radtts.py

# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: MIT
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.
import torch
from torch import nn
from common import Encoder, LengthRegulator, ConvAttention
from common import Invertible1x1ConvLUS, Invertible1x1Conv
from common import AffineTransformationLayer, LinearNorm, ExponentialClass
from common import get_mask_from_lengths
from attribute_prediction_model import get_attribute_prediction_model
from alignment import mas_width1 as mas


class FlowStep(nn.Module):
    def __init__(self,  n_mel_channels, n_context_dim, n_layers,
                 affine_model='simple_conv', scaling_fn='exp',
                 matrix_decomposition='', affine_activation='softplus',
                 use_partial_padding=False, cache_inverse=False):
        super(FlowStep, self).__init__()
        if matrix_decomposition == 'LUS':
            self.invtbl_conv = Invertible1x1ConvLUS(n_mel_channels, cache_inverse=cache_inverse)
        else:
            self.invtbl_conv = Invertible1x1Conv(n_mel_channels, cache_inverse=cache_inverse)

        self.affine_tfn = AffineTransformationLayer(
            n_mel_channels, n_context_dim, n_layers, affine_model=affine_model,
            scaling_fn=scaling_fn, affine_activation=affine_activation,
            use_partial_padding=use_partial_padding)

    def enable_inverse_cache(self):
        self.invtbl_conv.cache_inverse=True


    def forward(self, z, context, inverse=False, seq_lens=None):
        if inverse:  # for inference z-> mel
            z = self.affine_tfn(z, context, inverse, seq_lens=seq_lens)
            z = self.invtbl_conv(z, inverse)
            return z
        else:  # training mel->z
            z, log_det_W = self.invtbl_conv(z)
            z, log_s = self.affine_tfn(z, context, seq_lens=seq_lens)
            return z, log_det_W, log_s


class RADTTS(torch.nn.Module):
    def __init__(self, n_speakers, n_speaker_dim, n_text, n_text_dim, n_flows,
                 n_conv_layers_per_step, n_mel_channels, n_hidden,
                 mel_encoder_n_hidden, dummy_speaker_embedding, n_early_size,
                 n_early_every, n_group_size, affine_model, dur_model_config,
                 f0_model_config, energy_model_config, v_model_config=None,
                 include_modules='dec', scaling_fn='exp',
                 matrix_decomposition='', learn_alignments=False,
                 affine_activation='softplus', attn_use_CTC=True,
                 use_speaker_emb_for_alignment=False, use_context_lstm=False,
                 context_lstm_norm=None, text_encoder_lstm_norm=None,
                 n_f0_dims=0, n_energy_avg_dims=0,
                 context_lstm_w_f0_and_energy=True,
                 use_first_order_features=False, unvoiced_bias_activation='',
                 ap_pred_log_f0=False, **kwargs):
        super(RADTTS, self).__init__()
        assert(n_early_size % 2 == 0)
        self.do_mel_descaling = kwargs.get('do_mel_descaling', True)
        self.n_mel_channels = n_mel_channels
        self.n_f0_dims = n_f0_dims  # >= 1 to trains with f0
        self.n_energy_avg_dims = n_energy_avg_dims  # >= 1 trains with energy
        self.decoder_use_partial_padding = kwargs.get(
            'decoder_use_partial_padding', True)
        self.n_speaker_dim = n_speaker_dim
        assert(self.n_speaker_dim % 2 == 0)
        self.speaker_embedding = torch.nn.Embedding(
            n_speakers, self.n_speaker_dim)
        self.embedding = torch.nn.Embedding(n_text, n_text_dim)
        self.flows = torch.nn.ModuleList()
        self.encoder = Encoder(encoder_embedding_dim=n_text_dim,
                               norm_fn=nn.InstanceNorm1d,
                               lstm_norm_fn=text_encoder_lstm_norm)
        self.dummy_speaker_embedding = dummy_speaker_embedding
        self.learn_alignments = learn_alignments
        self.affine_activation = affine_activation
        self.include_modules = include_modules
        self.attn_use_CTC = bool(attn_use_CTC)
        self.use_speaker_emb_for_alignment = use_speaker_emb_for_alignment
        self.use_context_lstm = bool(use_context_lstm)
        self.context_lstm_norm = context_lstm_norm
        self.context_lstm_w_f0_and_energy = context_lstm_w_f0_and_energy
        self.length_regulator = LengthRegulator()
        self.use_first_order_features = bool(use_first_order_features)
        self.decoder_use_unvoiced_bias = kwargs.get(
            'decoder_use_unvoiced_bias', True)
        self.ap_pred_log_f0 = ap_pred_log_f0
        self.ap_use_unvoiced_bias = kwargs.get('ap_use_unvoiced_bias', True)
        self.attn_straight_through_estimator = kwargs.get(
            'attn_straight_through_estimator', False)
        if 'atn' in include_modules or 'dec' in include_modules:
            if self.learn_alignments:
                if self.use_speaker_emb_for_alignment:
                    self.attention = ConvAttention(
                        n_mel_channels, n_text_dim + self.n_speaker_dim)
                else:
                    self.attention = ConvAttention(n_mel_channels, n_text_dim)

