prepare train.py to automated training

set cryptography lib to the certain version
add manual_run for manual operations, right now it is onnx conversion and upload

annotation-queue/offset.yaml

@@ -1 +1 @@
-offset_queue: 0
+offset_queue: 108

manual_run.py

@@ -0,0 +1,31 @@
+import shutil
+from datetime import datetime
+from os import path
+
+from constants import models_dir, prefix, date_format, MODELS_FOLDER
+from api_client import ApiClient
+from augmentation import Augmentator
+from exports import export_onnx
+from security import Security
+
+# Augmentator().augment_annotations()
+#train_dataset(from_scratch=True)
+
+# resume_training('/azaion/dev/ai-training/runs/detect/train12/weights/last.pt')
+
+result_dir = '/azaion/dev/ai-training/runs/detect/train12'
+model_dir = path.join(models_dir, f'{prefix}2025-05-18')
+shutil.copytree(result_dir, model_dir)
+
+model_path = path.join(models_dir, f'{prefix[:-1]}.pt')
+shutil.copy(path.join(model_dir, 'weights', 'best.pt'), model_path)
+
+api_client = ApiClient()
+onnx_path = export_onnx(model_path)
+print(f'Conversion done: onnx path: {onnx_path}')
+
+with open(onnx_path, 'rb') as binary_file:
+    onnx_bytes = binary_file.read()
+
+key = Security.get_model_encryption_key()
+api_client.upload_big_small_resource(onnx_bytes, onnx_path, MODELS_FOLDER, key)

preprocess-train.py

@@ -1,8 +0,0 @@
-from augmentation import Augmentator
-from train import train_dataset, convert2rknn, resume_training
-
-# Augmentator().augment_annotations()
-#train_dataset(from_scratch=True)
-
-
-resume_training('/azaion/dev/ai-training/runs/detect/train12/weights/last.pt')

requirements.txt

@@ -8,7 +8,7 @@ albumentations
 opencv-python
 matplotlib
 PyYAML
-cryptography
+cryptography==44.0.2
 numpy==1.26.4
 requests
 pyyaml

train.py

@@ -37,33 +37,24 @@ old_images_percentage = 75
 DEFAULT_CLASS_NUM = 80
 total_files_copied = 0
 
-def form_dataset(from_date: datetime):
-    makedirs(today_dataset, exist_ok=True)
+
+def form_dataset():
+    shutil.rmtree(today_dataset, ignore_errors=True)
+    makedirs(today_dataset)
     images = []
-    old_images = []
     with scandir(processed_images_dir) as imd:
         for image_file in imd:
             if not image_file.is_file():
                 continue
-            mod_time = datetime.fromtimestamp(image_file.stat().st_mtime).replace(hour=0, minute=0, second=0, microsecond=0)
-            if from_date is None:
             images.append(image_file)
-            elif mod_time > from_date:
-                images.append(image_file)
-            else: # gather old images as well in order to avoid overfitting on the only new data.
-                old_images.append(image_file)
-
-    random.shuffle(old_images)
-    old_images_size = int(len(old_images) * old_images_percentage / 100.0)
-
-    print(f'Got {len(images)} new images and {old_images_size} of old images (to prevent overfitting). Shuffling them...')
-    images.extend(old_images[:old_images_size])
 
+    print(f'Got {len(images)} images. Start shuffling...')
     random.shuffle(images)
 
     train_size = int(len(images) * train_set / 100.0)
     valid_size = int(len(images) * valid_set / 100.0)
 
+    print(f'Start copying...')
     copy_annotations(images[:train_size], 'train')
     copy_annotations(images[train_size:train_size + valid_size], 'valid')
     copy_annotations(images[train_size + valid_size:], 'test')
@@ -140,37 +131,6 @@ def create_yaml():
         f.writelines([f'{line}\n' for line in lines])
 
