GPT-SoVITS/GPT_SoVITS/AR/models/t2s_model.py

# modified from https://github.com/yangdongchao/SoundStorm/blob/master/soundstorm/s1/AR/models/t2s_model.py
# reference: https://github.com/lifeiteng/vall-e
import torch
from tqdm import tqdm

from AR.models.utils import make_pad_mask
from AR.models.utils import (
    topk_sampling,
    sample,
    logits_to_probs,
    multinomial_sample_one_no_sync,
    dpo_loss,
    make_reject_y,
    get_batch_logps
)
from AR.modules.embedding import SinePositionalEmbedding
from AR.modules.embedding import TokenEmbedding
from AR.modules.transformer import LayerNorm
from AR.modules.transformer import TransformerEncoder
from AR.modules.transformer import TransformerEncoderLayer
from torch import nn
from torch.nn import functional as F
from torchmetrics.classification import MulticlassAccuracy

try:
    from flash_attn import flash_attn_with_kvcache
except ImportError:
    flash_attn_with_kvcache = None

default_config = {
    "embedding_dim": 512,
    "hidden_dim": 512,
    "num_head": 8,
    "num_layers": 12,
    "num_codebook": 8,
    "p_dropout": 0.0,
    "vocab_size": 1024 + 1,
    "phoneme_vocab_size": 512,
    "EOS": 1024,
}


class Text2SemanticDecoder(nn.Module):
    def __init__(self, config, norm_first=False, top_k=3):
        super(Text2SemanticDecoder, self).__init__()
        self.model_dim = config["model"]["hidden_dim"]
        self.embedding_dim = config["model"]["embedding_dim"]
        self.num_head = config["model"]["head"]
        self.num_layers = config["model"]["n_layer"]
        self.norm_first = norm_first
        self.vocab_size = config["model"]["vocab_size"]
        self.phoneme_vocab_size = config["model"]["phoneme_vocab_size"]
        self.p_dropout = config["model"]["dropout"]
        self.EOS = config["model"]["EOS"]
        self.norm_first = norm_first
        assert self.EOS == self.vocab_size - 1
        # should be same as num of kmeans bin
        # assert self.EOS == 1024
        self.bert_proj = nn.Linear(1024, self.embedding_dim)
        self.ar_text_embedding = TokenEmbedding(
            self.embedding_dim, self.phoneme_vocab_size, self.p_dropout
        )
        self.ar_text_position = SinePositionalEmbedding(
            self.embedding_dim, dropout=0.1, scale=False, alpha=True
        )
        self.ar_audio_embedding = TokenEmbedding(
            self.embedding_dim, self.vocab_size, self.p_dropout
        )
        self.ar_audio_position = SinePositionalEmbedding(
            self.embedding_dim, dropout=0.1, scale=False, alpha=True
        )

        self.h = TransformerEncoder(
            TransformerEncoderLayer(
                d_model=self.model_dim,
                nhead=self.num_head,
                dim_feedforward=self.model_dim * 4,
                dropout=0.1,
                batch_first=True,
                norm_first=norm_first,
            ),
            num_layers=self.num_layers,
            norm=LayerNorm(self.model_dim) if norm_first else None,
        )

        self.ar_predict_layer = nn.Linear(self.model_dim, self.vocab_size, bias=False)
        self.loss_fct = nn.CrossEntropyLoss(reduction="sum")

        self.ar_accuracy_metric = MulticlassAccuracy(
            self.vocab_size,
            top_k=top_k,
            average="micro",
            multidim_average="global",
            ignore_index=self.EOS,
        )

    def make_input_data(self, x, x_lens, y, y_lens, bert_feature):
        x = self.ar_text_embedding(x)
        x = x + self.bert_proj(bert_feature.transpose(1, 2))
        x = self.ar_text_position(x)
        x_mask = make_pad_mask(x_lens)

        y_mask = make_pad_mask(y_lens)
        y_mask_int = y_mask.type(torch.int64)
        codes = y.type(torch.int64) * (1 - y_mask_int)

