bug fixing and improvements

- set onnxruntime execution_mode to sequential
- bugfix and improve ai model handling
- set used batch_size range to 1,2,4,8,16,32

graxpert/ai_model_handling.py

@@ -28,7 +28,7 @@
     try:
         os.rename(ai_models_dir, bge_ai_models_dir)
     except Exception as e:
-        logging.error(f"Renaming {ai_models_dir} to {bge_ai_models_dir} failed. {bge_ai_models_dir} will be newly created. Consider deleting obsolete {ai_models_dir}.")
+        logging.error(f"Renaming {ai_models_dir} to {bge_ai_models_dir} failed. {bge_ai_models_dir} will be newly created. Consider deleting obsolete {ai_models_dir} manually.")
 
 os.makedirs(bge_ai_models_dir, exist_ok=True)
 
@@ -64,7 +64,11 @@ def list_remote_versions(bucket_name):
 
 def list_local_versions(ai_models_dir):
     try:
-        model_dirs = [{"path": os.path.join(ai_models_dir, f), "version": f} for f in os.listdir(ai_models_dir) if re.search(r"\d\.\d\.\d", f)]  # match semantic version
+        model_dirs = [
+            {"path": os.path.join(ai_models_dir, f), "version": f}
+            for f in os.listdir(ai_models_dir)
+            if re.search(r"\d\.\d\.\d", f) and len(os.listdir(os.path.join(ai_models_dir, f))) > 0  # match semantic version
+        ]
         return model_dirs
     except Exception as e:
         logging.exception(e)
@@ -122,11 +126,11 @@ def cleanup_orphaned_local_versions(orphaned_local_versions):
             logging.exception(e)
 
 
-def download_version(ai_models_dir, bucket_name, remote_version, progress=None):
+def download_version(ai_models_dir, bucket_name, target_version, progress=None):
     try:
         remote_versions = list_remote_versions(bucket_name)
         for r in remote_versions:
-            if remote_version == r["version"]:
+            if target_version == r["version"]:
                 remote_version = r
                 break
 
@@ -144,11 +148,11 @@ def download_version(ai_models_dir, bucket_name, remote_version, progress=None):
 
         with zipfile.ZipFile(ai_model_zip, "r") as zip_ref:
             zip_ref.extractall(ai_model_dir)
-        
+
         if not os.path.isfile(ai_model_file):
             raise ValueError(f"Could not find ai 'model.onnx' file after extracting {ai_model_zip}")
         os.remove(ai_model_zip)
-        
+
     except Exception as e:
         # try to delete (rollback) ai_model_dir in case of errors
         logging.exception(e)
@@ -163,7 +167,6 @@ def validate_local_version(ai_models_dir, local_version):
 
 
 def get_execution_providers_ordered():
-    supported_providers = ["DmlExecutionProvider", "CoreMLExecutionProvider", "CUDAExecutionProvider",
-                           "CPUExecutionProvider"]
+    supported_providers = ["DmlExecutionProvider", "CoreMLExecutionProvider", "CUDAExecutionProvider", "CPUExecutionProvider"]
 
     return [provider for provider in supported_providers if provider in ort.get_available_providers()]

graxpert/denoising.py

@@ -6,20 +6,31 @@
 import onnxruntime as ort
 
 from graxpert.ai_model_handling import get_execution_providers_ordered
-from graxpert.application.eventbus import eventbus
 from graxpert.application.app_events import AppEvents
+from graxpert.application.eventbus import eventbus
 from graxpert.ui.ui_events import UiEvents
 
-def denoise(image, ai_path, strength, batch_size=3, window_size=256, stride=128, progress=None):
+
+def denoise(image, ai_path, strength, batch_size=4, window_size=256, stride=128, progress=None):
 
     logging.info("Starting denoising")
 
+    if batch_size < 1:
+        logging.info(f"mapping batch_size of {batch_size} to 1")
+        batch_size = 1
+    elif batch_size > 32:
+        logging.info(f"mapping batch_size of {batch_size} to 32")
+        batch_size = 32
+    elif not (batch_size & (batch_size - 1) == 0):  # check if batch_size is power of two
+        logging.info(f"mapping batch_size of {batch_size} to {2 ** (batch_size).bit_length() // 2}")
+        batch_size = 2 ** (batch_size).bit_length() // 2  # map batch_size to power of two
+
     input = copy.deepcopy(image)
 
     global cached_denoised_image
     if cached_denoised_image is not None:
         return blend_images(input, cached_denoised_image, strength)
-    
+
     num_colors = image.shape[-1]
     if num_colors == 1:
         image = np.array([image[:, :, 0], image[:, :, 0], image[:, :, 0]])
@@ -51,9 +62,7 @@ def denoise(image, ai_path, strength, batch_size=3, window_size=256, stride=128,
     output = copy.deepcopy(image)
 
     providers = get_execution_providers_ordered()
-    ort_options = ort.SessionOptions()
-    ort_options.execution_mode = ort.ExecutionMode.ORT_PARALLEL
-    session = ort.InferenceSession(ai_path, providers=providers, sess_options=ort_options)
+    session = ort.InferenceSession(ai_path, providers=providers)
 
     logging.info(f"Available inference providers : {providers}")
     logging.info(f"Used inference providers : {session.get_providers()}")
@@ -84,7 +93,7 @@ def denoise(image, ai_path, strength, batch_size=3, window_size=256, stride=128,
 
         if not input_tiles:
             continue
-        
+
         input_tiles = np.array(input_tiles)
 
         output_tiles = []
@@ -131,15 +140,18 @@ def denoise(image, ai_path, strength, batch_size=3, window_size=256, stride=128,
 
     return output
 
+
 def blend_images(original_image, denoised_image, strength):
-    blend = denoised_image * strength + original_image * (1-strength)
+    blend = denoised_image * strength + original_image * (1 - strength)
     return np.clip(blend, 0, 1)
 
