Awesome

Simultaneous Perturbation Stochastic Approximation (SPSA)

Python implementation of the SPSA algorithm [1]. This is a minimisation algorithm based on gradient descent. The advantage of SPSA is that the complexity does not scale too much with number of parameters, as only two function evaluations are required per iteration regardless of the number of variables. It has also been shown to improve the training time of neural networks in crtain cases, by substituting backpropagation for SPSA [2].

Documentation

SPSA(f, theta, n_iter, extra_params = False, theta_min = None, theta_max = None, report = False, constats = constats, return_progress = False)

• Parameters:

  • f: Function to be minimised (func)
  • theta: Initial parameter guess (np.array)
  • n_iter: Number of iterations (int)
  • extra_params: Extra parameters taken by f (np.array)
  • theta_min: Minimum value of theta (np.array)
  • theta_max: Maximum value of theta (np.array)
  • report: Print progress. If False, nothing is printed. If n (int), every n iterations print the iteration number, function value and parameter values (bool / int)
  • constats: Constants needed for the gradient descent (dict). Default is {"alpha": 0.602, "gamma": 0.101, "a": 0.6283185307179586, "c": 0.1, "A": False}
  • return_progress: If False, nothing is else is returned. If n (int), return the iteration number, increasing by n, and the function value at each iteration (bool / int)

• Returns:

  • theta: Optimum parameters values to minimise f (np.array)
  • f(theta): Minimum value found (float)
  • If return_progress == True:
    • progress: Array with all the function values at each return_progress iteration (np.array)

• Carries out the SPSA algorithm

plot_progress(progress, title = False, xlabel = False, ylabel = False, save = False)

• Parameters:
  • progress: Third output from SPSA (np.array)
  • title: Graph title (str)
  • xlabel: Label for the x axis. Use r"$$" for latex formatting (str)
  • ylabel: Label for the y axis. Use r"$$" for latex formatting (str)
  • save: If not False, save the graph with the name given (bool / str)
• Plots the function value against iteration number

References

[1] Spall, J. C. An Overview of the Simultaneous Perturbation Method for Efficient Optimization. Johns Hopkins APL Technical Digest. 1998; 4 (19): 482-492.
[2] Wulff, Benjamin & Schücker, Jannis & Bauckhage, Christian. (2018). SPSA for Layer-Wise Training of Deep Networks. 10.1007/978-3-030-01424-7_55.