fix(ui): Set Zero values to Null instead of 0.0 (#161)

Author: hipsterusername

pyproject.toml

@@ -69,7 +69,7 @@ dependencies = [
 
 [tool.ruff]
 src = ["src"]
-select = ["E", "F", "W", "C9", "N8", "I"]
+lint.select = ["E", "F", "W", "C9", "N8", "I"]
 target-version = "py39"
 line-length = 120
 

src/invoke_training/pipelines/_experimental/sd_dpo_lora/config.py

@@ -95,11 +95,12 @@ class SdDirectPreferenceOptimizationLoraConfig(BasePipelineConfig):
 
     text_encoder_learning_rate: float | None = None
     """The learning rate to use for the text encoder model. If set, this overrides the optimizer's default learning
-    rate.
+    rate. Set to null or 0 to use the optimizer's default learning rate.
     """
 
     unet_learning_rate: float | None = None
     """The learning rate to use for the UNet model. If set, this overrides the optimizer's default learning rate.
+    Set to null or 0 to use the optimizer's default learning rate.
     """
 
     lr_scheduler: Literal[
@@ -194,7 +195,7 @@ class SdDirectPreferenceOptimizationLoraConfig(BasePipelineConfig):
     """
 
     max_grad_norm: float | None = None
-    """Max gradient norm for clipping. Set to None for no clipping.
+    """Max gradient norm for clipping. Set to null or 0 for no clipping.
     """
 
     validation_prompts: list[str] = []

src/invoke_training/pipelines/flux/lora/config.py

@@ -52,12 +52,12 @@ class FluxLoraConfig(BasePipelineConfig):
 
     text_encoder_learning_rate: float | None = 1e-4
     """The learning rate to use for the text encoder model. If set, this overrides the optimizer's default learning
-    rate.
+    rate. Set to null or 0 to use the optimizer's default learning rate.
     """
 
     transformer_learning_rate: float | None = 4e-4
     """The learning rate to use for the transformer model. If set, this overrides the optimizer's default learning
-    rate.
+    rate. Set to null or 0 to use the optimizer's default learning rate.
     """
 
     lr_scheduler: Literal[
@@ -176,7 +176,7 @@ class FluxLoraConfig(BasePipelineConfig):
     """
 
     max_grad_norm: float | None = None
-    """Max gradient norm for clipping. Set to None for no clipping.
+    """Max gradient norm for clipping. Set to null or 0 for no clipping.
     """
 
     validation_prompts: list[str] = []

src/invoke_training/pipelines/stable_diffusion/lora/config.py

@@ -59,11 +59,12 @@ class SdLoraConfig(BasePipelineConfig):
 
     text_encoder_learning_rate: float | None = None
     """The learning rate to use for the text encoder model. If set, this overrides the optimizer's default learning
-    rate.
+    rate. Set to null or 0 to use the optimizer's default learning rate.
     """
 
     unet_learning_rate: float | None = None
     """The learning rate to use for the UNet model. If set, this overrides the optimizer's default learning rate.
+    Set to null or 0 to use the optimizer's default learning rate.
     """
 
     lr_scheduler: Literal[
@@ -188,7 +189,7 @@ class SdLoraConfig(BasePipelineConfig):
     """
 
     max_grad_norm: float | None = None
-    """Max gradient norm for clipping. Set to None for no clipping.
+    """Max gradient norm for clipping. Set to null or 0 for no clipping.
     """
 
     validation_prompts: list[str] = []

src/invoke_training/pipelines/stable_diffusion/textual_inversion/config.py

@@ -160,7 +160,7 @@ class SdTextualInversionConfig(BasePipelineConfig):
     """
 
     max_grad_norm: float | None = None
-    """Maximum gradient norm for gradient clipping. Set to `None` for no clipping.
+    """Maximum gradient norm for gradient clipping. Set to `null` or 0 for no clipping.
     """
 
     validation_prompts: list[str] = []

src/invoke_training/pipelines/stable_diffusion_xl/finetune/config.py

@@ -125,7 +125,7 @@ class SdxlFinetuneConfig(BasePipelineConfig):
     """
 
     max_grad_norm: float | None = None
-    """Max gradient norm for clipping. Set to None for no clipping.
+    """Max gradient norm for clipping. Set to null or 0 for no clipping.
     """
 
