Changelog

All notable changes to this project are documented here.

Version 2.1.0 (2025)

🚀 Major Release: Hybrid Models & Pre-trained Customization

New Features:

  • ✨ Hybrid Builder: Customize pre-trained torchvision models with patches

  • ✨ Weight Utilities: Smart partial loading and weight transfer

  • ✨ 12 New Blocks: CBAM, ECA, DropPath, Mish, GeM, MBConv, and more

  • ✨ Enhanced YAML Recipes: Macros, inheritance, and JSONSchema validation

  • ✨ CLI Tools: Build, benchmark, validate, and export from command line

  • ✨ 40+ Supported Backbones: ResNet, EfficientNet, ConvNeXt, MobileNet, ViT, Swin

New Modules:

  • torchvision_customizer.hybrid - Hybrid model building

  • torchvision_customizer.hybrid.weight_utils - Weight loading utilities

  • torchvision_customizer.hybrid.extractor - Backbone structure extraction

  • torchvision_customizer.recipe.schema - JSONSchema validation

  • torchvision_customizer.recipe.yaml_loader - Enhanced YAML loading

  • torchvision_customizer.cli - Command-line interface

New Blocks:

  • CBAMBlock - Convolutional Block Attention Module

  • ECABlock - Efficient Channel Attention

  • DropPath - Stochastic Depth regularization

  • Mish - Self-regularizing activation

  • GeM - Generalized Mean Pooling

  • MBConv - Mobile Inverted Bottleneck

  • FusedMBConv - Fused MBConv (EfficientNetV2)

  • LayerScale - Per-channel scaling

  • CoordConv - Coordinate Convolution

  • SqueezeExcitation - Enhanced SE block

  • ConvBNAct - Conv + BN + Activation combo

  • GELUActivation - GELU with approximate mode

CLI Commands:

tvc build --yaml model.yaml
tvc benchmark --yaml model.yaml
tvc validate --yaml model.yaml
tvc export --yaml model.yaml --format onnx
tvc list-backbones
tvc list-blocks
tvc create-recipe --template hybrid_resnet_se

Improvements:

  • 📚 Complete documentation rewrite

  • 🎨 Better README with visual design

  • ✅ Comprehensive test suite for hybrid module

  • 🔧 Registry now includes 30+ blocks

Version 2.0.0 (2025)

🔄 Major Release: 3-Tier API Architecture

Complete paradigm shift from monolithic CustomCNN to component-based architecture.

Breaking Changes:

  • ❌ Removed CustomCNN class

  • ❌ Removed torchvision_customizer.models module

  • ⚠️ Minimum Python version: 3.9 (previously 3.13)

New Features:

  • ✨ Tier 1: Component Registry - Centralized block discovery

  • ✨ Tier 2: Architecture Recipes - Declarative model definitions

  • ✨ Tier 3: Model Composer - Fluent API with >> operators

  • ✨ Parametric Templates - ResNet, VGG, MobileNet, DenseNet, EfficientNet

API Examples:

# Registry (Tier 1)
block = registry.get('residual')(64, 64)

# Recipes (Tier 2)
recipe = Recipe(stem="conv(64)", stages=["residual(64) x 2"], head="linear(10)")
model = build_recipe(recipe)

# Composer (Tier 3)
model = Stem(64) >> Stage(128, blocks=2) >> Head(10)

Version 1.0.0 (2025-01-17)

Initial Release

Features:

  • ✨ 90+ pre-built components (blocks, layers, models)

  • ✨ 16 modules for comprehensive CNN building

  • ✨ Multiple APIs (Simple, Builder, Advanced)

  • ✨ Configuration-based model definition (YAML/JSON)

  • ✨ Model analysis utilities (FLOPs, memory, summary)

  • ✨ 8 complete development steps with full documentation

  • ✨ 113 working examples

  • ✨ 100% type hints coverage

  • ✨ Comprehensive API documentation

Testing:

  • ✅ 382 tests with 100% pass rate

  • ✅ 100% code coverage

  • ✅ Cross-platform compatibility

  • ✅ GPU and CPU support

Roadmap

Version 2.2.0 (Planned)

  • ONNX optimization and INT8 quantization

  • AutoML architecture search integration

  • Multi-GPU training utilities (DDP wrappers)

  • Pre-built fine-tuning pipelines

  • Hugging Face Hub integration

Version 2.3.0 (Planned)

  • Knowledge distillation utilities

  • Pruning and model compression

  • TensorRT export support

  • Mobile deployment tools

Version 3.0.0 (Future)

  • 3D networks for video/medical imaging

  • Graph neural networks

  • Multi-modal architectures

  • Advanced quantization-aware training

Release Timeline

  • v1.0.0: 2025-01-17 ✅

  • v2.0.0: 2025 ✅

  • v2.1.0: 2025 ✅ (Current)

  • v2.2.0: Q2 2025

  • v2.3.0: Q3 2025

  • v3.0.0: Q4 2025+

Breaking Changes Policy

  • All breaking changes happen in major versions (v2.0.0, v3.0.0, etc.)

  • Deprecations are announced 1 major version in advance

  • Minor versions (v2.1.0, v2.2.0, etc.) are backward compatible

  • Patch versions (v2.1.1, v2.1.2, etc.) fix bugs without API changes

Migration Guides

v1.x → v2.0

# Old (v1.x)
model = CustomCNN(layers=50, architecture='resnet')

# New (v2.0+)
model = resnet(layers=50, num_classes=1000)
# or
model = Stem(64) >> Stage(128, blocks=2) >> Head(1000)

v2.0 → v2.1

v2.1 is fully backward compatible with v2.0. All v2.0 code works unchanged.

New v2.1 features are additive:

# v2.0 code still works
model = Stem(64) >> Stage(128, blocks=2) >> Head(10)

# v2.1 adds hybrid models
model = HybridBuilder().from_torchvision("resnet50", ...)

Contributing

Contributions welcome! See Contributing Guide for guidelines.

License

MIT License - See LICENSE file for details.