eTorch

eTorch provides the emergent GUI and other support for PyTorch networks, including an interactive 3D NetView for visualizing network dynamics, and other GUI elements for controlling the model and plotting training and testing performance, etc.

The key idea for the NetView is that each etorch.Layer stores the state variables as a etensor.Float32, which are just copied via Python code from the torch.FloatTensor state values recorded from running the network.

The etor python-side library provides a State object that handles the recording of state during the forward pass through a torch model. You just need to call the rec method for each step that you want to record. The set_net method is called with the torch.Network to record state to.

Here's the forward code for the alexnet example:

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        self.est.rec(x, "Image.Act")
        for i, f in enumerate(self.features):
            x = f(x)
            self.est.rec(x, self.fnames[i])
        x = self.avgpool(x)
        self.est.rec(x, "AP.Act")
        x = torch.flatten(x, 1)
        for i, f in enumerate(self.classifier):
            x = f(x)
            cnm = self.cnames[i]
            if len(cnm) > 0:
                self.est.rec(x, cnm)
        return x

Where self.fnames are the layer.variable names to save for each step in updating the features sequence, while self.cnames is the equivalent for the classifier.

As usual, the best way to see everything is to check out the examples:

• etra25 is a torch version of the leabra/examples/ra25 random associator, using all of the same overall program infrastructure. This provides full training and testing control, weight visualization, etc, and is a good reference for various code examples. It is very much Go-based in design, having been translated from Go source originally, so it may not be as familiar for typical Python users, but it does show what kind of overall complete GUI you can create.

• alexnet is the standard torchvision alexnet example, showing how large convolutional neural networks look with the visualization. Because these models are so ... visual, you can really see what each step is doing. This setup is only for testing, and doesn't show weights.

Installation

See the https://github.com/emer/etorch/tree/main/python directory for instructions on building an etorch program that is just like the python3 executable, but also includes all of the Go-based infrastructure that enables etorch to work.

You can also use the pyleabra executable from https://github.com/emer/leabra/tree/master/python, which includes etorch to facilitate interoperability between leabra and torch models.

Interoperating between Go and Python

See etable pyet for example code for converting between the Go etable.Table and numpy, torch, and pandas table structures, using the pyet Python library that is installed with etorch.