            self.n_flows = n_flows
            self.n_group_size = n_group_size

            n_flowstep_cond_dims = (
                self.n_speaker_dim +
                (n_text_dim + n_f0_dims + n_energy_avg_dims) * n_group_size)

            if self.use_context_lstm:
                n_in_context_lstm = (
                    self.n_speaker_dim + n_text_dim * n_group_size)
                n_context_lstm_hidden = int(
                    (self.n_speaker_dim + n_text_dim * n_group_size) / 2)

                if self.context_lstm_w_f0_and_energy:
                    n_in_context_lstm = (
                        n_f0_dims + n_energy_avg_dims + n_text_dim)
                    n_in_context_lstm *= n_group_size
                    n_in_context_lstm += self.n_speaker_dim

                    n_context_hidden = (
                        n_f0_dims + n_energy_avg_dims + n_text_dim)
                    n_context_hidden = n_context_hidden * n_group_size / 2
                    n_context_hidden = self.n_speaker_dim + n_context_hidden
                    n_context_hidden = int(n_context_hidden)

                    n_flowstep_cond_dims = (
                        self.n_speaker_dim + n_text_dim * n_group_size)

                self.context_lstm = torch.nn.LSTM(
                    input_size=n_in_context_lstm,
                    hidden_size=n_context_lstm_hidden, num_layers=1,
                    batch_first=True, bidirectional=True)

                if context_lstm_norm is not None:
                    if 'spectral' in context_lstm_norm:
                        print("Applying spectral norm to context encoder LSTM")
                        lstm_norm_fn_pntr = torch.nn.utils.spectral_norm
                    elif 'weight' in context_lstm_norm:
                        print("Applying weight norm to context encoder LSTM")
                        lstm_norm_fn_pntr = torch.nn.utils.weight_norm

                    self.context_lstm = lstm_norm_fn_pntr(
                        self.context_lstm, 'weight_hh_l0')
                    self.context_lstm = lstm_norm_fn_pntr(
                        self.context_lstm, 'weight_hh_l0_reverse')

            if self.n_group_size > 1:
                self.unfold_params = {'kernel_size': (n_group_size, 1),
                                      'stride': n_group_size,
                                      'padding': 0, 'dilation': 1}
                self.unfold = nn.Unfold(**self.unfold_params)

            self.exit_steps = []
            self.n_early_size = n_early_size
            n_mel_channels = n_mel_channels*n_group_size

            for i in range(self.n_flows):
                if i > 0 and i % n_early_every == 0:  # early exitting
                    n_mel_channels -= self.n_early_size
                    self.exit_steps.append(i)

                self.flows.append(FlowStep(
                    n_mel_channels, n_flowstep_cond_dims,
                    n_conv_layers_per_step, affine_model, scaling_fn,
                    matrix_decomposition, affine_activation=affine_activation,
                    use_partial_padding=self.decoder_use_partial_padding))

        if 'dpm' in include_modules:
            dur_model_config['hparams']['n_speaker_dim'] = n_speaker_dim
            self.dur_pred_layer = get_attribute_prediction_model(
                dur_model_config)