     validation_prompts: list[str] = []

src/invoke_training/pipelines/stable_diffusion_xl/lora/config.py

@@ -59,11 +59,12 @@ class SdxlLoraConfig(BasePipelineConfig):
 
     text_encoder_learning_rate: float | None = None
     """The learning rate to use for the text encoder model. If set, this overrides the optimizer's default learning
-    rate.
+    rate. Set to null or 0 to use the optimizer's default learning rate.
     """
 
     unet_learning_rate: float | None = None
     """The learning rate to use for the UNet model. If set, this overrides the optimizer's default learning rate.
+    Set to null or 0 to use the optimizer's default learning rate.
     """
 
     lr_scheduler: Literal[
@@ -188,7 +189,7 @@ class SdxlLoraConfig(BasePipelineConfig):
     """
 
     max_grad_norm: float | None = None
-    """Max gradient norm for clipping. Set to None for no clipping.
+    """Max gradient norm for clipping. Set to null or 0 for no clipping.
     """
 
     validation_prompts: list[str] = []

src/invoke_training/pipelines/stable_diffusion_xl/lora_and_textual_inversion/config.py

@@ -84,16 +84,18 @@ class SdxlLoraAndTextualInversionConfig(BasePipelineConfig):
 
     optimizer: AdamOptimizerConfig | ProdigyOptimizerConfig = AdamOptimizerConfig()
 
-    text_encoder_learning_rate: float = 1e-5
-    """The learning rate to use for the text encoder model.
+    text_encoder_learning_rate: float | None = 1e-5
+    """The learning rate to use for the text encoder model. Set to null or 0 to use the optimizer's default learning
+    rate.
     """
 
-    unet_learning_rate: float = 1e-4
-    """The learning rate to use for the UNet model.
+    unet_learning_rate: float | None = 1e-4
+    """The learning rate to use for the UNet model. Set to null or 0 to use the optimizer's default learning rate.
     """
 
-    textual_inversion_learning_rate: float = 1e-3
-    """The learning rate to use for textual inversion training of the embeddings.
+    textual_inversion_learning_rate: float | None = 1e-3
+    """The learning rate to use for textual inversion training of the embeddings. Set to null or 0 to use the
+    optimizer's default learning rate.
     """
 
     lr_scheduler: Literal[
@@ -187,7 +189,7 @@ class SdxlLoraAndTextualInversionConfig(BasePipelineConfig):
     """
 
     max_grad_norm: float | None = None
-    """Max gradient norm for clipping. Set to None for no clipping.
+    """Max gradient norm for clipping. Set to null or 0 for no clipping.
     """
 
     validation_prompts: list[str] = []

src/invoke_training/ui/config_groups/flux_lora_config_group.py

@@ -67,7 +67,7 @@ def __init__(self):
             with gr.Row():
                 self.transformer_learning_rate = gr.Number(
                     label="Transformer Learning Rate",
-                    info="The transformer learning rate. If None, then it is inherited from the base optimizer "
+                    info="The transformer learning rate. Set to 0 or leave empty to inherit from the base optimizer "
                     "learning rate.",
                     interactive=True,
                 )
@@ -103,7 +103,7 @@ def __init__(self):
                 )
                 self.max_grad_norm = gr.Number(
                     label="Max Gradient Norm",
-                    info="Max gradient norm for clipping. Set to None for no clipping.",
+                    info="Max gradient norm for clipping. Set to 0 or leave empty for no clipping (null).",
                     interactive=True,
                 )
                 self.train_batch_size = gr.Number(
@@ -313,13 +313,9 @@ def safe_pop(component, default=None):
             # Set basic properties
             new_config.model = safe_pop(self.model, new_config.model)
             new_config.train_transformer = safe_pop(self.train_transformer, new_config.train_transformer)
-            new_config.train_text_encoder = safe_pop(self.train_text_encoder, new_config.train_text_encoder)
-            new_config.transformer_learning_rate = safe_pop(
-                self.transformer_learning_rate, new_config.transformer_learning_rate
-            )
-            new_config.text_encoder_learning_rate = safe_pop(
-                self.text_encoder_learning_rate, new_config.text_encoder_learning_rate
-            )
+            # Note: train_text_encoder and text_encoder_learning_rate are not supported for Flux LoRA
+            transformer_lr_value = safe_pop(self.transformer_learning_rate, new_config.transformer_learning_rate)
+            new_config.transformer_learning_rate = None if transformer_lr_value == 0 else transformer_lr_value
             new_config.gradient_accumulation_steps = safe_pop(
                 self.gradient_accumulation_steps, new_config.gradient_accumulation_steps
             )
@@ -330,7 +326,8 @@ def safe_pop(component, default=None):
 
