CNN Comparator: An Interactive Web Platform for Convolutional Neural Network Architecture Comparison

CNN Comparator is an open-source web application focused on comparing the performance of different convolutional neural network (CNN) architectures in image classification tasks. Using a front-end and back-end separation architecture, it supports performance evaluation and visual comparison of multiple classic CNN models. It aims to address the pain points of low efficiency in model selection and experience-dependent decision-making in deep learning practice, helping developers and researchers quickly make model choices based on objective data.

In the field of deep learning, choosing the right CNN architecture is crucial for image classification tasks. However, faced with numerous options like ResNet, VGG, and Inception, developers often need to write a lot of repetitive code to train and test models, which is inefficient and error-prone. CNN Comparator was designed to solve this pain point: through an intuitive interface and automated processes, users can easily load models, train and evaluate them on the same dataset, and visualize the comparison results.

The project uses a three-tier architecture:

1. AI Backend: Responsible for model loading, training process management, inference prediction, and performance metric calculation, supporting multiple deep learning frameworks;
2. Logic Backend: Handles business logic, including request routing, dataset management, result aggregation and caching, and user authentication;
3. Frontend: Provides an intuitive model selection interface, real-time training monitoring, and interactive visual report generation to ensure a good user experience.

The core features of CNN Comparator include:

• Multi-model Parallel Comparison: Supports model series like ResNet, VGG, and MobileNet, enabling fair comparison under the same dataset and configuration;
• Standardized Evaluation Metrics: Provides metrics such as accuracy, precision, recall, F1 score, confusion matrix, training time, and parameter count;
• Visual Reports: Displays comparison results via bar charts, line charts, heatmaps, and ROC curves;
• Dataset Management: Built-in common datasets like CIFAR-10, supporting custom uploads, data augmentation, and preprocessing.

The project provides a complete Docker Compose configuration, including ai-backend, logic-backend, frontend, and database services. Simply execute docker-compose up -d to start the full service stack, which brings benefits like environment consistency, easy scalability, simplified dependencies, and convenient migration.

CNN Comparator is suitable for multiple scenarios:

1. Model Selection Decision: Quickly compare the performance of candidate models and make data-based decisions;
2. Education and Research: Help students understand the performance differences between different architectures;
3. Model Optimization Verification: Verify the effect of improved models;
4. Mobile Deployment Evaluation: Compare the balance between accuracy and speed of lightweight models.

Current Limitations: Only supports image classification tasks, limited model architectures, and incomplete distributed training; Improvement Directions: Expand support for Transformer architectures, add automatic hyperparameter search, integrate model compression tools, support multi-GPU training, and add model interpretability analysis (e.g., Grad-CAM).

CNN Comparator is a concise and practical open-source tool that accurately addresses the pain points of model selection and lowers the threshold for evaluation. Its clear three-tier architecture and Docker deployment reflect good engineering practices, making it suitable for beginners, developers, and researchers. With future feature expansions, it is expected to become a more comprehensive model comparison and evaluation platform.