        self.use_unvoiced_bias = False
        self.use_vpred_module = False
        self.ap_use_voiced_embeddings = kwargs.get(
            'ap_use_voiced_embeddings', True)

        if self.decoder_use_unvoiced_bias or self.ap_use_unvoiced_bias:
            assert (unvoiced_bias_activation in {'relu', 'exp'})
            self.use_unvoiced_bias = True
            if unvoiced_bias_activation == 'relu':
                unvbias_nonlin = nn.ReLU()
            elif unvoiced_bias_activation == 'exp':
                unvbias_nonlin = ExponentialClass()
            else:
                exit(1)  # we won't reach here anyway due to the assertion
            self.unvoiced_bias_module = nn.Sequential(
                LinearNorm(n_text_dim, 1), unvbias_nonlin)

        # all situations in which the vpred module is necessary
        if self.ap_use_voiced_embeddings or self.use_unvoiced_bias or 'vpred' in include_modules:
            self.use_vpred_module = True

        if self.use_vpred_module:
            v_model_config['hparams']['n_speaker_dim'] = n_speaker_dim
            self.v_pred_module = get_attribute_prediction_model(v_model_config)
            # 4 embeddings, first two are scales, second two are biases
            if self.ap_use_voiced_embeddings:
                self.v_embeddings = torch.nn.Embedding(4, n_text_dim)

        if 'apm' in include_modules:
            f0_model_config['hparams']['n_speaker_dim'] = n_speaker_dim
            energy_model_config['hparams']['n_speaker_dim'] = n_speaker_dim
            if self.use_first_order_features:
                f0_model_config['hparams']['n_in_dim'] = 2
                energy_model_config['hparams']['n_in_dim'] = 2
                if 'spline_flow_params' in f0_model_config['hparams'] and f0_model_config['hparams']['spline_flow_params'] is not None:
                    f0_model_config['hparams']['spline_flow_params']['n_in_channels'] = 2
                if 'spline_flow_params' in energy_model_config['hparams'] and energy_model_config['hparams']['spline_flow_params'] is not None:
                    energy_model_config['hparams']['spline_flow_params']['n_in_channels'] = 2
            else:
                if 'spline_flow_params' in f0_model_config['hparams'] and f0_model_config['hparams']['spline_flow_params'] is not None:
                    f0_model_config['hparams']['spline_flow_params']['n_in_channels'] = f0_model_config['hparams']['n_in_dim']
                if 'spline_flow_params' in energy_model_config['hparams'] and energy_model_config['hparams']['spline_flow_params'] is not None:
                    energy_model_config['hparams']['spline_flow_params']['n_in_channels'] = energy_model_config['hparams']['n_in_dim']

            self.f0_pred_module = get_attribute_prediction_model(
                f0_model_config)
            self.energy_pred_module = get_attribute_prediction_model(
                energy_model_config)

    def is_attribute_unconditional(self):
        """
        returns true if the decoder is conditioned on neither energy nor F0
        """
        return self.n_f0_dims == 0 and self.n_energy_avg_dims == 0

    def encode_speaker(self, spk_ids):
        spk_ids = spk_ids * 0 if self.dummy_speaker_embedding else spk_ids
        spk_vecs = self.speaker_embedding(spk_ids)
        return spk_vecs

    def encode_text(self, text, in_lens):
        # text_embeddings: b x len_text x n_text_dim
        text_embeddings = self.embedding(text).transpose(1, 2)
        # text_enc: b x n_text_dim x encoder_dim (512)
        if in_lens is None:
            text_enc = self.encoder.infer(text_embeddings).transpose(1, 2)
        else:
            text_enc = self.encoder(text_embeddings, in_lens).transpose(1, 2)

        return text_enc, text_embeddings

    def preprocess_context(self, context, speaker_vecs, out_lens=None, f0=None,
                           energy_avg=None):