+
 def reset_cached_denoised_image(event):
     global cached_denoised_image
     cached_denoised_image = None
 
+
 cached_denoised_image = None
 eventbus.add_listener(AppEvents.LOAD_IMAGE_REQUEST, reset_cached_denoised_image)
 eventbus.add_listener(AppEvents.CALCULATE_REQUEST, reset_cached_denoised_image)
-eventbus.add_listener(UiEvents.APPLY_CROP_REQUEST, reset_cached_denoised_image)
+eventbus.add_listener(UiEvents.APPLY_CROP_REQUEST, reset_cached_denoised_image)

graxpert/main.py

@@ -49,6 +49,9 @@ def collect_available_versions(ai_models_dir, bucket_name):
 def bge_version_type(arg_value, pat=re.compile(r"^\d+\.\d+\.\d+$")):
     return version_type(bge_ai_models_dir, bge_bucket_name, arg_value, pat=re.compile(r"^\d+\.\d+\.\d+$"))
 
+def denoise_version_type(arg_value, pat=re.compile(r"^\d+\.\d+\.\d+$")):
+    return version_type(denoise_ai_models_dir, denoise_bucket_name, arg_value, pat=re.compile(r"^\d+\.\d+\.\d+$"))
+
 
 def version_type(ai_models_dir, bucket_name, arg_value, pat=re.compile(r"^\d+\.\d+\.\d+$")):
 
@@ -227,7 +230,7 @@ def main():
             nargs="?",
             required=False,
             default=None,
-            type=bge_version_type,
+            type=denoise_version_type,
             help='Version of the Denoising AI model, default: "latest"; available locally: [{}], available remotely: [{}]'.format(
                 ", ".join(available_denoise_versions[0]), ", ".join(available_denoise_versions[1])
             ),
@@ -239,7 +242,7 @@ def main():
             required=False,
             default=None,
             type=float,
-            help='Strength of the desired denoising effect, default: "1.0"',
+            help='Strength of the desired denoising effect, default: "0.5"',
         )
         denoise_parser.add_argument(
             "-batch_size",
@@ -248,7 +251,7 @@ def main():
             required=False,
             default=None,
             type=int,
-            help='Number of image tiles which Graxpert will denoise in parallel. Be careful: increasing this value might result in out-of-memory errors. Valid Range: 1..50, default: "3"',
+            help='Number of image tiles which Graxpert will denoise in parallel. Be careful: increasing this value might result in out-of-memory errors. Valid Range: 1..32, default: "4"',
         )
 
         if "-h" in sys.argv or "--help" in sys.argv:

graxpert/preferences.py

@@ -40,7 +40,7 @@ class Prefs:
     denoise_ai_version: AnyStr = None
     graxpert_version: AnyStr = graxpert_version
     denoise_strength: float = 0.5
-    ai_batch_size: int = 3
+    ai_batch_size: int = 4
 
 
 def app_state_2_prefs(prefs: Prefs, app_state: AppState) -> Prefs:

graxpert/ui/right_menu.py

@@ -187,6 +187,7 @@ def __init__(self, master, **kwargs):
         self.denoise_ai_version.trace_add("write", lambda a, b, c: eventbus.emit(AppEvents.DENOISE_AI_VERSION_CHANGED, {"denoise_ai_version": self.denoise_ai_version.get()}))
 
         # ai settings
+        self.ai_batch_size_options = ["1","2","4","8","16","32"]
         self.ai_batch_size = tk.IntVar()
         self.ai_batch_size.set(graxpert.prefs.ai_batch_size)
         self.ai_batch_size.trace_add("write", lambda a, b, c: eventbus.emit(AppEvents.AI_BATCH_SIZE_CHANGED, {"ai_batch_size": self.ai_batch_size.get()}))
@@ -239,7 +240,8 @@ def lang_change(lang):
         GraXpertOptionMenu(self, variable=self.denoise_ai_version, values=self.denoise_ai_options).grid(**self.default_grid())
 
         # ai settings
-        ValueSlider(self, variable=self.ai_batch_size, variable_name=_("AI Batch Size"), min_value=1, max_value=50, precision=0).grid(**self.default_grid())
+        CTkLabel(self, text=_("AI inference batch size"), font=self.heading_font2).grid(column=0, row=self.nrow(), pady=pady, sticky=tk.N)
+        GraXpertOptionMenu(self, variable=self.ai_batch_size, values=self.ai_batch_size_options).grid(**self.default_grid())
 
     def setup_layout(self):
         self.columnconfigure(0, weight=1)