Ptn Linear Model

If PtnLinear is one dense layer, this is the entire multi-layer perceptron in a single node. You type a list of dimensions, it builds the whole stack, and you get a ready-to-train PTMODEL. For the pack's canonical use case - MNIST/Fashion-MNIST classification with a 784-pixel input - this is the node from the fashion_mnist_train.json workflow. When you want to test an MLP baseline before reaching for a conv net or transformer, this is the fastest way to stand one up.

How it works

PtnLinearModel builds a DenseModel: every hidden layer is nn.Linear followed by a ReLU activation, and the final layer is a bare nn.Linear with no activation - you're expected to pair it with a loss that brings its own, like Ptn BCE With Logits Loss or Ptn Cross Entropy Loss. Internally it also flattens the input (inputs.view(-1, dim_list[0])) before the first layer, which saves you a separate flattening step as long as your input's feature count matches dim_list[0].

The inputs that matter

• dim_list (default "[784,10]") - a Python-list string of dimensions. It holds the input size first, then each layer's output size. [784,128,10] means input 784, hidden 128, output 10. The list is always one element longer than the number of layers.
• bias_list (default "[True]") - a string list of True/False, one per layer, controlling bias on each dense layer.
• num_layers (default 1) - how many layers to build. Must satisfy len(dim_list) - 1 == num_layers, or the node throws a validation error telling you exactly that.

Output is a single PTMODEL. Note the string-vs-int gotcha: you're typing "[784,10]" with brackets, not 784,10 - the node ast.literal_evals the string, so malformed lists are a runtime error.

How you'd use it

The Fashion-MNIST example chains PtnPreFlatten before this node so 28×28 images arrive as 784-vectors, then feeds the model into Pt Train Classification Model with an optimizer, a loss, and Pt Save Model on the output. For a quick experiment, [784,128,10] with num_layers=2 and bias_list=[True,True] is a sensible two-layer start.

Installing

Standard for the pack. ComfyUI Manager → search "Pt-Wrapper", or:

cd ComfyUI/custom_nodes
git clone https://github.com/HowToSD/ComfyUI-Pt-Wrapper

Restart. First launch installs the pack's heavy requirements (scikit-learn, pandas, transformers, etc.) - slow once, fine after.

Where people get burned

Nearly every failure is a list/count mismatch: dim_list shorter than num_layers + 1, or bias_list the wrong length. The node is good about raising a clear ValueError, but the fix is on you. Also remember hidden layers all get ReLU, so if you wanted a different activation (or none) you'd chain individual PtnLinear layers instead of using this convenience node.