Computation-graph visualization - tensorplay.utils.viz

Rendering a graph

import tensorplay as tp
from tensorplay.utils.viz import make_dot

x = tp.randn(5, 5, requires_grad=True)
w = tp.randn(5, 5, requires_grad=True)
loss = ((x @ w).relu()).sum()

dot = make_dot(loss, params={"x": x, "w": w})
dot.render("graph", format="png")   # writes graph.png in the working directory

Node colors carry the structure:

• light blue — leaf tensors with requires_grad=True; those listed in params are labeled by the name you passed,

• light green — the output tensor handed to make_dot,

• white — the recorded operations, labeled by operation name.

Render backends

make_dot picks whichever backend is installed:

• the graphviz package returns a real graphviz.Digraph — call .render(filename, format=...) to write an image, or .pipe() for the raw bytes. Rasterizing to a file also needs the dot program on your PATH.

• otherwise networkx + matplotlib draw a hierarchical layout and return a wrapper with the same .render(filename, format="png") call.

• with neither installed, calling make_dot raises RuntimeError.

Reading the graph as text

The same chain is walkable without any rendering dependency: every tensor with a gradient history exposes its grad_fn, and every recorded operation exposes its name and the operations it consumed.

fn = loss.grad_fn
while fn is not None:
    print(fn.name)
    fn = fn.next_functions[0][0] if fn.next_functions else None