Abschluss-Projekt/backend/data

YOLOX Base Configuration System

Overview

This directory contains base experiment configurations for YOLOX models. These configurations define "protected" parameters that are preserved during transfer learning from COCO-pretrained models.

How It Works

Transfer Learning Flow

1. COCO Transfer Learning (transfer_learning = 'coco'):

   • Loads base configuration from data/yolox_*.py based on selected_model
   • Base parameters are protected and used as defaults
   • User settings from the form only override what's explicitly set
   • Result: Best of both worlds - proven COCO settings + your customizations

2. Sketch/Custom Training (transfer_learning = 'sketch'):

   • No base configuration loaded
   • Uses only user-defined parameters from the training form
   • Full control over all settings

Base Configuration Files

• yolox_s.py - YOLOX-Small (depth=0.33, width=0.50)
• yolox_m.py - YOLOX-Medium (depth=0.67, width=0.75)
• yolox_l.py - YOLOX-Large (depth=1.0, width=1.0)
• yolox_x.py - YOLOX-XLarge (depth=1.33, width=1.25)

Protected Parameters

These parameters are defined in base configs and preserved unless explicitly overridden:

Model Architecture:

• depth - Model depth multiplier
• width - Model width multiplier
• activation - Activation function (silu)

Training Hyperparameters:

• basic_lr_per_img - Learning rate per image
• scheduler - LR scheduler (yoloxwarmcos)
• warmup_epochs - Warmup epochs
• max_epoch - Maximum training epochs
• no_aug_epochs - No augmentation epochs
• min_lr_ratio - Minimum LR ratio

Optimizer:

• momentum - SGD momentum
• weight_decay - Weight decay

Augmentation:

• mosaic_prob - Mosaic probability
• mixup_prob - Mixup probability
• hsv_prob - HSV augmentation probability
• flip_prob - Flip probability
• degrees - Rotation degrees
• translate - Translation
• shear - Shear
• mosaic_scale - Mosaic scale range
• mixup_scale - Mixup scale range
• enable_mixup - Enable mixup

Input/Output:

• input_size - Training input size
• test_size - Testing size
• random_size - Random size range

Evaluation:

• eval_interval - Evaluation interval
• print_interval - Print interval

Customizing Base Configurations

Adding a New Model

Create a new file data/yolox_MODELNAME.py:

#!/usr/bin/env python3
# -*- coding:utf-8 -*-
# Base configuration for YOLOX-MODELNAME

class BaseExp:
    """Base experiment configuration for YOLOX-MODELNAME"""
    
    # Define protected parameters
    depth = 1.0
    width = 1.0
    # ... other parameters

Modifying Parameters

Edit the corresponding yolox_*.py file and update the BaseExp class attributes.

Example: To change YOLOX-S max epochs:

# In data/yolox_s.py
class BaseExp:
    max_epoch = 500  # Changed from 300
    # ... other parameters

Parameter Priority

The merge logic follows this priority (highest to lowest):

1. User form values (if explicitly set, not None)
2. Base config values (if transfer_learning='coco')
3. Default fallbacks (hardcoded minimums)

Example

COCO Transfer Learning

User sets in form: max_epoch=100, depth=0.5
Base config (yolox_s.py) has: depth=0.33, width=0.50, max_epoch=300

Result: depth=0.5 (user override), width=0.50 (base), max_epoch=100 (user override)

Sketch Training

User sets in form: max_epoch=100, depth=0.5
No base config loaded

Result: depth=0.5 (user), max_epoch=100 (user), width=1.0 (default fallback)

Debugging

To see which base config was loaded, check Flask logs:

Loaded base config for yolox-s: ['depth', 'width', 'activation', ...]

If base config fails to load:

Warning: Could not load base config for yolox-s: [error message]
Falling back to custom settings only