import os import shutil import importlib.util from models.training import Training from models.TrainingProject import TrainingProject def load_base_config(selected_model): """Load base configuration for a specific YOLOX model""" model_name = selected_model.lower().replace('-', '_').replace('.pth', '') base_config_path = os.path.join(os.path.dirname(__file__), '..', 'data', f'{model_name}.py') if not os.path.exists(base_config_path): raise Exception(f'Base configuration not found for model: {model_name} at {base_config_path}') # Load the module dynamically spec = importlib.util.spec_from_file_location(f"base_config_{model_name}", base_config_path) module = importlib.util.module_from_spec(spec) spec.loader.exec_module(module) # Extract all attributes from BaseExp class base_exp = module.BaseExp() base_config = {} for attr in dir(base_exp): if not attr.startswith('_'): base_config[attr] = getattr(base_exp, attr) return base_config def generate_yolox_exp(training_id): """Generate YOLOX exp.py file""" # Fetch training row from DB training = Training.query.get(training_id) if not training: training = Training.query.filter_by(project_details_id=training_id).first() if not training: raise Exception(f'Training not found for trainingId or project_details_id: {training_id}') # If transfer_learning is 'coco', generate exp using base config + custom settings if training.transfer_learning == 'coco': exp_content = generate_yolox_inference_exp(training_id, use_base_config=True) return {'type': 'custom', 'expContent': exp_content} # If transfer_learning is 'sketch', generate custom exp.py if training.transfer_learning == 'sketch': exp_content = generate_yolox_inference_exp(training_id, use_base_config=False) return {'type': 'custom', 'expContent': exp_content} raise Exception(f'Unknown transfer_learning type: {training.transfer_learning}') def save_yolox_exp(training_id, out_path): """Save YOLOX exp.py to specified path""" exp_result = generate_yolox_exp(training_id) if exp_result['type'] == 'custom' and 'expContent' in exp_result: with open(out_path, 'w') as f: f.write(exp_result['expContent']) return out_path elif exp_result['type'] == 'default' and 'expPath' in exp_result: # Optionally copy the file if outPath is different if exp_result['expPath'] != out_path: shutil.copyfile(exp_result['expPath'], out_path) return out_path else: raise Exception('Unknown expResult type or missing content') def generate_yolox_inference_exp(training_id, options=None, use_base_config=False): """Generate inference exp.py using DB values Args: training_id: The training/project_details ID options: Optional overrides for data paths use_base_config: If True, load base config and only override with user-defined values """ if options is None: options = {} training = Training.query.get(training_id) if not training: training = Training.query.filter_by(project_details_id=training_id).first() if not training: raise Exception(f'Training not found for trainingId or project_details_id: {training_id}') # Always use the training_id (project_details_id) for annotation file names project_details_id = training.project_details_id data_dir = options.get('data_dir', '/home/kitraining/To_Annotate/') train_ann = options.get('train_ann', f'coco_project_{training_id}_train.json') val_ann = options.get('val_ann', f'coco_project_{training_id}_valid.json') test_ann = options.get('test_ann', f'coco_project_{training_id}_test.json') # Get num_classes from TrainingProject.classes JSON num_classes = 80 try: training_project = TrainingProject.query.get(project_details_id) if training_project and training_project.classes: classes_arr = training_project.classes if isinstance(classes_arr, str): import json classes_arr = json.loads(classes_arr) if isinstance(classes_arr, list): num_classes = len([c for c in classes_arr if c not in [None, '']]) elif isinstance(classes_arr, dict): num_classes = len([k for k, v in classes_arr.items() if v not in [None, '']]) except Exception as e: print(f'Could not determine num_classes from TrainingProject.classes: {e}') # Initialize config dictionary config = {} # If using base config (transfer learning from COCO), load protected parameters first if use_base_config and training.selected_model: try: base_config = load_base_config(training.selected_model) config.update(base_config) print(f'Loaded base config for {training.selected_model}: {list(base_config.keys())}') except Exception as e: print(f'Warning: Could not load base config for {training.selected_model}: {e}') print('Falling back to custom settings only') # Override with user-defined values from training table (only if they exist and are not None) user_overrides = { 'depth': training.depth, 'width': training.width, 'input_size': training.input_size, 'mosaic_scale': training.mosaic_scale, 'test_size': training.test_size, 'enable_mixup': training.enable_mixup, 'max_epoch': training.max_epoch, 'warmup_epochs': training.warmup_epochs, 'warmup_lr': training.warmup_lr, 'basic_lr_per_img': training.basic_lr_per_img, 'scheduler': training.scheduler, 'no_aug_epochs': training.no_aug_epochs, 'min_lr_ratio': training.min_lr_ratio, 'ema': training.ema, 'weight_decay': training.weight_decay, 'momentum': training.momentum, 'print_interval': training.print_interval, 'eval_interval': training.eval_interval, 'test_conf': training.test_conf, 'nms_thre': training.nms_thre, 'mosaic_prob': training.mosaic_prob, 'mixup_prob': training.mixup_prob, 'hsv_prob': training.hsv_prob, 'flip_prob': training.flip_prob, 'degrees': training.degrees, 'translate': training.translate, 'shear': training.shear, 'mixup_scale': training.mixup_scale, 'activation': training.activation, } # Only override if value is explicitly set (not None) for key, value in user_overrides.items(): if value is not None: config[key] = value # Apply any additional options overrides config.update(options) # Set defaults for any missing required parameters config.setdefault('depth', 1.00) config.setdefault('width', 1.00) config.setdefault('input_size', [640, 640]) config.setdefault('mosaic_scale', [0.1, 2]) config.setdefault('random_size', [10, 20]) config.setdefault('test_size', [640, 640]) config.setdefault('enable_mixup', False) config.setdefault('exp_name', 'inference_exp') # Build exp content exp_content = f'''#!/usr/bin/env python3 # -*- coding:utf-8 -*- # Copyright (c) Megvii, Inc. and its affiliates. import os from yolox.exp import Exp as MyExp class Exp(MyExp): def __init__(self): super(Exp, self).__init__() self.data_dir = "{data_dir}" self.train_ann = "{train_ann}" self.val_ann = "{val_ann}" self.test_ann = "{test_ann}" self.num_classes = {num_classes} ''' # Set pretrained_ckpt if transfer_learning is 'coco' if training.transfer_learning and isinstance(training.transfer_learning, str) and training.transfer_learning.lower() == 'coco': yolox_base_dir = '/home/kitraining/Yolox/YOLOX-main' selected_model = training.selected_model.replace('.pth', '') if training.selected_model else '' if selected_model: exp_content += f" self.pretrained_ckpt = r'{yolox_base_dir}/pretrained/{selected_model}.pth'\n" # Format arrays def format_value(val): if isinstance(val, (list, tuple)): return '(' + ', '.join(map(str, val)) + ')' elif isinstance(val, bool): return str(val) elif isinstance(val, str): return f'"{val}"' else: return str(val) # Add all config parameters to exp for key, value in config.items(): if key not in ['exp_name']: # exp_name is handled separately exp_content += f" self.{key} = {format_value(value)}\n" # Add exp_name at the end (uses dynamic path) exp_content += f''' self.exp_name = os.path.split(os.path.realpath(__file__))[1].split(".")[0] ''' return exp_content def save_yolox_inference_exp(training_id, out_path, options=None): """Save inference exp.py to custom path""" exp_content = generate_yolox_inference_exp(training_id, options, use_base_config=False) with open(out_path, 'w') as f: f.write(exp_content) return out_path