        if self.n_group_size > 1:
            # unfolding zero-padded values
            context = self.unfold(context.unsqueeze(-1))
            if f0 is not None:
                f0 = self.unfold(f0[:, None, :, None])
            if energy_avg is not None:
                energy_avg = self.unfold(energy_avg[:, None, :, None])
        speaker_vecs = speaker_vecs[..., None].expand(-1, -1, context.shape[2])
        context_w_spkvec = torch.cat((context, speaker_vecs), 1)

        if self.use_context_lstm:
            if self.context_lstm_w_f0_and_energy:
                if f0 is not None:
                    context_w_spkvec = torch.cat((context_w_spkvec, f0), 1)

                if energy_avg is not None:
                    context_w_spkvec = torch.cat(
                        (context_w_spkvec, energy_avg), 1)

            unfolded_out_lens = (out_lens // self.n_group_size).long().cpu()
            unfolded_out_lens_packed = nn.utils.rnn.pack_padded_sequence(
                context_w_spkvec.transpose(1, 2), unfolded_out_lens,
                batch_first=True, enforce_sorted=False)
            self.context_lstm.flatten_parameters()
            context_lstm_packed_output, _ = self.context_lstm(
                unfolded_out_lens_packed)
            context_lstm_padded_output, _ = nn.utils.rnn.pad_packed_sequence(
                context_lstm_packed_output, batch_first=True)
            context_w_spkvec = context_lstm_padded_output.transpose(1, 2)

        if not self.context_lstm_w_f0_and_energy:
            if f0 is not None:
                context_w_spkvec = torch.cat((context_w_spkvec, f0), 1)

            if energy_avg is not None:
                context_w_spkvec = torch.cat((context_w_spkvec, energy_avg), 1)

        return context_w_spkvec

    def enable_inverse_cache(self):
        for flow_step in self.flows:
            flow_step.enable_inverse_cache()

    def fold(self, mel):
        """Inverse of the self.unfold(mel.unsqueeze(-1)) operation used for the
        grouping or "squeeze" operation on input

        Args:
            mel: B x C x T tensor of temporal data
        """
        mel = nn.functional.fold(
            mel, output_size=(mel.shape[2]*self.n_group_size, 1),
            **self.unfold_params).squeeze(-1)
        return mel

    def binarize_attention(self, attn, in_lens, out_lens):
        """For training purposes only. Binarizes attention with MAS. These will
        no longer recieve a gradient
        Args:
            attn: B x 1 x max_mel_len x max_text_len
        """
        b_size = attn.shape[0]
        with torch.no_grad():
            attn_cpu = attn.data.cpu().numpy()
            attn_out = torch.zeros_like(attn)
            for ind in range(b_size):
                hard_attn = mas(attn_cpu[ind, 0, :out_lens[ind], :in_lens[ind]])
                attn_out[ind, 0, :out_lens[ind], :in_lens[ind]] = torch.tensor(
                    hard_attn, device=attn.get_device())
        return attn_out

    def get_first_order_features(self, feats, out_lens, dilation=1):
        """
        feats: b x max_length
        out_lens: b-dim
        """
        # add an extra column
        feats_extended_R = torch.cat(
            (feats, torch.zeros_like(feats[:, 0:dilation])), dim=1)
        feats_extended_L = torch.cat(
            (torch.zeros_like(feats[:, 0:dilation]), feats), dim=1)
        dfeats_R = feats_extended_R[:, dilation:] - feats
        dfeats_L = feats - feats_extended_L[:, 0:-dilation]

        return (dfeats_R + dfeats_L) * 0.5

    def apply_voice_mask_to_text(self, text_enc, voiced_mask):
        """
        text_enc: b x C x N
        voiced_mask: b x N
        """
        voiced_mask = voiced_mask.unsqueeze(1)
        voiced_embedding_s = self.v_embeddings.weight[0:1, :, None]
        unvoiced_embedding_s = self.v_embeddings.weight[1:2, :, None]
        voiced_embedding_b = self.v_embeddings.weight[2:3, :, None]
        unvoiced_embedding_b = self.v_embeddings.weight[3:4, :, None]
        scale = torch.sigmoid(voiced_embedding_s*voiced_mask + unvoiced_embedding_s*(1-voiced_mask))
        bias = 0.1*torch.tanh(voiced_embedding_b*voiced_mask + unvoiced_embedding_b*(1-voiced_mask))
        return text_enc*scale+bias

