fix bugs

cog_test_train.sh

@@ -7,6 +7,6 @@ cog predict --gpus $GPU_ID \
     -i concept_mode="face" \
     -i sd_model_version="sdxl" \
     -i max_train_steps="360" \
-    -i caption_model="gpt4-v" \
+    -i caption_model="blip" \
     -i debug="False" \
     -i seed="0"

predict.py

@@ -86,10 +86,6 @@ class Predictor(BasePredictor):
             description="Number of steps between saving checkpoints. Set to very very high number to disable checkpointing, because you don't need intermediate checkpoints.",
             default=10000,
         ),
-        is_lora: bool = Input(
-            description="Whether to use LoRA training. If set to False, will use full fine tuning",
-            default=True,
-        ),
         unet_lr: float = Input(
             description="final learning rate of unet (after warmup)",
             default=0.001,
@@ -144,15 +140,13 @@ class Predictor(BasePredictor):
             train_batch_size=train_batch_size,
             max_train_steps=max_train_steps,
             checkpointing_steps=checkpointing_steps,
-            is_lora=is_lora,
-            prodigy_d_coef=prodigy_d_coef,
             ti_lr=ti_lr,
+            unet_lr=unet_lr,
             lora_rank=lora_rank,
             caption_model=caption_model,
             n_tokens=n_tokens,
             verbose=verbose,
             debug=debug,
-            text_encoder_lora_optimizer=text_encoder_lora_optimizer,
             freeze_ti_after_completion_f=freeze_ti_after_completion_f,
             token_warmup_steps=token_warmup_steps
         )

trainer/preprocess.py

@@ -521,7 +521,14 @@ def gpt4_v_caption_dataset(
         }
 
         response = requests.post("https://api.openai.com/v1/chat/completions", headers=headers, json=payload)
-        return index, response.json()["choices"][0]["message"]["content"]
+        
+        try:
+            result = response.json()["choices"][0]["message"]["content"]
+        except:
+            print(response.json())
+            result = ""
+
+        return index, result
 
     with concurrent.futures.ThreadPoolExecutor(max_workers=batch_size) as executor:
         future_to_index = {executor.submit(fetch_caption, i, img): i for i, img in enumerate(images) if captions[i] is None}

trainer/utils/utils.py

@@ -37,6 +37,8 @@ def replace_in_string(s, replacements):
     return s
 
 def fix_prompt(prompt: str):
+    if not prompt:
+        return prompt
     # Remove extra commas and spaces, and fix space before punctuation
     prompt = re.sub(r"\s+", " ", prompt)  # Replace multiple spaces with a single space
     prompt = re.sub(r",,", ",", prompt)  # Replace double commas with a single comma

training_args_face.json

@@ -1,7 +1,7 @@
 {
     "output_dir": "lora_models/xander_best",
-    "sd_model_version": "sd15",
-    "lora_training_urls": "/home/rednax/Documents/datasets/people/xander",
+    "sd_model_version": "sdxl",
+    "lora_training_urls": "https://storage.googleapis.com/public-assets-xander/A_workbox/lora_training_sets/xander_5.zip",
     "concept_mode": "face",
     "seed": 0,
     "resolution": 512,