Pt Compute Loss

Here's a first for a node guide: the node's own source code says you probably shouldn't use it. Pt Compute Loss exists to compute a loss value from a model's output, the ground-truth target, and a loss object - and the author's comment is blunt: "The user does not normally need to use this node. Instead the user should use training nodes that accept a loss object."

Why the disclaimer? Because the pack's training nodes (PtTrainModel and friends) take a loss object directly and handle the loss computation internally. If you're running the standard train-a-classifier flow, you'll pick a PtnCrossEntropyLoss or PtnMSELoss node, wire it into the training node, and never touch this one. Pt Compute Loss is for the edge case where you're building a custom training loop or a validation step yourself and you want the loss value handed to you as a tensor.

Inputs

• input_tens - the model output (y_hat), a TENSOR.
• target_tens - the ground truth / labels, a TENSOR.
• loss - a PTLOSS object from one of the pack's Ptn*Loss nodes (MSE, cross-entropy, L1, KL-div, etc.).

Output: a TENSOR holding the computed loss - a scalar (0-d) tensor for most loss objects with default reduction.

Where people get burned

The classic beginner trap: feeding in input_tens and target_tens with incompatible shapes. Loss functions are picky. Cross-entropy wants class-index targets (or logits), MSE wants matching shapes - a misaligned batch or a target tensor that's one-hot when the loss expects class indices will error or, worse, silently produce a wrong-looking number. Check shapes with PtShowSize before you blame the node.

Also note this is a forward pass only - it computes the loss, it doesn't backprop. In this pack, gradients are the training nodes' job. If you find yourself wiring this into a custom loop and nothing trains, that's the missing piece.

Installing it

Ships in ComfyUI-Pt-Wrapper by Hide Inada (HowToSD). ComfyUI Manager → search "ComfyUI-Pt-Wrapper" → restart, or:

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

The pack installs a heavyweight requirements.txt - transformers, datasets, scikit-learn, scipy, gensim, pandas, peft, accelerate, and more. If Manager's install fails partway, pip install -r requirements.txt inside the clone. No model downloads for this node itself; the example training workflows auto-fetch datasets like CIFAR-10 and Fashion-MNIST.