PyTorch

Layer, model and training nodes.

18 nodes. Right-click any node in the editor to read this documentation in the app.

Torch

〰️ Activation Layer

id torch_activation_layer · Torch · Python export: yes

Element-wise activation function, chosen from a list: ReLU, Sigmoid or Tanh.

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.

Outputs

Port	Type	Description
layer	Layer	The chosen activation.

Fields

Field	Type	Default	Choices
Activation	select	relu	relu, sigmoid, tanh

➕ Add Merge

id torch_add_merge · Torch · Python export: yes

Sum two or more branches element-wise — the residual / skip connection: output = main(x) + shortcut(x). Branch shapes must match.

Inputs

Port	Type	Description
branch_1	Layer	Branch 1 to add (required).
branch_2	Layer	Branch 2 to add (required).
branch_3	Layer	Branch 3 to add (optional).
branch_4	Layer	Branch 4 to add (optional).

Outputs

Port	Type	Description
layer	Layer	Sum of the connected branches.

📶 BatchNorm2D Layer

id torch_batchnorm2d_layer · Torch · Python export: yes

Batch normalisation over the channels of a [batch, channels, H, W] tensor. Channels must equal the preceding Conv2D's out_channels.

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.
num_features	int	Number of channels (overrides the num_features field). Overrides the Channels (= out_channels of prev Conv) field when connected.

Outputs

Port	Type	Description
layer	Layer	BatchNorm2d layer.

Fields

Field	Type	Default	Choices
Channels (= out_channels of prev Conv)	int	16

🔀 Concat Merge

id torch_cat_merge · Torch · Python export: yes

Concatenate two or more branches along a dimension (Inception / DenseNet style). Dim 1 is the channel axis of [B, C, H, W] tensors; other dimensions must match.

Inputs

Port	Type	Description
branch_1	Layer	Branch 1 to concatenate (required).
branch_2	Layer	Branch 2 to concatenate (required).
branch_3	Layer	Branch 3 to concatenate (optional).
branch_4	Layer	Branch 4 to concatenate (optional).

Outputs

Port	Type	Description
layer	Layer	The branches concatenated along dim.

Fields

Field	Type	Default	Choices
Concat dim	int	1

🔲 Conv2D Layer

id torch_conv2d_layer · Torch · Python export: yes

2-D convolution over [batch, channels, height, width] images. Padding 'same' keeps height and width unchanged.

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.
in_channels	int	Channels coming in (overrides the in_channels field). Overrides the In channels field when connected.
out_channels	int	Channels going out (overrides the out_channels field). Overrides the Out channels field when connected.
kernel_size	int	Square kernel side length (overrides the kernel_size field). Overrides the Kernel size field when connected.

Outputs

Port	Type	Description
layer	Layer	Conv2d layer, ready to chain or feed a Sequential / Graph Model.

Fields

Field	Type	Default	Choices
In channels	int	3
Out channels	int	16
Kernel size	int	3
Padding	select	0	0, 1, 2, 3, same

🎲 Dropout Layer

id torch_dropout_layer · Torch · Python export: yes

Randomly zeroes a fraction p of activations during training to reduce over-fitting; does nothing at evaluation time.

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.
p	float	Probability of zeroing each activation (0-1) (overrides the p field). Overrides the Drop probability field when connected.

Outputs

Port	Type	Description
layer	Layer	Dropout layer.

Fields

Field	Type	Default	Choices
Drop probability	float	0.5

📦 Export ONNX

id torch_export_onnx · Torch · Python export: yes

Export the model to an ONNX file, tracing it with a random dummy input of the given shape (batch, channels, height, width).

Inputs

Port	Type	Description
model	Module	Model to export.

Outputs

Port	Type	Description
path	str	Path of the exported .onnx file.

Fields

Field	Type	Default	Choices
Path	text	model.onnx
Dummy input shape	text	1,3,32,32

📏 Flatten Layer

id torch_flatten_layer · Torch · Python export: yes

Flatten everything except the batch dimension, e.g. [B, C, H, W] -> [B, CHW]. Put it between conv and linear layers.

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.

Outputs

Port	Type	Description
layer	Layer	Flatten layer.

🕸 Graph Model

id torch_graph_model · Torch · Python export: yes

Assemble a whole network from wired-together layer nodes. Connect the FINAL layer(s) of your design; the node traces every upstream connection (chains, residual skips, concatenations) and wraps the graph as one model.

Inputs

Port	Type	Description
output_1	Layer	Final layer of the network.
output_2	Layer	Additional output layer (optional).
output_3	Layer	Additional output layer (optional).
output_4	Layer	Additional output layer (optional).

Outputs

Port	Type	Description
model	Module	The assembled model (single output tensor, or a tuple for several outputs).

➖ Linear Layer

id torch_linear_layer · Torch · Python export: yes

Fully-connected layer: y = xW + b, mapping in_features inputs to out_features outputs.

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.
in_features	int	Input size (overrides the in_features field). Overrides the In features field when connected.
out_features	int	Output size (overrides the out_features field). Overrides the Out features field when connected.

