DocsPatterns
Architecture Patterns
Pre-built neural network architecture patterns and templates to accelerate model development.
Overview
The Patterns Library provides:
- Pre-built Templates: Common architectures ready to use
- Best Practices: Proven network configurations
- Customizable: Modify any pattern to fit your needs
- Documented: Each pattern includes usage notes
Using Patterns
- Open the Node Editor (
Ctrl+1) - Click File → Load Pattern or press
Ctrl+Shift+P - Browse patterns by category
- Click a pattern to preview
- Click Insert to add to canvas
- Connect your data input and output
Classification Patterns
Simple MLPBeginner
3-layer fully connected network for tabular data.
Input → Dense(128) → ReLU → Dense(64) → ReLU → Dense(n_classes)
LeNet-5Classic
Classic CNN for image classification.
Conv2D(6) → Pool → Conv2D(16) → Pool → Dense(120) → Dense(84)
VGG BlockBuilding Block
Reusable VGG-style convolutional block.
Conv2D(64) → Conv2D(64) → MaxPool2D → BatchNorm
ResNet BlockAdvanced
Residual block with skip connection.
Input → Conv → BN → ReLU → Conv → BN → Add(Input) → ReLU
Sequence Patterns
Simple LSTMBeginner
Basic LSTM for sequence classification.
Embedding → LSTM(64) → Dense(32) → Dense(n_classes)
Bidirectional LSTMIntermediate
Bidirectional LSTM for better context.
Embedding → Bidirectional(LSTM) → Dropout → Dense
Seq2SeqAdvanced
Encoder-decoder for sequence translation.
Encoder: LSTM(hidden) Decoder: LSTM(hidden) → Dense
Transformer EncoderAdvanced
Self-attention based encoder.
PosEncoding → MultiHeadAttn → LayerNorm → FFN → LayerNorm
Autoencoder Patterns
Simple Autoencoder
Dense autoencoder for dimensionality reduction.
Encoder: Dense(128) → Dense(32) Decoder: Dense(128) → Dense(input_dim)
Variational Autoencoder
VAE with reparameterization trick.
Encoder → μ, σ → Sample(z) → Decoder → Reconstruction
Conv Autoencoder
Convolutional autoencoder for images.
Conv → Pool → Conv → Pool → ConvT → ConvT → Conv
Denoising Autoencoder
Learns to remove noise from inputs.
AddNoise → Encoder → Latent → Decoder → Clean
Skip Connection Patterns
Residual (Add)
ResNet-style: add input to output.
x + F(x)
Dense (Concat)
DenseNet-style: concatenate features.
[x, F(x)]
U-Net Skip
Encoder to decoder connections.
Concat(enc, dec)
Creating Custom Patterns
- Build your architecture in the Node Editor
- Select all nodes you want to save (
Ctrl+A) - Right-click → Save as Pattern
- Enter pattern name and description
- Choose category
- Pattern is saved to
~/.cyxwiz/patterns/
Pattern Categories
| Category | Use Case |
|---|---|
| Classification | Image, text, tabular classification |
| Regression | Continuous value prediction |
| Sequence | Time series, NLP, translation |
| Generative | Autoencoders, GANs, VAEs |
| Segmentation | U-Net, semantic segmentation |
| Building Blocks | Reusable components (ResBlock, etc.) |