    def forward(self, mel, speaker_ids, text, in_lens, out_lens,
                binarize_attention=False, attn_prior=None,
                f0=None, energy_avg=None, voiced_mask=None, p_voiced=None):
        speaker_vecs = self.encode_speaker(speaker_ids)
        text_enc, text_embeddings = self.encode_text(text, in_lens)

        log_s_list, log_det_W_list, z_mel = [], [], []
        attn = None
        attn_soft = None
        attn_hard = None
        if 'atn' in self.include_modules or 'dec' in self.include_modules:
            # make sure to do the alignments before folding
            attn_mask = get_mask_from_lengths(in_lens)[..., None] == 0

            text_embeddings_for_attn = text_embeddings
            if self.use_speaker_emb_for_alignment:
                speaker_vecs_expd = speaker_vecs[:, :, None].expand(
                    -1, -1, text_embeddings.shape[2])
                text_embeddings_for_attn = torch.cat(
                    (text_embeddings_for_attn, speaker_vecs_expd.detach()), 1)

            # attn_mask shld be 1 for unsd t-steps in text_enc_w_spkvec tensor
            attn_soft, attn_logprob = self.attention(
                mel, text_embeddings_for_attn, out_lens, attn_mask,
                key_lens=in_lens, attn_prior=attn_prior)

            if binarize_attention:
                attn = self.binarize_attention(attn_soft, in_lens, out_lens)
                attn_hard = attn
                if self.attn_straight_through_estimator:
                    attn_hard = attn_soft + (attn_hard - attn_soft).detach()
            else:
                attn = attn_soft

            context = torch.bmm(text_enc, attn.squeeze(1).transpose(1, 2))

        f0_bias = 0
        # unvoiced bias forward pass
        if self.use_unvoiced_bias:
            f0_bias = self.unvoiced_bias_module(context.permute(0, 2, 1))
            f0_bias = -f0_bias[..., 0]
            f0_bias = f0_bias * (~voiced_mask.bool()).float()

        # mel decoder forward pass
        if 'dec' in self.include_modules:
            if self.n_group_size > 1:
                # might truncate some frames at the end, but that's ok
                # sometimes referred to as the "squeeeze" operation
                # invert this by calling self.fold(mel_or_z)
                mel = self.unfold(mel.unsqueeze(-1))
            z_out = []
            # where context is folded
            # mask f0 in case values are interpolated

            if f0 is None:
                f0_aug = None
            else:
                if self.decoder_use_unvoiced_bias:
                    f0_aug = f0 * voiced_mask + f0_bias
                else:
                    f0_aug = f0 * voiced_mask

            context_w_spkvec = self.preprocess_context(
                    context, speaker_vecs, out_lens, f0_aug,
                    energy_avg)

            log_s_list, log_det_W_list, z_out = [], [], []
            unfolded_seq_lens = out_lens//self.n_group_size
            for i, flow_step in enumerate(self.flows):
                if i in self.exit_steps:
                    z = mel[:, :self.n_early_size]
                    z_out.append(z)
                    mel = mel[:, self.n_early_size:]
                mel, log_det_W, log_s = flow_step(
                    mel, context_w_spkvec, seq_lens=unfolded_seq_lens)
                log_s_list.append(log_s)
                log_det_W_list.append(log_det_W)

            z_out.append(mel)
            z_mel = torch.cat(z_out, 1)

        # duration predictor forward pass
        duration_model_outputs = None
        if 'dpm' in self.include_modules:
            if attn_hard is None:
                attn_hard = self.binarize_attention(
                    attn_soft, in_lens, out_lens)

            # convert hard attention to durations
            attn_hard_reduced = attn_hard.sum(2)[:, 0, :]
            duration_model_outputs = self.dur_pred_layer(
                torch.detach(text_enc),
                torch.detach(speaker_vecs),
                torch.detach(attn_hard_reduced.float()), in_lens)