        # Training
        # AR Decoder
        y, targets = self.pad_y_eos(codes, y_mask_int, eos_id=self.EOS)
        x_len = x_lens.max()
        y_len = y_lens.max()
        y_emb = self.ar_audio_embedding(y)
        y_pos = self.ar_audio_position(y_emb)

        xy_padding_mask = torch.concat([x_mask, y_mask], dim=1)

        ar_xy_padding_mask = xy_padding_mask

        x_attn_mask = F.pad(
            torch.zeros((x_len, x_len), dtype=torch.bool, device=x.device),
            (0, y_len),
            value=True,
        )

        y_attn_mask = F.pad(
            torch.triu(
                torch.ones(y_len, y_len, dtype=torch.bool, device=x.device),
                diagonal=1,
            ),
            (x_len, 0),
            value=False,
        )

        xy_attn_mask = torch.concat([x_attn_mask, y_attn_mask], dim=0)
        bsz, src_len = x.shape[0], x_len + y_len
        _xy_padding_mask = (
            ar_xy_padding_mask.view(bsz, 1, 1, src_len)
            .expand(-1, self.num_head, -1, -1)
            .reshape(bsz * self.num_head, 1, src_len)
        )
        xy_attn_mask = xy_attn_mask.logical_or(_xy_padding_mask)
        new_attn_mask = torch.zeros_like(xy_attn_mask, dtype=x.dtype)
        new_attn_mask.masked_fill_(xy_attn_mask, float("-inf"))
        xy_attn_mask = new_attn_mask
        # x 和完整的 y 一次性输入模型
        xy_pos = torch.concat([x, y_pos], dim=1)

        return xy_pos, xy_attn_mask, targets

    def forward(self, x, x_lens, y, y_lens, bert_feature):
        """
        x: phoneme_ids
        y: semantic_ids
        """

        reject_y, reject_y_lens = make_reject_y(y, y_lens)

        xy_pos, xy_attn_mask, targets = self.make_input_data(x, x_lens, y, y_lens, bert_feature)

        xy_dec, _ = self.h(
            (xy_pos, None),
            mask=xy_attn_mask,
        )
        x_len = x_lens.max()
        logits = self.ar_predict_layer(xy_dec[:, x_len:])

        ###### DPO #############
        reject_xy_pos, reject_xy_attn_mask, reject_targets = self.make_input_data(x, x_lens, reject_y, reject_y_lens, bert_feature)

        reject_xy_dec, _ = self.h(
            (reject_xy_pos, None),
            mask=reject_xy_attn_mask,
        )
        x_len = x_lens.max()
        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len:])

        # loss
        # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum

        loss_1 = F.cross_entropy(logits.permute(0, 2, 1), targets, reduction="sum")
        acc = self.ar_accuracy_metric(logits.permute(0, 2, 1).detach(), targets).item()

        A_logits, R_logits = get_batch_logps(logits, reject_logits, targets, reject_targets)
        loss_2, _, _ = dpo_loss(A_logits, R_logits, 0, 0, 0.2, reference_free=True)

        loss = loss_1 + loss_2

        return loss, acc

    def forward_old(self, x, x_lens, y, y_lens, bert_feature):
        """
        x: phoneme_ids
        y: semantic_ids
        """
        x = self.ar_text_embedding(x)
        x = x + self.bert_proj(bert_feature.transpose(1, 2))
        x = self.ar_text_position(x)
        x_mask = make_pad_mask(x_lens)

        y_mask = make_pad_mask(y_lens)
        y_mask_int = y_mask.type(torch.int64)
        codes = y.type(torch.int64) * (1 - y_mask_int)

        # Training
        # AR Decoder
        y, targets = self.pad_y_eos(codes, y_mask_int, eos_id=self.EOS)
        x_len = x_lens.max()
        y_len = y_lens.max()
        y_emb = self.ar_audio_embedding(y)
        y_pos = self.ar_audio_position(y_emb)

        xy_padding_mask = torch.concat([x_mask, y_mask], dim=1)
        ar_xy_padding_mask = xy_padding_mask

        x_attn_mask = F.pad(
            torch.zeros((x_len, x_len), dtype=torch.bool, device=x.device),
            (0, y_len),
            value=True,
        )
        y_attn_mask = F.pad(
            torch.triu(
                torch.ones(y_len, y_len, dtype=torch.bool, device=x.device),
                diagonal=1,
            ),
            (x_len, 0),
            value=False,
        )
        xy_attn_mask = torch.concat([x_attn_mask, y_attn_mask], dim=0)
        bsz, src_len = x.shape[0], x_len + y_len
        _xy_padding_mask = (
            ar_xy_padding_mask.view(bsz, 1, 1, src_len)
            .expand(-1, self.num_head, -1, -1)
            .reshape(bsz * self.num_head, 1, src_len)
        )
        xy_attn_mask = xy_attn_mask.logical_or(_xy_padding_mask)
        new_attn_mask = torch.zeros_like(xy_attn_mask, dtype=x.dtype)
        new_attn_mask.masked_fill_(xy_attn_mask, float("-inf"))
        xy_attn_mask = new_attn_mask
        # x 和完整的 y 一次性输入模型
        xy_pos = torch.concat([x, y_pos], dim=1)
        xy_dec, _ = self.h(
            (xy_pos, None),
            mask=xy_attn_mask,
        )
        logits = self.ar_predict_layer(xy_dec[:, x_len:]).permute(0, 2, 1)
        # loss
        # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
        loss = F.cross_entropy(logits, targets, reduction="sum")
        acc = self.ar_accuracy_metric(logits.detach(), targets).item()
        return loss, acc

