Node Types Reference
Complete reference for all 80+ node types available in the CyxWiz Node Editor.
Data Pipeline Nodes
Load data from the Dataset Registry. First node in any training pipeline.
Outputs
data(Tensor) - Feature datalabels(Labels) - Target labels
Parameters
dataset_name- Name of loaded dataset
Batch iterator with shuffle and drop_last options.
| Parameter | Default | Description |
|---|---|---|
| batch_size | 32 | Samples per batch |
| shuffle | true | Shuffle each epoch |
| drop_last | false | Drop incomplete batch |
Split data into train/validation/test sets.
| Parameter | Default | Description |
|---|---|---|
| train_ratio | 0.7 | Training set ratio |
| val_ratio | 0.15 | Validation set ratio |
| shuffle | true | Shuffle before split |
| stratify | true | Maintain class ratios |
Core Layers
| Node | Purpose | Key Parameters |
|---|---|---|
| Dense | Fully connected layer | units, use_bias, activation |
| Conv1D | 1D convolution for sequences | filters, kernel_size, stride, padding |
| Conv2D | 2D convolution for images | filters, kernel_size, stride, padding, groups |
| Conv3D | 3D convolution for volumes/video | Same as Conv2D, but 3D |
| MaxPool2D | Max pooling | kernel_size, stride, padding |
| AvgPool2D | Average pooling | kernel_size, stride, padding |
| GlobalMaxPool | Global max pooling | - |
| GlobalAvgPool | Global average pooling | - |
Normalization Layers
Batch normalization for training stability.
num_features, momentum=0.1, eps=1e-5, affine=true
Layer normalization for transformers.
normalized_shape, eps=1e-5
Group normalization.
num_groups=32, num_channels
Instance normalization.
Similar to BatchNorm
Recurrent Layers
| Node | Description | Parameters |
|---|---|---|
| RNN | Simple RNN layer | hidden_size, num_layers, bidirectional, dropout |
| LSTM | Long Short-Term Memory | hidden_size, num_layers, bidirectional, proj_size |
| GRU | Gated Recurrent Unit | hidden_size, num_layers, bidirectional |
| Embedding | Token embedding lookup | num_embeddings, embedding_dim, padding_idx |
Attention & Transformer
Multi-head attention mechanism.
embed_dim=512, num_heads=8, dropout=0.0
Self-attention (Q=K=V from single input).
Full transformer encoder block.
d_model=512, nhead=8, dim_feedforward=2048, num_layers=6
Add positional information.
d_model=512, max_len=5000, dropout=0.1
Activation Functions
| Node | Formula | Use Case |
|---|---|---|
| ReLU | max(0, x) | Most common, fast |
| LeakyReLU | max(0.01x, x) | Prevent dying neurons |
| GELU | x * Phi(x) | Transformers |
| Swish | x * sigmoid(x) | EfficientNet |
| Sigmoid | 1/(1+e^-x) | Binary output |
| Tanh | (e^x-e^-x)/(e^x+e^-x) | Range [-1,1] |
| Softmax | e^xi / sum(e^xj) | Classification |
Loss Functions
| Node | Use Case | Formula |
|---|---|---|
| MSELoss | Regression | mean((y-y')^2) |
| CrossEntropyLoss | Multi-class | -sum(y*log(y')) |
| BCELoss | Binary | -[y*log(y') + ...] |
| L1Loss | Robust regression | mean(abs(y-y')) |
| HuberLoss | Robust | MSE if small, L1 if large |
Optimizers
| Node | Algorithm | Good For |
|---|---|---|
| SGD | Stochastic Gradient Descent | Simple, well-tuned |
| Adam | Adaptive moments | Most tasks |
| AdamW | Adam with weight decay | Transformers |
| RMSprop | Root mean square prop | RNNs |
Output Node
Required: Every training graph must end with an Output node.
Inputs
prediction(Tensor) - Model outputloss(Loss) - Loss value for training