Add LoRA+ to optimizer params

lycoris/kohya/__init__.py

@@ -147,6 +147,15 @@ def create_network(
         # dylora didn't support scale weight norm yet
         delattr(type(network), "apply_max_norm_regularization")
 
+    loraplus_lr_ratio = kwargs.get("loraplus_lr_ratio", None)
+    loraplus_unet_lr_ratio = kwargs.get("loraplus_unet_lr_ratio", None)
+    loraplus_text_encoder_lr_ratio = kwargs.get("loraplus_text_encoder_lr_ratio", None)
+    loraplus_lr_ratio = float(loraplus_lr_ratio) if loraplus_lr_ratio is not None else None
+    loraplus_unet_lr_ratio = float(loraplus_unet_lr_ratio) if loraplus_unet_lr_ratio is not None else None
+    loraplus_text_encoder_lr_ratio = float(loraplus_text_encoder_lr_ratio) if loraplus_text_encoder_lr_ratio is not None else None
+    if loraplus_lr_ratio is not None or loraplus_unet_lr_ratio is not None or loraplus_text_encoder_lr_ratio is not None:
+        network.set_loraplus_lr_ratio(loraplus_lr_ratio, loraplus_unet_lr_ratio, loraplus_text_encoder_lr_ratio)
+
     return network
 
 
@@ -320,6 +329,10 @@ def __init__(
         self.rank_dropout = rank_dropout
         self.module_dropout = module_dropout
 
+        self.loraplus_lr_ratio = None
+        self.loraplus_unet_lr_ratio = None
+        self.loraplus_text_encoder_lr_ratio = None
+
         self.use_tucker = use_tucker
 
         def create_single_module(
@@ -572,29 +585,80 @@ def merge_to(self, text_encoder, unet, weights_sd, dtype, device):
         self.loras = self.text_encoder_loras + self.unet_loras
         super().merge_to(1)
 
-    def prepare_optimizer_params(self, text_encoder_lr, unet_lr, learning_rate):
-        def enumerate_params(loras):
-            params = []
-            for lora in loras:
-                params.extend(lora.parameters())
-            return params
+    def set_loraplus_lr_ratio(self, loraplus_lr_ratio, loraplus_unet_lr_ratio, loraplus_text_encoder_lr_ratio):
+        self.loraplus_lr_ratio = loraplus_lr_ratio
+        self.loraplus_unet_lr_ratio = loraplus_unet_lr_ratio
+        self.loraplus_text_encoder_lr_ratio = loraplus_text_encoder_lr_ratio
 
+    def prepare_optimizer_params(self, text_encoder_lr, unet_lr, default_lr):
         self.requires_grad_(True)
+
         all_params = []
+        lr_descriptions = []
+
+        def assemble_params(loras, lr, ratio):
+            param_groups = {"lora": {}, "plus": {}}
+            for lora in loras:
+                for name, param in lora.named_parameters():
+                    if ratio is not None and (
+                        "lora_up" in name or
+                        # LoHa
+                        "hada_w1_b" in name or
+                        "hada_w2_b" in name or
+                        # Lokr
+                        "lokr_w2" in name or
+                        "lokr_w1_b" in name or
+                        "lokr_w2_b" in name or
+                        # GLora
+                        name == "b1.weight" or
+                        name == "b2.weight"
+                    ):
+                        param_groups["plus"][f"{lora.lora_name}.{name}"] = param
+                    else:
+                        param_groups["lora"][f"{lora.lora_name}.{name}"] = param
+
+            params = []
+            descriptions = []
+            for key in param_groups.keys():
+                param_data = {"params": param_groups[key].values()}
+
+                if len(param_data["params"]) == 0:
+                    continue
+
+                if lr is not None:
+                    if key == "plus":
+                        param_data["lr"] = lr * ratio
+                    else:
+                        param_data["lr"] = lr
+
+                if param_data.get("lr", None) == 0 or param_data.get("lr", None) is None:
+                    logger.info("NO LR skipping!")
+                    continue
+
+                params.append(param_data)
+                descriptions.append("plus" if key == "plus" else "")
+
+            return params, descriptions
 
         if self.text_encoder_loras:
-            param_data = {"params": enumerate_params(self.text_encoder_loras)}
-            if text_encoder_lr is not None:
-                param_data["lr"] = text_encoder_lr
-            all_params.append(param_data)
+            params, descriptions = assemble_params(
+                self.text_encoder_loras,
+                text_encoder_lr if text_encoder_lr is not None else default_lr,
+                self.loraplus_text_encoder_lr_ratio or self.loraplus_lr_ratio,
+            )
+            all_params.extend(params)
+            lr_descriptions.extend(["textencoder" + (" " + d if d else "") for d in descriptions])
 
         if self.unet_loras:
-            param_data = {"params": enumerate_params(self.unet_loras)}
-            if unet_lr is not None:
-                param_data["lr"] = unet_lr
-            all_params.append(param_data)
+            params, descriptions = assemble_params(
+                self.unet_loras,
+                unet_lr if unet_lr is not None else default_lr,
+                self.loraplus_unet_lr_ratio or self.loraplus_lr_ratio,
+            )
+            all_params.extend(params)
+            lr_descriptions.extend(["unet" + (" " + d if d else "") for d in descriptions])
 
-        return all_params
+        return all_params, lr_descriptions
 
     def save_weights(self, file, dtype, metadata):
         if metadata is not None and len(metadata) == 0: