1.1 KiB
Building a Model from Scratch
ComfyUI Pt Wrapper provides composite models that can be easily customized by adjusting parameters. However, you can also build a model from scratch by sequentially combining basic models such as Conv2D and Linear.
To chain two models, use the Pt Chained Model node. This node also supports specifying post-processing after the forward pass of the second model (e.g., activation functions). To define the post-processing step, connect a Ptf node (e.g., Ptf ReLU).
In some cases, preprocessing may be required before feeding data into the model. For instance, you can use the Pt Pre Add Channel Axis node to add a channel dimension to the input tensor or the Pt Pre Flatten node to flatten convolutional output before passing it to a linear layer.
The following example workflows demonstrate this approach:
./examples/workflows/fashion_mnist_eval_conv_chained_model.json: Constructs a multi-layer convolutional network from scratch../examples/workflows/fashion_mnist_train_chained_model.json: Uses the trained model for evaluation.