Merge pull request #75 from frodo821/master

モデルのマージに球面線形補間を使えるようにした

webui_merge.py

@@ -102,6 +102,25 @@ def merge_style(model_name_a, model_name_b, weight, output_name, style_triple_li
     return output_style_path, list(new_style2id.keys())
 
 
+def lerp_tensors(t, v0, v1):
+    return v0 * (1 - t) + v1 * t
+
+def slerp_tensors(t, v0, v1, dot_thres = 0.998):
+    device = v0.device
+    v0c = v0.cpu().numpy()
+    v1c = v1.cpu().numpy()
+
+    dot = np.sum(v0c * v1c / (np.linalg.norm(v0c) * np.linalg.norm(v1c)))
+
+    if abs(dot) > dot_thres:
+        return lerp_tensors(t, v0, v1)
+    
+    th0 = np.arccos(dot)
+    sin_th0 = np.sin(th0)
+    th_t = th0 * t
+
+    return torch.from_numpy(v0c * np.sin(th0 - th_t) / sin_th0 + v1c * np.sin(th_t) / sin_th0).to(device)
+
 def merge_models(
     model_path_a,
     model_path_b,
@@ -110,6 +129,7 @@ def merge_models(
     speech_style_weight,
     tempo_weight,
     output_name,
+    use_slerp_instead_of_lerp,
 ):
     """model Aを起点に、model Bの各要素を重み付けしてマージする。
     safetensors形式を前提とする。"""
@@ -137,8 +157,8 @@ def merge_models(
         else:
             continue
         merged_model_weight[key] = (
-            model_a_weight[key] * (1 - weight) + model_b_weight[key] * weight
-        )
+            slerp_tensors if use_slerp_instead_of_lerp else lerp_tensors
+        )(weight, model_a_weight[key], model_b_weight[key])
 
     merged_model_path = os.path.join(
         assets_root, output_name, f"{output_name}.safetensors"
@@ -171,6 +191,7 @@ def merge_models_gr(
     voice_pitch_weight,
     speech_style_weight,
     tempo_weight,
+    use_slerp_instead_of_lerp,
 ):
     if output_name == "":
         return "Error: 新しいモデル名を入力してください。"
@@ -182,6 +203,7 @@ def merge_models_gr(
         speech_style_weight,
         tempo_weight,
         output_name,
+        use_slerp_instead_of_lerp,
     )
     return f"Success: モデルを{merged_model_path}に保存しました。"
 
@@ -246,7 +268,12 @@ def load_styles_gr(model_name_a, model_name_b):
     with open(config_path_b, encoding="utf-8") as f:
         config_b = json.load(f)
     styles_b = list(config_b["data"]["style2id"].keys())
-    return gr.Textbox(value=", ".join(styles_a)), gr.Textbox(value=", ".join(styles_b))
+
+    return gr.Textbox(value=", ".join(styles_a)), gr.Textbox(value=", ".join(styles_b)), gr.TextArea(
+        label="スタイルのマージリスト",
+        placeholder=f"{DEFAULT_STYLE}, {DEFAULT_STYLE},{DEFAULT_STYLE}\nAngry, Angry, Angry",
+        value='\n'.join(f"{sty_a}, {sty_b}, {sty_a if sty_a != sty_b else ''}{sty_b}" for sty_a in styles_a for sty_b in styles_b),
+    )
 
 
 initial_md = """
@@ -359,6 +386,10 @@ with gr.Blocks(theme=GRADIO_THEME) as app:
                 maximum=1,
                 step=0.1,
             )
+            use_slerp_instead_of_lerp = gr.Checkbox(
+                label="lerpのかわりにslerpを使う",
+                value=False,
+            )
         with gr.Column(variant="panel"):
             gr.Markdown("## モデルファイル（safetensors）のマージ")
             model_merge_button = gr.Button("モデルファイルのマージ", variant="primary")
@@ -414,7 +445,7 @@ with gr.Blocks(theme=GRADIO_THEME) as app:
     load_style_button.click(
         load_styles_gr,
         inputs=[model_name_a, model_name_b],
-        outputs=[styles_a, styles_b],
+        outputs=[styles_a, styles_b, style_triple_list],
     )
 
     model_merge_button.click(
@@ -429,6 +460,7 @@ with gr.Blocks(theme=GRADIO_THEME) as app:
             voice_pitch_slider,
             speech_style_slider,
             tempo_slider,
+            use_slerp_instead_of_lerp,
         ],
         outputs=[info_model_merge],
     )