // 2 * 3 x = pad1d(x, (pad, pad + le * hl - x.shape[-1]), mode="reflect") z = spectro(x, nfft, hl)[..., :-1, :] assert z.shape[-1] == le + 4, (z.shape, x.shape, le) z = z[..., 2: 2 + le] return z def _ispec(self, z, length=None, scale=0): hl = self.hop_length // (4**scale) z = F.pad(z, (0, 0, 0, 1)) z = F.pad(z, (2, 2)) pad = hl // 2 * 3 le = hl * int(math.ceil(length / hl)) + 2 * pad x = ispectro(z, hl, length=le) x = x[..., pad: pad + length] return x def _magnitude(self, z): # return the magnitude of the spectrogram, except when cac is True, # in which case we just move the complex dimension to the channel one. if self.cac: B, C, Fr, T = z.shape m = torch.view_as_real(z).permute(0, 1, 4, 2, 3) m = m.reshape(B, C * 2, Fr, T) else: m = z.abs() return m def _mask(self, z, m): # Apply masking given the mixture spectrogram `z` and the estimated mask `m`. # If `cac` is True, `m` is actually a full spectrogram and `z` is ignored. niters = self.wiener_iters if self.cac: B, S, C, Fr, T = m.shape out = m.view(B, S, -1, 2, Fr, T).permute(0, 1, 2, 4, 5, 3) out = torch.view_as_complex(out.contiguous()) return out if self.training: niters = self.end_iters if niters < 0: z = z[:, None] return z / (1e-8 + z.abs()) * m else: return self._wiener(m, z, niters) def _wiener(self, mag_out, mix_stft, niters): # apply wiener filtering from OpenUnmix. init = mix_stft.dtype wiener_win_len = 300 residual = self.wiener_residual B, S, C, Fq, T = mag_out.shape mag_out = mag_out.permute(0, 4, 3, 2, 1) mix_stft = torch.view_as_real(mix_stft.permute(0, 3, 2, 1)) outs = [] for sample in range(B): pos = 0 out = [] for pos in range(0, T, wiener_win_len): frame = slice(pos, pos + wiener_win_len) z_out = wiener( mag_out[sample, frame], mix_stft[sample, frame], niters, residual=residual, ) out.append(z_out.transpose(-1, -2)) outs.append(torch.cat(out, dim=0)) out = torch.view_as_complex(torch.stack(outs, 0)) out = out.permute(0, 4, 3, 2, 1).contiguous() if residual: out = out[:, :-1] assert list(out.shape) == [B, S, C, Fq, T] return out.to(init) def valid_length(self, length: int): """ Return a length that is appropriate for evaluation. In our case, always return the training length, unless it is smaller than the given length, in which case this raises an error. """ if not self.use_train_segment: return length training_length = int(self.segment * self.samplerate) if training_length < length: raise ValueError( f"Given length {length} is longer than " f"training length {training_length}") return training_length def forward(self, mix): length = mix.shape[-1] length_pre_pad = None if self.use_train_segment: if self.training: self.segment = Fraction(mix.shape[-1], self.samplerate) else: training_length = int(self.segment * self.samplerate) if mix.shape[-1] < training_length: length_pre_pad = mix.shape[-1] mix = F.pad(mix, (0, training_length - length_pre_pad)) z = self._spec(mix) mag = self._magnitude(z).to(mix.device) x = mag B, C, Fq, T = x.shape # unlike previous Demucs, we always normalize because it is easier. mean = x.mean(dim=(1, 2, 3), keepdim=True) std = x.std(dim=(1, 2, 3), keepdim=True) x = (x - mean) / (1e-5 + std) # x will be the freq. branch input. # Prepare the time branch input. xt = mix meant = xt.mean(dim=(1, 2), keepdim=True) stdt = xt.std(dim=(1, 2), keepdim=True) xt = (xt - meant) / (1e-5 + stdt) # okay, this is a giant mess I know... saved = [] # skip connections, freq. saved_t = [] # skip connections, time. lengths = [] # saved lengths to properly remove padding, freq branch. lengths_t = [] # saved lengths for time branch. for idx, encode in enumerate(self.encoder): lengths.append(x.shape[-1]) inject = None if idx < len(self.tencoder): # we have not yet merged branches. lengths_t.append(xt.shape[-1]) tenc = self.tencoder[idx] xt = tenc(xt) if not tenc.empty: # save for skip connection saved_t.append(xt) else: # tenc contains just the first conv., so that now time and freq. # branches have the same shape and can be merged. inject = xt x = encode(x, inject) if idx == 0 and self.freq_emb is not None: # add frequency embedding to allow for non equivariant convolutions # over the frequency axis. frs = torch.arange(x.shape[-2], device=x.device) emb = self.freq_emb(frs).t()[None, :, :, None].expand_as(x) x = x + self.freq_emb_scale * emb saved.append(x) if self.crosstransformer: if self.bottom_channels: b, c, f, t = x.shape x = rearrange(x, "b c f t-> b c (f t)") x = self.channel_upsampler(x) x = rearrange(x, "b c (f t)-> b c f t", f=f) xt = self.channel_upsampler_t(xt) x, xt = self.crosstransformer(x, xt) if self.bottom_channels: x = rearrange(x, "b c f t-> b c (f t)") x = self.channel_downsampler(x) x = rearrange(x, "b c (f t)-> b c f t", f=f) xt = self.channel_downsampler_t(xt) for idx, decode in enumerate(self.decoder): skip = saved.pop(-1) x, pre = decode(x, skip, lengths.pop(-1)) # `pre` contains the output just before final transposed convolution, # which is used when the freq. and time branch separate. offset = self.depth - len(self.tdecoder) if idx >= offset: tdec = self.tdecoder[idx - offset] length_t = lengths_t.pop(-1) if tdec.empty: assert pre.shape[2] == 1, pre.shape pre = pre[:, :, 0] xt, _ = tdec(pre, None, length_t) else: skip = saved_t.pop(-1) xt, _ = tdec(xt, skip, length_t) # Let's make sure we used all stored skip connections. assert len(saved) == 0 assert len(lengths_t) == 0 assert len(saved_t) == 0 S = len(self.sources) x = x.view(B, S, -1, Fq, T) x = x * std[:, None] + mean[:, None] # to cpu as non-cuda GPUs don't support complex numbers # demucs issue #435 ##432 # NOTE: in this case z already is on cpu # TODO: remove this when mps supports complex numbers device_type = x.device.type device_load = f"{device_type}:{x.device.index}" if not device_type == 'mps' else device_type x_is_other_gpu = not device_type in ["cuda", "cpu"] if x_is_other_gpu: x = x.cpu() zout = self._mask(z, x) if self.use_train_segment: if self.training: x = self._ispec(zout, length) else: x = self._ispec(zout, training_length) else: x = self._ispec(zout, length) # back to other device if x_is_other_gpu: x = x.to(device_load) if self.use_train_segment: if self.training: xt = xt.view(B, S, -1, length) else: xt = xt.view(B, S, -1, training_length) else: xt = xt.view(B, S, -1, length) xt = xt * stdt[:, None] + meant[:, None] x = xt + x if length_pre_pad: x = x[..., :length_pre_pad] return x