        # f0, energy, vpred predictors forward pass
        f0_model_outputs = None
        energy_model_outputs = None
        vpred_model_outputs = None
        if 'apm' in self.include_modules:
            if attn_hard is None:
                attn_hard = self.binarize_attention(
                    attn_soft, in_lens, out_lens)

            # convert hard attention to durations
            if binarize_attention:
                text_enc_time_expanded = context.clone()
            else:
                text_enc_time_expanded = torch.bmm(
                    text_enc, attn_hard.squeeze(1).transpose(1, 2))

            if self.use_vpred_module:
                # unvoiced bias requires  voiced mask prediction
                vpred_model_outputs = self.v_pred_module(
                    torch.detach(text_enc_time_expanded),
                    torch.detach(speaker_vecs),
                    torch.detach(voiced_mask), out_lens)

                # affine transform context using voiced mask
                if self.ap_use_voiced_embeddings:
                    text_enc_time_expanded = self.apply_voice_mask_to_text(
                        text_enc_time_expanded, voiced_mask)

            # whether to use the unvoiced bias in the attribute predictor
            # circumvent in-place modification
            f0_target = f0.clone()
            if self.ap_use_unvoiced_bias:
                f0_target = torch.detach(f0_target * voiced_mask + f0_bias)
            else:
                f0_target = torch.detach(f0_target)

            # fit to log f0 in f0 predictor
            f0_target[voiced_mask.bool()] = torch.log(
                f0_target[voiced_mask.bool()])
            f0_target = f0_target / 6  # scale to ~ [0, 1] in log space
            energy_avg = energy_avg * 2 - 1  # scale to ~ [-1, 1]

            if self.use_first_order_features:
                df0 = self.get_first_order_features(f0_target, out_lens)
                denergy_avg = self.get_first_order_features(
                    energy_avg, out_lens)

                f0_voiced = torch.cat(
                    (f0_target[:, None],  df0[:, None]), dim=1)
                energy_avg = torch.cat(
                        (energy_avg[:, None], denergy_avg[:, None]), dim=1)

                f0_voiced = f0_voiced * 3  # scale to ~ 1 std
                energy_avg = energy_avg * 3  # scale to ~ 1 std
            else:
                f0_voiced = f0_target * 2  # scale to ~ 1 std
                energy_avg = energy_avg * 1.4  # scale to ~ 1 std

            f0_model_outputs = self.f0_pred_module(
                    text_enc_time_expanded, torch.detach(speaker_vecs),
                    f0_voiced, out_lens)

            energy_model_outputs = self.energy_pred_module(
                    text_enc_time_expanded, torch.detach(speaker_vecs),
                    energy_avg, out_lens)

        outputs = {'z_mel': z_mel,
                   'log_det_W_list': log_det_W_list,
                   'log_s_list': log_s_list,
                   'duration_model_outputs': duration_model_outputs,
                   'f0_model_outputs': f0_model_outputs,
                   'energy_model_outputs': energy_model_outputs,
                   'vpred_model_outputs': vpred_model_outputs,
                   'attn_soft': attn_soft,
                   'attn': attn,
                   'text_embeddings': text_embeddings,
                   'attn_logprob': attn_logprob
                   }

        return outputs

    def infer(self, speaker_id, text, sigma, sigma_dur=0.8, sigma_f0=0.8,
              sigma_energy=0.8, token_dur_scaling=1.0, token_duration_max=100,
              speaker_id_text=None, speaker_id_attributes=None, dur=None,
              f0=None, energy_avg=None, voiced_mask=None, f0_mean=0.0,
              f0_std=0.0, energy_mean=0.0, energy_std=0.0):
        batch_size = text.shape[0]
        n_tokens = text.shape[1]
        spk_vec = self.encode_speaker(speaker_id)
        spk_vec_text, spk_vec_attributes = spk_vec, spk_vec
        if speaker_id_text is not None:
            spk_vec_text = self.encode_speaker(speaker_id_text)
        if speaker_id_attributes is not None:
            spk_vec_attributes = self.encode_speaker(speaker_id_attributes)

        txt_enc, txt_emb = self.encode_text(text, None)