             new_config.lora_rank_dim = safe_pop(self.lora_rank_dim, new_config.lora_rank_dim)
             new_config.min_snr_gamma = safe_pop(self.min_snr_gamma, new_config.min_snr_gamma)
-            new_config.max_grad_norm = safe_pop(self.max_grad_norm, new_config.max_grad_norm)
+            max_grad_norm_value = safe_pop(self.max_grad_norm, new_config.max_grad_norm)
+            new_config.max_grad_norm = None if max_grad_norm_value == 0 else max_grad_norm_value
             new_config.train_batch_size = safe_pop(self.train_batch_size, new_config.train_batch_size)
             new_config.weight_dtype = safe_pop(self.weight_dtype, new_config.weight_dtype)
             new_config.mixed_precision = safe_pop(self.mixed_precision, new_config.mixed_precision)

src/invoke_training/ui/config_groups/sd_lora_config_group.py

@@ -109,12 +109,13 @@ def __init__(self):
             with gr.Row():
                 self.unet_learning_rate = gr.Number(
                     label="UNet Learning Rate",
-                    info="The UNet learning rate. If None, then it is inherited from the base optimizer learning rate.",
+                    info="The UNet learning rate. Set to 0 or leave empty to inherit from the base optimizer "
+                    "learning rate.",
                     interactive=True,
                 )
                 self.text_encoder_learning_rate = gr.Number(
                     label="Text Encoder Learning Rate",
-                    info="The text encoder learning rate. If None, then it is inherited from the base optimizer "
+                    info="The text encoder learning rate. Set to 0 or leave empty to inherit from the base optimizer "
                     "learning rate.",
                     interactive=True,
                 )
@@ -153,7 +154,7 @@ def __init__(self):
                 )
                 self.max_grad_norm = gr.Number(
                     label="Max Gradient Norm",
-                    info="Max gradient norm for clipping. Set to None for no clipping.",
+                    info="Max gradient norm for clipping. Set to 0 or leave empty for no clipping (null).",
                     interactive=True,
                 )
                 self.train_batch_size = gr.Number(
@@ -223,13 +224,16 @@ def update_config_with_ui_component_data(
         new_config.hf_variant = ui_data.pop(self.hf_variant) or None
         new_config.max_checkpoints = ui_data.pop(self.max_checkpoints)
         new_config.train_unet = ui_data.pop(self.train_unet)
-        new_config.unet_learning_rate = ui_data.pop(self.unet_learning_rate)
+        unet_lr_value = ui_data.pop(self.unet_learning_rate)
+        new_config.unet_learning_rate = None if unet_lr_value == 0 else unet_lr_value
         new_config.train_text_encoder = ui_data.pop(self.train_text_encoder)
-        new_config.text_encoder_learning_rate = ui_data.pop(self.text_encoder_learning_rate)
+        text_encoder_lr_value = ui_data.pop(self.text_encoder_learning_rate)
+        new_config.text_encoder_learning_rate = None if text_encoder_lr_value == 0 else text_encoder_lr_value
         new_config.lr_scheduler = ui_data.pop(self.lr_scheduler)
         new_config.lr_warmup_steps = ui_data.pop(self.lr_warmup_steps)
         new_config.use_masks = ui_data.pop(self.use_masks)
-        new_config.max_grad_norm = ui_data.pop(self.max_grad_norm)
+        max_grad_norm_value = ui_data.pop(self.max_grad_norm)
+        new_config.max_grad_norm = None if max_grad_norm_value == 0 else max_grad_norm_value
         new_config.train_batch_size = ui_data.pop(self.train_batch_size)
         new_config.cache_text_encoder_outputs = ui_data.pop(self.cache_text_encoder_outputs)
         new_config.cache_vae_outputs = ui_data.pop(self.cache_vae_outputs)