update args

.gitignore

@@ -12,3 +12,5 @@ tests/
 train.py
 debug/*
 !debug/*.py
+
+training_args_x_*.json

trainer/config.py

@@ -23,6 +23,7 @@ class TrainingConfig(BaseModel):
     ti_weight_decay: float = 3e-4
     lora_weight_decay: float = 0.002
     l1_penalty: float = 0.1
+    noise_offset: float = 0.05
     snr_gamma: float = 5.0
     lora_rank: int = 12
     use_dora: bool = False
@@ -55,6 +56,7 @@ class TrainingConfig(BaseModel):
     lr_num_cycles: int = 1
     lr_power: float = 1.0
     dataloader_num_workers: int = 0
+    training_attributes: dict = {}
 
     def save_as_json(self, file_path: str) -> None:
         with open(file_path, 'w') as f:

trainer/dataset_and_utils.py

@@ -66,6 +66,8 @@ def plot_loss(losses, save_path='losses.png', window_length=31, polyorder=3):
     # plt.yscale('log')  # Uncomment if log scale is desired
     plt.xlabel('Step')
     plt.ylabel('Training Loss')
+    ymin, ymax = plt.ylim()
+    plt.ylim(min(ymin, 0), ymax)
     plt.legend()
     plt.savefig(save_path)
     plt.close()

trainer/utils/inference.py

@@ -80,6 +80,6 @@ def render_images(training_pipeline, render_size, lora_path, train_step, seed, i
     img_grid_path = make_validation_img_grid(lora_path)
 
     if not reload_entire_pipeline: # restore the training scheduler
-        pipeline.scheduler = training_scheduler
+        training_pipeline.scheduler = training_scheduler
 
     return validation_prompts_raw

trainer_pti.py

@@ -52,7 +52,11 @@ def main(
         seed = config.seed,
     )
 
-    instance_data_dir=os.path.join(input_dir, "captions.csv")
+    # Update the training attributes with some info from the pre-processing:
+    config.training_attributes["n_training_imgs"] = n_imgs
+    config.training_attributes["trigger_text"] = trigger_text
+    config.training_attributes["segmentation_prompt"] = segmentation_prompt
+    config.training_attributes["captions"] = captions
 
     if config.allow_tf32:
         torch.backends.cuda.matmul.allow_tf32 = True
@@ -161,8 +165,9 @@ def main(
         #unet.add_adapter(unet_lora_config)
         
         unet = get_peft_model(unet, unet_lora_config)
+        pipe.unet = unet
         print_trainable_parameters(unet, name = 'unet')
-
+        
         unet_lora_parameters = list(filter(lambda p: p.requires_grad, unet.parameters()))
 
         params_to_optimize = [
@@ -213,7 +218,7 @@ def main(
         )
         
     train_dataset = PreprocessedDataset(
-        instance_data_dir,
+        os.path.join(input_dir, "captions.csv"),
         tokenizer_one,
         tokenizer_two,
         vae,
@@ -349,10 +354,9 @@ def main(
             # Sample noise that we'll add to the latents:
             noise = torch.randn_like(vae_latent)
 
-            noise_offset = 0.05 # TODO, turn this into an input arg and do a grid search
-            if noise_offset > 0.0:
+            if config.noise_offset > 0.0:
                 # https://www.crosslabs.org//blog/diffusion-with-offset-noise
-                noise += noise_offset * torch.randn(
+                noise += config.noise_offset * torch.randn(
                     (noise.shape[0], noise.shape[1], 1, 1), device=noise.device)
 
             bsz = vae_latent.shape[0]

training_args.json

@@ -3,7 +3,7 @@
     "name": "unnamed",
     "lora_training_urls": "https://storage.googleapis.com/public-assets-xander/A_workbox/lora_training_sets/clipx_tiny.zip",
     "concept_mode": "style",
-    "sd_model_version": "sd15",
+    "sd_model_version": "sdxl",
     "seed": 0,
     "resolution": 960,
     "train_batch_size": 4,
@@ -17,7 +17,7 @@
     "l1_penalty": 0.05,
     "snr_gamma": 5.0,
     "lora_rank": 12,
-    "use_dora": false,
+    "use_dora": true,
     "caption_prefix": "in the style of TOK, ",
     "caption_model": "blip",
     "left_right_flip_augmentation": true,