    # 需要看下这个函数和 forward 的区别以及没有 semantic 的时候 prompts 输入什么
    def infer(
            self,
            x,
            x_lens,
            prompts,
            bert_feature,
            top_k: int = -100,
            early_stop_num: int = -1,
            temperature: float = 1.0,
    ):
        x = self.ar_text_embedding(x)
        x = x + self.bert_proj(bert_feature.transpose(1, 2))
        x = self.ar_text_position(x)

        # AR Decoder
        y = prompts
        prefix_len = y.shape[1]
        x_len = x.shape[1]
        x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
        stop = False
        for _ in tqdm(range(1500)):
            y_emb = self.ar_audio_embedding(y)
            y_pos = self.ar_audio_position(y_emb)
            # x 和逐渐增长的 y 一起输入给模型
            xy_pos = torch.concat([x, y_pos], dim=1)
            y_len = y.shape[1]
            x_attn_mask_pad = F.pad(
                x_attn_mask,
                (0, y_len),
                value=True,
            )
            y_attn_mask = F.pad(
                torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
                (x_len, 0),
                value=False,
            )
            xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
                y.device
            )

            xy_dec, _ = self.h(
                (xy_pos, None),
                mask=xy_attn_mask,
            )
            logits = self.ar_predict_layer(xy_dec[:, -1])
            samples = topk_sampling(
                logits, top_k=top_k, top_p=1.0, temperature=temperature
            )

            if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
                print("use early stop num:", early_stop_num)
                stop = True

            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
                # print(torch.argmax(logits, dim=-1)[0] == self.EOS, samples[0, 0] == self.EOS)
                stop = True
            if stop:
                if prompts.shape[1] == y.shape[1]:
                    y = torch.concat([y, torch.zeros_like(samples)], dim=1)
                    print("bad zero prediction")
                print(f"T2S Decoding EOS [{prefix_len} -> {y.shape[1]}]")
                break
            # 本次生成的 semantic_ids 和之前的 y 构成新的 y
            # print(samples.shape)#[1,1]#第一个1是bs
            # import os
            # os._exit(2333)
            y = torch.concat([y, samples], dim=1)
        return y

    def pad_y_eos(self, y, y_mask_int, eos_id):
        targets = F.pad(y, (0, 1), value=0) + eos_id * F.pad(
            y_mask_int, (0, 1), value=1
        )
        # 错位
        return targets[:, :-1], targets[:, 1:]

    def infer_one_step(self, x, xy_attn_mask, k_cache, v_cache, cache_seqlens):
        hidden_dim = x.shape[-1]

        for layer_id in range(self.num_layers):
            layer = self.h.layers[layer_id]

            q, k, v = F.linear(
                x,
                layer.self_attn.in_proj_weight,
                layer.self_attn.in_proj_bias
            ).chunk(3, dim=-1)

            batch_size = q.shape[0]
            q_len = q.shape[1]

            if flash_attn_with_kvcache is None:
                past_k = k_cache[layer_id]
                past_v = v_cache[layer_id]

                if past_k is not None:
                    k = torch.cat([past_k, k], 1)
                    v = torch.cat([past_v, v], 1)
                k_cache[layer_id] = k
                v_cache[layer_id] = v
                kv_len = k.shape[1]

                q = q.view(batch_size, q_len, layer.self_attn.num_heads, -1).transpose(1, 2)
                k = k.view(batch_size, kv_len, layer.self_attn.num_heads, -1).transpose(1, 2)
                v = v.view(batch_size, kv_len, layer.self_attn.num_heads, -1).transpose(1, 2)

                if xy_attn_mask is None:
                    attn = F.scaled_dot_product_attention(q, k, v)
                else:
                    attn = F.scaled_dot_product_attention(q, k, v, ~xy_attn_mask)

                attn = attn.permute(2, 0, 1, 3).reshape(-1, hidden_dim)
            else:
                q = q.view(batch_size, q_len, layer.self_attn.num_heads, -1)
                k = k.view(batch_size, q_len, layer.self_attn.num_heads, -1)
                v = v.view(batch_size, q_len, layer.self_attn.num_heads, -1)

