Use np.full?

[oyamad]

In several pieces of code, an array of a constant value is created by something like

Q = np.ones((num_nodes, num_nodes))
Q = Q * np.inf

This can be simplified with np.full:

Q = np.full((num_nodes, num_nodes), np.inf)

[mmcky]

thanks @oyamad I am all for this.

@shlff would you have time to go through the lectures and make a PR?

[jstac]

Thanks @oyamad , I didn't know about np.full.

@shlff please go ahead when you have time. In this and other lectures, you can search for np.ones and see if it's being used in the pattern that @oyamad mentioned (creation of a constant vector).

[shlff]

Thanks @jstac , @oyamad and @mmcky .

I went through all lectures in lecture-python.myst and find similar patterns in the following lectures:

• career: v = np.ones((cw.grid_size, cw.grid_size)) * 100
• geom_series: f_2 = np.ones(T_max+1)*infinite_lease(g, r, x_0)
• inventory_dynamics: X = np.ones(num_firms) * x_init
• lqcontrol: income_r = np.ones(T-K) * s
• markov_asset: n = 5 and P = 0.0125 * np.ones((n, n))
• mccall_correlated: f_init = np.log(js.c) * np.ones(len(js.z_grid))
• multi_hyper: Σ = np.ones((c, c)) * n * (N - n) / (N - 1) / N ** 2
• multivariate_normal: μ_IQ = np.ones(n+1) * μθ and μb = np.ones(T) * 𝛼0
• odu: e = np.ones((N, T+1)) and π = np.ones((N, T+1)) * 1e-3
• short_path: Q = np.ones((num_nodes, num_nodes)) and Q = Q * np.inf (This one is exactly what @oyamad mentioned.)
• time_series_with_matrices: b = np.ones(T) * 𝛼0 and b_steady = np.ones(T) * 𝛼0
• wealth_dynamics: ψ_0 = np.ones(num_households) * wdy.y_mean and ψ_0 = np.ones(num_households) * wdy.y_mean and ψ_0 = np.ones(num_households) * wdy.y_mean

I am not really sure which one is better: Q = np.ones((num_nodes, num_nodes)) * np.inf or Q = np.full((num_nodes, num_nodes), np.inf).

But maybe we can implement some simple experiments to see compare the speed and stability (e.g., under numba) for these two cases?

[jstac]

Thanks @shlff .

A few of these do not fit the pattern @oyamad identified: using np.ones to create an array and then multiplying it by a constant to create a constant vector.

Can you please cut down this list to just those examples?

(Once we have a list fitting this pattern, we will change them all to use np.full, since dealing with this pattern is the purpose of that function.)

[shlff]

Thanks @jstac . Please find the updated list above.

Looking forward to your comments.

[jstac]

Thanks @shlff , good work. That list is correct so please proceed.

[shlff]

Thanks @jstac . In the next step I will create a PR to replace all patterns in that list with np.full, suggested by @oyamad .

[oyamad]

There are a few more:

• inventory_dynamics.md

• lake_model.md

• complex_and_trig.md

• multivariate_normal.md

• samuelson.md
  Use np.full_like.

• markov_asset.md

  P += np.diag(0.95 - 0.0125 * np.ones(5))

  This does not seem to be very clever. It should be, for example,

  P[range(n), range(n)] += 1 - P.sum(1)

  or

  P[np.diag_indices(n)] += 1 - P.sum(1)

[shlff]

Thanks @oyamad . I will check the pattern through the lectures you mentioned above and create another PR to modify it.