plot_gradient_flow.py

import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
import numpy as np


def plot_grad_flow(model):
    ave_grads = []
    max_grads = []
    layers = []
    for name, param in model.named_parameters():
        if param.requires_grad and "bias" not in name:
            if param.grad is None:
                print("[!] No grad:", name)
                continue
            layers.append(name)
            ave_grads.append(param.grad.abs().mean())
            max_grads.append(param.grad.abs().max())

    plt.bar(np.arange(len(max_grads)), max_grads, alpha=0.1, lw=1, color="c")
    plt.bar(np.arange(len(max_grads)), ave_grads, alpha=0.1, lw=1, color="b")
    plt.hlines(0, 0, len(ave_grads) + 1, lw=2, color="k")
    plt.xticks(range(0, len(ave_grads), 1), layers, rotation="vertical")
    plt.xlim(left=0, right=len(ave_grads))
    plt.ylim(bottom=-0.001, top=0.02)
    plt.xlabel("Layers")
    plt.ylabel("average gradient")
    plt.title("Gradient flow")
    plt.grid(True)
    plt.legend(
        [Line2D([0], [0], color=x, lw=4) for x in "cbk"],
        ["max-gradient", "mean-gradient", "zero-gradient"],
    )


if __name__ == "__main__":
    model = ...
    inputs = ...
    labels = ...
    criterion = ...

    outputs = model.forward(inputs)
    loss = criterion(outputs, labels)
    loss.backward()
    plot_grad_flow(model)