                if xy_attn_mask is None:
                    attn = flash_attn_with_kvcache(q, k_cache[layer_id], v_cache[layer_id], k, v, cache_seqlens=cache_seqlens, causal=True)
                else:
                    # NOTE: there's a slight difference with the result produced by SDPA.
                    x_len = (~xy_attn_mask).sum(1)[0].item()

                    attn_x = flash_attn_with_kvcache(
                        q[:, :x_len],
                        k_cache[layer_id],
                        v_cache[layer_id],
                        k[:, :x_len],
                        v[:, :x_len],
                        cache_seqlens=cache_seqlens,
                        causal=False
                    )

                    attn_y = flash_attn_with_kvcache(
                        q[:, x_len:],
                        k_cache[layer_id],
                        v_cache[layer_id],
                        k[:, x_len:],
                        v[:, x_len:],
                        cache_seqlens=cache_seqlens + x_len,
                        causal=True
                    )

                    attn = torch.cat([attn_x, attn_y], dim=1)
                attn = attn.view(-1, hidden_dim)

            attn_out = F.linear(attn, layer.self_attn.out_proj.weight, layer.self_attn.out_proj.bias)

            x = layer.norm1(x + attn_out, None)

            x = layer.norm2(x + layer.linear2(F.relu(layer.linear1(x))), None)

        xy_dec = x

        logits = self.ar_predict_layer(
            xy_dec[:, -1]
        )

        return logits

    def infer_panel(
            self,
            x,  #####全部文本token
            x_lens,
            prompts,  ####参考音频token
            bert_feature,
            top_k: int = -100,
            top_p: int = 100,
            early_stop_num: int = -1,
            temperature: float = 1.0,
    ):
        x = self.ar_text_embedding(x)
        x = x + self.bert_proj(bert_feature.transpose(1, 2))
        x = self.ar_text_position(x)

        # AR Decoder
        y = prompts

        x_len = x.shape[1]
        x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
        stop = False
        # print(1111111,self.num_layers)

        if flash_attn_with_kvcache is not None:
            k_cache = [torch.empty(x.shape[0], 2048, 16, 32, dtype=x.dtype, device=x.device) for _ in range(self.num_layers)]
            v_cache = [torch.empty(x.shape[0], 2048, 16, 32, dtype=x.dtype, device=x.device) for _ in range(self.num_layers)]
        else:
            k_cache = [None] * self.num_layers
            v_cache = [None] * self.num_layers
        ###################  first step ##########################
        if y is not None:
            y_emb = self.ar_audio_embedding(y)
            y_len = y_emb.shape[1]
            prefix_len = y.shape[1]
            y_pos = self.ar_audio_position(y_emb)
            xy_pos = torch.concat([x, y_pos], dim=1)
            ref_free = False
        else:
            y_emb = None
            y_len = 0
            prefix_len = 0
            y_pos = None
            xy_pos = x
            y = torch.zeros(x.shape[0], 0, dtype=torch.int, device=x.device)
            ref_free = True

        x_attn_mask_pad = F.pad(
            x_attn_mask,
            (0, y_len),  ###xx的纯0扩展到xx纯0+xy纯1，(x,x+y)
            value=True,
        )
        y_attn_mask = F.pad(  ###yy的右上1扩展到左边xy的0,(y,x+y)
            torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
            (x_len, 0),
            value=False,
        )
        xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
            x.device
        )

        cache_seqlens = torch.zeros(x.shape[0], dtype=torch.int32, device=x.device)

        for idx in tqdm(range(1500)):
            logits = self.infer_one_step(xy_pos, xy_attn_mask, k_cache, v_cache, cache_seqlens)

            if idx == 0:
                cache_seqlens += xy_pos.shape[1]
            else:
                cache_seqlens += 1
            xy_attn_mask = None

            if idx == 0:
                logits = logits[:, :-1]
            samples = sample(
                logits[0], y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
            )[0].unsqueeze(0)

            y = torch.concat([y, samples], dim=1)

            if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
                print("use early stop num:", early_stop_num)
                stop = True

            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
                stop = True
            if stop:
                if y.shape[1] == 0:
                    y = torch.concat([y, torch.zeros_like(samples)], dim=1)
                    print("bad zero prediction")
                print(f"T2S Decoding EOS [{prefix_len} -> {y.shape[1]}]")
                break

            ####################### update next step ###################################
            y_emb = self.ar_audio_embedding(y[:, -1:])
            xy_pos = y_emb * self.ar_audio_position.x_scale + self.ar_audio_position.alpha * self.ar_audio_position.pe[:, prompts.shape[1] + idx]

        if ref_free:
            return y[:, :-1], 0
        return y[:, :-1], idx - 1