Outputs

Port	Type	Description
layer	Layer	Linear layer, ready to chain or feed a Sequential / Graph Model.

Fields

Field	Type	Default	Choices
In features	int	128
Out features	int	10

⬇️ MaxPool2D Layer

id torch_maxpool2d_layer · Torch · Python export: yes

2-D max pooling: keeps the largest value in each window, shrinking height and width (stride 2 halves them).

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.
kernel_size	int	Pooling window side length (overrides the kernel_size field). Overrides the Kernel size field when connected.
stride	int	Step between windows (overrides the stride field). Overrides the Stride field when connected.

Outputs

Port	Type	Description
layer	Layer	MaxPool2d layer.

Fields

Field	Type	Default	Choices
Kernel size	int	2
Stride	int	2

📚 MNIST Dataset

id mnist_dataset · Torch · Python export: yes

The MNIST handwritten-digit training set as a DataLoader (batches of 32: images shaped [32, 1, 28, 28], integer labels 0-9). Downloads to ./data on first use, so it needs network access once.

Outputs

Port	Type	Description
dataset	DataLoader	Batches of (image tensor, label) — wire into Train Image Classifier.

📈 ReLU Layer

id torch_relu_layer · Torch · Python export: yes

ReLU activation: max(0, x), element-wise.

Inputs

Port	Type	Description
upstream	Layer	Previous layer. Leave unconnected for the model's first layer.

Outputs

Port	Type	Description
layer	Layer	ReLU activation.

💾 Save Model

id torch_save_model · Torch · Python export: yes

Save the model's weights (state_dict) to a .pt file and pass the model through unchanged.

Inputs

Port	Type	Description
model	Module	Model whose weights to save.

Outputs

Port	Type	Description
model	Module	The input model, unchanged.

Fields

Field	Type	Default	Choices
Path	text	model.pt

🧱 Sequential

id torch_sequential · Torch · Python export: yes

Chain layers, in port order, into a plain nn.Sequential model. Wire layers into layer_1, layer_2, ... For branching / residual designs use Graph Model instead.

Inputs

Port	Type	Description
layer_1	Layer	Layer number 1 in the chain (optional).
layer_2	Layer	Layer number 2 in the chain (optional).
layer_3	Layer	Layer number 3 in the chain (optional).
layer_4	Layer	Layer number 4 in the chain (optional).
layer_5	Layer	Layer number 5 in the chain (optional).
layer_6	Layer	Layer number 6 in the chain (optional).
layer_7	Layer	Layer number 7 in the chain (optional).
layer_8	Layer	Layer number 8 in the chain (optional).

Outputs

Port	Type	Description
model	Module	nn.Sequential of the connected layers.

🧪 Simple CNN

id torch_simple_cnn · Torch · Python export: yes

A complete small image classifier in one node, for when you don't want to wire layers by hand: Conv2D (same padding) -> ReLU -> 2x2 MaxPool -> Flatten -> Linear. Set the image channels and size, the number of filters and kernel size, and the number of classes; the layer sizes are worked out for you. For anything deeper or branching, build the network from the layer nodes.

Inputs

Port	Type	Description
in_channels	int	Image channels (1 = greyscale, 3 = RGB) (overrides the in_channels field). Overrides the Image channels field when connected.
image_size	int	Image height and width in pixels (square images) (overrides the image_size field). Overrides the Image size (px) field when connected.
conv_channels	int	Number of convolution filters (overrides the conv_channels field). Overrides the Conv filters field when connected.
kernel_size	int	Square kernel side length (overrides the kernel_size field). Overrides the Kernel size field when connected.
num_classes	int	Number of output classes (overrides the num_classes field). Overrides the Classes field when connected.

Outputs

Port	Type	Description
model	Module	nn.Sequential classifier — wire into Train Image Classifier.

Fields

Field	Type	Default	Choices
Image channels	int	1
Image size (px)	int	28
Conv filters	int	16
Kernel size	int	3
Classes	int	10

🔥 Tensor

id torch_tensor · Torch · Python export: yes

Convert an array, DataFrame or list into a float32 torch tensor.

Inputs

Port	Type	Description
data	ndarray | DataFrame | list	Data to convert.

Outputs

Port	Type	Description
tensor	Tensor	float32 tensor with the same shape as the data.

🏋️ Train Image Classifier

id torch_train_image_classifier · Torch · Python export: yes

Train a model as an image classifier: Adam optimiser, cross-entropy loss, one pass over the DataLoader per epoch. Reports progress per epoch / batch and outputs the trained model.

Inputs

Port	Type	Description
model	Module	Model to train (Sequential or Graph Model).
dataset	DataLoader	Batches of (inputs, integer class labels).
lr	float	Adam learning rate (overrides the lr field). Overrides the Learning rate field when connected.
epochs	int	Number of passes over the data (overrides the epochs field). Overrides the Epochs field when connected.

Outputs

Port	Type	Description
model	Module	The model after training (same object, updated in place).

Fields

Field	Type	Default	Choices
Epochs	int	1
Learning rate	float	0.001