add DPO training

GPT_SoVITS/AR/models/t2s_model.py

@@ -8,6 +8,9 @@ from AR.models.utils import (
     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
@@ -85,11 +88,104 @@ class Text2SemanticDecoder(nn.Module):
             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)
@@ -231,6 +327,7 @@ class Text2SemanticDecoder(nn.Module):
         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,
     ):
@@ -305,7 +402,7 @@ class Text2SemanticDecoder(nn.Module):
             if(idx==0):###第一次跑不能EOS否则没有了
                 logits = logits[:, :-1]  ###刨除1024终止符号的概率
             samples = sample(
-                logits[0], y, top_k=top_k, top_p=1.0, repetition_penalty=1.35
+                logits[0], y, top_k=top_k, top_p=top_p, repetition_penalty=1.05, temperature=temperature
             )[0].unsqueeze(0)
             if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
                 print("use early stop num:", early_stop_num)