        if dur is None:
            # get token durations
            z_dur = torch.cuda.FloatTensor(batch_size, 1, n_tokens)
            z_dur = z_dur.normal_() * sigma_dur

            dur = self.dur_pred_layer.infer(z_dur, txt_enc, spk_vec_text)
            if dur.shape[-1] < txt_enc.shape[-1]:
                to_pad = txt_enc.shape[-1] - dur.shape[2]
                pad_fn = nn.ReplicationPad1d((0, to_pad))
                dur = pad_fn(dur)
            dur = dur[:, 0]
            dur = dur.clamp(0, token_duration_max)
            dur = dur * token_dur_scaling if token_dur_scaling > 0 else dur
            dur = (dur + 0.5).floor().int()

        out_lens = dur.sum(1).long().cpu() if dur.shape[0] != 1 else [dur.sum(1)]
        max_n_frames = max(out_lens)

        out_lens = torch.LongTensor(out_lens).to(txt_enc.device)

        # get attributes f0, energy, vpred, etc)
        txt_enc_time_expanded = self.length_regulator(
            txt_enc.transpose(1, 2), dur).transpose(1, 2)

        if not self.is_attribute_unconditional():
            # if explicitly modeling attributes
            if voiced_mask is None:
                if self.use_vpred_module:
                    # get logits
                    voiced_mask = self.v_pred_module.infer(
                        None, txt_enc_time_expanded, spk_vec_attributes)
                    voiced_mask = (torch.sigmoid(voiced_mask[:, 0]) > 0.5)
                    voiced_mask = voiced_mask.float()

            ap_txt_enc_time_expanded = txt_enc_time_expanded
            # voice mask augmentation only used for attribute prediction
            if self.ap_use_voiced_embeddings:
                ap_txt_enc_time_expanded = self.apply_voice_mask_to_text(
                    txt_enc_time_expanded, voiced_mask)

            f0_bias = 0
            # unvoiced bias forward pass
            if self.use_unvoiced_bias:
                f0_bias = self.unvoiced_bias_module(
                    txt_enc_time_expanded.permute(0, 2, 1))
                f0_bias = -f0_bias[..., 0]
                f0_bias = f0_bias * (~voiced_mask.bool()).float()

            if f0 is None:
                n_f0_feature_channels = 2 if self.use_first_order_features else 1
                z_f0 = torch.cuda.FloatTensor(
                    batch_size, n_f0_feature_channels, max_n_frames).normal_() * sigma_f0
                f0 = self.infer_f0(
                    z_f0, ap_txt_enc_time_expanded, spk_vec_attributes,
                    voiced_mask, out_lens)[:, 0]

            if f0_mean > 0.0:
                vmask_bool = voiced_mask.bool()
                f0_mu, f0_sigma = f0[vmask_bool].mean(), f0[vmask_bool].std()
                f0[vmask_bool] = (f0[vmask_bool] - f0_mu) / f0_sigma
                f0_std = f0_std if f0_std > 0 else f0_sigma
                f0[vmask_bool] = f0[vmask_bool] * f0_std + f0_mean

            if energy_avg is None:
                n_energy_feature_channels = 2 if self.use_first_order_features else 1
                z_energy_avg = torch.cuda.FloatTensor(
                    batch_size, n_energy_feature_channels, max_n_frames).normal_() * sigma_energy
                energy_avg = self.infer_energy(
                    z_energy_avg, ap_txt_enc_time_expanded, spk_vec, out_lens)[:, 0]

            # replication pad, because ungrouping with different group sizes
            # may lead to mismatched lengths
            if energy_avg.shape[1] < out_lens[0]:
                to_pad = out_lens[0] - energy_avg.shape[1]
                pad_fn = nn.ReplicationPad1d((0, to_pad))
                f0 = pad_fn(f0[None])[0]
                energy_avg = pad_fn(energy_avg[None])[0]
            if f0.shape[1] < out_lens[0]:
                to_pad = out_lens[0] - f0.shape[1]
                pad_fn = nn.ReplicationPad1d((0, to_pad))
                f0 = pad_fn(f0[None])[0]

