From d3d5e385395f5c7b1b7ac880bf2c99bfb3eed696 Mon Sep 17 00:00:00 2001
From: Phil Wang <lucidrains@gmail.com>
Date: Thu, 31 Oct 2024 07:43:15 -0700
Subject: [PATCH] fix an issue with g2p and pydantic, and also add value
 residual learning paper from iclr 2025

---
 README.md                |  9 +++++++++
 e2_tts_pytorch/e2_tts.py | 14 ++++++++++++--
 pyproject.toml           |  5 +++--
 3 files changed, 24 insertions(+), 4 deletions(-)

diff --git a/README.md b/README.md
index b2fe95d..4ba5310 100644
--- a/README.md
+++ b/README.md
@@ -146,3 +146,12 @@ sampled = e2tts.sample(mel[:, :5], text = text)
     url     = {https://api.semanticscholar.org/CorpusID:267657558}
 }
 ```
+
+```bibtex
+@inproceedings{Zhou2024ValueRL,
+    title   = {Value Residual Learning For Alleviating Attention Concentration In Transformers},
+    author  = {Zhanchao Zhou and Tianyi Wu and Zhiyun Jiang and Zhenzhong Lan},
+    year    = {2024},
+    url     = {https://api.semanticscholar.org/CorpusID:273532030}
+}
+```
diff --git a/e2_tts_pytorch/e2_tts.py b/e2_tts_pytorch/e2_tts.py
index 91e3971..5a6bc3e 100644
--- a/e2_tts_pytorch/e2_tts.py
+++ b/e2_tts_pytorch/e2_tts.py
@@ -673,6 +673,11 @@ def forward(
 
         skips = []
 
+        # value residual
+
+        text_attn_first_values = None
+        attn_first_values = None
+
         # go through the layers
 
         for ind, (speech_modules, text_modules) in enumerate(self.layers):
@@ -704,7 +709,10 @@ def forward(
 
                 text_embed = text_conv(text_embed, mask = mask) + text_embed
 
-                text_embed = text_attn(text_attn_norm(text_embed), rotary_pos_emb = text_rotary_pos_emb, mask = mask) + text_embed
+                text_attn_out, text_attn_inter = text_attn(text_attn_norm(text_embed), rotary_pos_emb = text_rotary_pos_emb, mask = mask, return_intermediates = True, value_residual = text_attn_first_values)
+                text_embed = text_attn_out + text_embed
+
+                text_attn_first_values = default(text_attn_first_values, text_attn_inter.values)
 
                 text_embed = text_ff(text_ff_norm(text_embed)) + text_embed
 
@@ -729,7 +737,9 @@ def forward(
 
             # attention and feedforward blocks
 
-            attn_out = attn(attn_norm(x, **norm_kwargs), rotary_pos_emb = rotary_pos_emb, mask = mask)
+            attn_out, attn_inter = attn(attn_norm(x, **norm_kwargs), rotary_pos_emb = rotary_pos_emb, mask = mask, return_intermediates = True, value_residual = attn_first_values)
+
+            attn_first_values = default(attn_first_values, attn_inter.values)
 
             x = x + maybe_attn_adaln_zero(attn_out, **norm_kwargs)
 
diff --git a/pyproject.toml b/pyproject.toml
index f20a644..58778e4 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "e2-tts-pytorch"
-version = "1.2.4"
+version = "1.4.1"
 description = "E2-TTS in Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -32,13 +32,14 @@ dependencies = [
     'g2p-en',
     'jaxtyping',
     'loguru',
+    'pydantic<2',
     'tensorboard',
     'torch>=2.0',
     'torchdiffeq',
     'torchaudio>=2.3.1',
     'tqdm>=4.65.0',
     'vocos',
-    'x-transformers>=1.31.14',
+    'x-transformers>=1.42.4',
 ]
 
 [project.urls]