Graph Visualization

MiniTorchBR can render the computation graph as an interactive HTML file using PyVis.

Basic Graph Rendering

import numpy as np
from MiniTorch.core.variable import Variable
from MiniTorch.utils.visualize import visualize_graph

x = Variable(np.array([1.0, 2.0]))
w = Variable(np.array([0.5, -0.5]))

y = (x * w).sum()
y.backward()

# Renders an interactive HTML file
visualize_graph(y, output_path="graph.html")

Open graph.html in any browser. Nodes represent variables and edges represent operations. You can drag, zoom, and hover for details.

Larger Network Graph

from MiniTorch.nn.linear import Linear
from MiniTorch.ops.relu import relu
from MiniTorch.ops.softmax_cross_entropy import softmax_cross_entropy

x_np = np.random.randn(4, 8)
x = Variable(x_np)

fc1 = Linear(8, 4)
fc2 = Linear(4, 2)

h = relu(fc1(x))
logits = fc2(h)
loss = softmax_cross_entropy(logits, np.array([0, 1, 0, 1]))
loss.backward()

visualize_graph(loss, output_path="network_graph.html")

Training Curves

from MiniTorch.utils.training_viz import plot_training_history

history = {
    "train_loss": [0.95, 0.72, 0.55, 0.42, 0.34],
    "val_loss":   [0.98, 0.75, 0.58, 0.46, 0.39],
    "train_acc":  [0.65, 0.76, 0.83, 0.87, 0.90],
    "val_acc":    [0.63, 0.74, 0.81, 0.85, 0.88],
}

plot_training_history(history)   # saves training_history.png

Weight Histograms

from MiniTorch.utils.training_viz import plot_weight_histograms

plot_weight_histograms(model)    # saves weight_histograms.png

Histograms show the distribution of weights in each layer — useful for diagnosing vanishing/exploding gradients.

Confusion Matrix

from MiniTorch.utils.training_viz import plot_confusion_matrix
import numpy as np

y_true = np.array([0, 1, 2, 1, 0, 2])
y_pred = np.array([0, 1, 2, 0, 0, 1])

plot_confusion_matrix(y_true, y_pred, class_names=["cat", "dog", "bird"])
# saves confusion_matrix.png