            if self.decoder_use_unvoiced_bias:
                context_w_spkvec = self.preprocess_context(
                    txt_enc_time_expanded, spk_vec, out_lens,
                    f0 * voiced_mask + f0_bias, energy_avg)
            else:
                context_w_spkvec = self.preprocess_context(
                    txt_enc_time_expanded, spk_vec, out_lens, f0*voiced_mask,
                    energy_avg)
        else:
            context_w_spkvec = self.preprocess_context(
                txt_enc_time_expanded, spk_vec, out_lens, None,
                None)

        residual = torch.cuda.FloatTensor(
            batch_size, 80 * self.n_group_size, max_n_frames // self.n_group_size)

        residual = residual.normal_() * sigma

        # map from z sample to data
        exit_steps_stack = self.exit_steps.copy()
        mel = residual[:, len(exit_steps_stack) * self.n_early_size:]
        remaining_residual = residual[:, :len(exit_steps_stack)*self.n_early_size]
        unfolded_seq_lens = out_lens//self.n_group_size
        for i, flow_step in enumerate(reversed(self.flows)):
            curr_step = len(self.flows) - i - 1
            mel = flow_step(mel, context_w_spkvec, inverse=True, seq_lens=unfolded_seq_lens)
            if len(exit_steps_stack) > 0 and curr_step == exit_steps_stack[-1]:
                # concatenate the next chunk of z
                exit_steps_stack.pop()
                residual_to_add = remaining_residual[
                    :, len(exit_steps_stack)*self.n_early_size:]
                remaining_residual = remaining_residual[
                    :, :len(exit_steps_stack)*self.n_early_size]
                mel = torch.cat((residual_to_add, mel), 1)

        if self.n_group_size > 1:
            mel = self.fold(mel)
        if self.do_mel_descaling:
            mel = mel * 2 - 5.5

        return {'mel': mel,
                'dur': dur,
                'f0': f0,
                'energy_avg': energy_avg,
                'voiced_mask': voiced_mask
                }

    def infer_f0(self, residual, txt_enc_time_expanded, spk_vec,
                 voiced_mask=None, lens=None):

        f0 = self.f0_pred_module.infer(
                residual, txt_enc_time_expanded, spk_vec, lens)

        if voiced_mask is not None and len(voiced_mask.shape) == 2:
            voiced_mask = voiced_mask[:, None]

        # constants
        if self.ap_pred_log_f0:
            if self.use_first_order_features:
                f0 = f0[:,0:1,:] / 3
            else:
                f0 = f0 / 2
            f0 = f0 * 6
        else:
            f0 = f0 / 6
            f0 = f0 / 640

        if voiced_mask is None:
            voiced_mask = f0 > 0.0
        else:
            voiced_mask = voiced_mask.bool()

        # due to grouping, f0 might be 1 frame short
        voiced_mask = voiced_mask[:,:,:f0.shape[-1]]
        if self.ap_pred_log_f0:
            # if variable is set, decoder sees linear f0
            # mask = f0 > 0.0 if voiced_mask is None else voiced_mask.bool()
            f0[voiced_mask] = torch.exp(f0[voiced_mask])
        f0[~voiced_mask] = 0.0
        return f0

    def infer_energy(self, residual, txt_enc_time_expanded, spk_vec, lens):
        energy = self.energy_pred_module.infer(
                residual, txt_enc_time_expanded, spk_vec, lens)

        # magic constants
        if self.use_first_order_features:
            energy = energy / 3
        else:
            energy = energy / 1.4
        energy = (energy + 1) / 2
        return energy

    def remove_norms(self):
        """Removes spectral and weightnorms from model. Call before inference
        """
        for name, module in self.named_modules():
            try:
                nn.utils.remove_spectral_norm(module, name='weight_hh_l0')
                print("Removed spectral norm from {}".format(name))
            except:
                pass
            try:
                nn.utils.remove_spectral_norm(module, name='weight_hh_l0_reverse')
                print("Removed spectral norm from {}".format(name))
            except:
                pass
            try:
                nn.utils.remove_weight_norm(module)
                print("Removed wnorm from {}".format(name))
            except:
                pass