 
-def revert_to_processed_data(date):
-    def revert_dir(src_dir, dest_dir):
-        for file in listdir(src_dir):
-            s = path.join(src_dir, file)
-            d = path.join(dest_dir, file)
-            replace(s, d)
-    date_dataset = path.join(datasets_dir, f'{prefix}{date}')
-    makedirs(processed_images_dir, exist_ok=True)
-    makedirs(processed_labels_dir, exist_ok=True)
-    for subset in ['test', 'train', 'valid']:
-        revert_dir(path.join(date_dataset, subset, 'images'), processed_images_dir)
-        revert_dir(path.join(date_dataset, subset, 'labels'), processed_labels_dir)
-    shutil.rmtree(date_dataset)
-
-
-def get_latest_model():
-    def convert(d: str):
-        if not d.startswith(prefix):
-            return None
-        dir_date = datetime.strptime(d.replace(prefix, ''), '%Y-%m-%d')
-        dir_model_path = path.join(models_dir, d, 'weights', 'best.pt')
-        return {'date': dir_date, 'path': dir_model_path}
-
-    dates = [convert(d) for d in next(os.walk(models_dir))[1]]
-    dates = list(filter(lambda x : x is not None, dates))
-    sorted_dates = list(sorted(dates, key=lambda x: x['date'] ))
-    if len(sorted_dates) == 0:
-        return None, None
-    last_model = sorted_dates[-1]
-    return last_model['date'], last_model['path']
-
 
 def resume_training(last_pt_path):
     model = YOLO(last_pt_path)
@@ -183,63 +143,38 @@ def resume_training(last_pt_path):
                           workers=24)
 
 
-def train_dataset(existing_date=None, from_scratch=False):
-    latest_date, latest_model = get_latest_model() if not from_scratch else None, None
-
-    if existing_date is not None:
-        cur_folder = f'{prefix}{existing_date}'
-        cur_dataset = path.join(datasets_dir, f'{prefix}{existing_date}')
-    else:
-        # if from_scratch and Path(today_dataset).exists():
-        #     shutil.rmtree(today_dataset)
-        # form_dataset(latest_date)
-        # create_yaml()
-        cur_folder = today_folder
-        cur_dataset = today_dataset
-
-    model_name = latest_model if latest_model is not None and path.isfile(latest_model) and not from_scratch else 'yolo11m.yaml'
-    print(f'Initial model: {model_name}')
+def train_dataset():
+    form_dataset()
+    create_yaml()
+    model_name = 'yolo11m.yaml'
     model = YOLO(model_name)
 
-    yaml = abspath(path.join(cur_dataset, 'data.yaml'))
-    results = model.train(data=yaml,
+    results = model.train(data=abspath(path.join(today_dataset, 'data.yaml')),
                            epochs=120,
                            batch=11,
                            imgsz=1280,
                            save_period=1,
                            workers=24)
 
-    model_dir = path.join(models_dir, cur_folder)
+    model_dir = path.join(models_dir, today_folder)
     shutil.copytree(results.save_dir, model_dir)
 
     model_path = path.join(models_dir, f'{prefix[:-1]}.pt')
     shutil.copy(path.join(model_dir, 'weights', 'best.pt'), model_path)
-    shutil.rmtree('runs')
     return model_path
 
 
-def convert2rknn():
-    subprocess.call(['bash', 'convert.sh'], cwd="./orangepi5")
-    latest_date, latest_model = get_latest_model()
-    model = YOLO(latest_model)
-    model.export(format="onnx")
-    pass
-
-
 def validate(model_path):
     model = YOLO(model_path)
-    metrics = model.val()
-    pass
+    print(model.val())
 
 
 if __name__ == '__main__':
-    model_path = train_dataset(from_scratch=True)
-    # validate(path.join('runs', 'detect', 'train7', 'weights', 'best.pt'))
-    # form_data_sample(500)
-    # convert2rknn()
-    api_client = ApiClient()
-    onnx_path = export_onnx('azaion.pt')
+    model_path = train_dataset()
+    validate(path.join('runs', 'detect', 'train7', 'weights', 'best.pt'))
+    onnx_path = export_onnx(model_path)
 
+    api_client = ApiClient()
     with open(onnx_path, 'rb') as binary_file:
         onnx_bytes = binary_file.read()