add learned value residual mixing + add adopt as default

e2_tts_pytorch/e2_tts.py

@@ -560,6 +560,8 @@ def __init__(
             )
 
         for ind in range(depth):
+            is_first_block = ind == 0
+
             is_later_half = ind >= (depth // 2)
             has_text = ind < text_depth
 
@@ -596,7 +598,7 @@ def __init__(
                 text_conv = DepthwiseConv(dim_text, kernel_size = kernel_size)
 
                 text_attn_norm = RMSNorm(dim_text)
-                text_attn = Attention(dim = dim_text, heads = text_heads, dim_head = text_dim_head, dropout = dropout, **attn_kwargs)
+                text_attn = Attention(dim = dim_text, heads = text_heads, dim_head = text_dim_head, dropout = dropout, learned_value_residual_mix = not is_first_block, **attn_kwargs)
 
                 text_ff_norm = RMSNorm(dim_text)
                 text_ff = FeedForward(dim = dim_text, glu = True, mult = text_ff_mult, dropout = dropout, **ff_kwargs)

e2_tts_pytorch/trainer.py

@@ -8,6 +8,7 @@
 
 import torch
 import torch.nn.functional as F
+from torch.optim import Optimizer
 from torch.utils.data import DataLoader, Dataset
 from torch.utils.tensorboard import SummaryWriter
 from torch.optim.lr_scheduler import LinearLR, SequentialLR
@@ -19,6 +20,8 @@
 from accelerate import Accelerator
 from accelerate.utils import DistributedDataParallelKwargs
 
+from adam_atan2_pytorch.adopt import Adopt
+
 from ema_pytorch import EMA
 
 from loguru import logger
@@ -133,9 +136,10 @@ class E2Trainer:
     def __init__(
         self,
         model: E2TTS,
-        optimizer,
-        num_warmup_steps=20000,
-        grad_accumulation_steps=1,
+        optimizer: Optimizer | None = None,
+        learning_rate = 7.5e-5,
+        num_warmup_steps = 20000,
+        grad_accumulation_steps = 1,
         duration_predictor: DurationPredictor | None = None,
         checkpoint_path = None,
         log_file = "logs.txt",
@@ -172,16 +176,23 @@ def __init__(
         self.use_switch_ema = use_switch_ema
 
         self.duration_predictor = duration_predictor
+
+        # optimizer
+
+        if not exists(optimizer):
+            optimizer = Adopt(model.parameters(), lr = learning_rate)
+
         self.optimizer = optimizer
+
         self.num_warmup_steps = num_warmup_steps
-        self.checkpoint_path = default(checkpoint_path, 'model.pth')
         self.mel_spectrogram = MelSpec(sampling_rate=self.target_sample_rate)
 
         self.ema_model, self.model, self.optimizer = self.accelerator.prepare(
             self.ema_model, self.model, self.optimizer
         )
         self.max_grad_norm = max_grad_norm
 
+        self.checkpoint_path = default(checkpoint_path, 'model.pth')
         self.writer = SummaryWriter(log_dir=tensorboard_log_dir)
 
     @property

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "e2-tts-pytorch"
-version = "1.5.2"
+version = "1.6.0"
 description = "E2-TTS in Pytorch"
 authors = [
     { name = "Phil Wang", email = "[email protected]" }
@@ -25,6 +25,7 @@ classifiers=[
 
 dependencies = [
     'accelerate>=0.33.0',
+    'adam-atan2-pytorch>=0.1.12',
     'beartype',
     'einops>=0.8.0',
     'einx>=0.3.0',
@@ -39,7 +40,7 @@ dependencies = [
     'torchaudio>=2.3.1',
     'tqdm>=4.65.0',
     'vocos',
-    'x-transformers>=1.42.4',
+    'x-transformers>=1.42.16',
 ]
 
 [project.urls]

train_example.py

@@ -1,7 +1,6 @@
 import torch
 from e2_tts_pytorch import E2TTS, DurationPredictor
 
-from torch.optim import Adam
 from datasets import load_dataset
 
 from e2_tts_pytorch.trainer import (
@@ -26,8 +25,6 @@
 
 train_dataset = HFDataset(load_dataset("MushanW/GLOBE")["train"])
 
-optimizer = Adam(e2tts.parameters(), lr=7.5e-5)
-
 trainer = E2Trainer(
     e2tts,
     optimizer,