Dependency Problems from Email

[whitead]

1. Dependencies - in all my trials, the code wouldn't work with late versions of pillow. I use:

"pillow==8.3.2"

instead of

"pillow>=8.3.2"

This isn't optimal but so far it works as long as I don't need to paste equations on Colab.

2. Under 3.8.1:

Leave one class out CV

N = len(soldata)
error = []
error_std = []
for c in unique_classes:
# slice out segments
test = soldata.loc[soldata["Group"] == c]
train = soldata.loc[soldata["Group"] != c]
test_x, test_y = test[feature_names].values, test["Solubility"].values
x, y = train[feature_names].values, train["Solubility"].values
# compute coefficients
w, *_ = np.linalg.lstsq(x, y)
# compute intercept (b)
b = np.mean(y - np.dot(x, w))
# compute test erropr
k_error.append(np.mean((np.dot(test_x, w) + b - test_y) ** 2))
error.append(np.mean(k_error))
error_std.append(np.std(k_error, ddof=1))
print(f"test error = {np.mean(error):.2f}")

There is something missing as k_error isn't initiated