为什么我会收到日志方法的“ufunc 循环不支持 numpy.ndarray 类型的参数 0”错误？

Question

First, I used np.array to perform operations on multiple matrices, and it was successful.

 import numpy as np import matplotlib.pyplot as plt f = np.array([[0.35, 0.65]]) e = np.array([[0.92, 0.08], [0.03, 0.97]]) r = np.array([[0.95, 0.05], [0.06, 0.94]]) d = np.array([[0.99, 0.01], [0.08, 0.92]]) c = np.array([[0, 1], [1, 0]]) D = np.sum(f@(e@r@d*c)) u = f@e I = np.sum(f@(e*np.log(e/u))) print(D) print(I)

Outcome:

 0.14538525 0.45687371996485304

Next, I tried to plot the result using one of the elements in the matrix as a variable, but an error occurred.

 import numpy as np import matplotlib.pyplot as plt t = np.arange(0.01, 0.99, 0.01) f = np.array([[0.35, 0.65]]) e = np.array([[1-t, t], [0.03, 0.97]]) r = np.array([[0.95, 0.05], [0.06, 0.94]]) d = np.array([[0.99, 0.01], [0.08, 0.92]]) c = np.array([[0, 1], [1, 0]]) D = np.sum(f@(e@r@d*c)) u = f@e I = np.sum(f@(e*np.log(e/u))) plt.plot(t, D) plt.plot(t, I) plt.show()

It shows the error below:

 AttributeError Traceback (most recent call last) AttributeError: 'numpy.ndarray' object has no attribute 'log' The above exception was the direct cause of the following exception: TypeError Traceback (most recent call last) <ipython-input-14-0856df964382> in <module>() 10 11 u = f@e ---> 12 I = np.sum(f@(e*np.log(e/u))) 13 14 plt.plot(t, D) TypeError: loop of ufunc does not support argument 0 of type numpy.ndarray which has no callable log method

There was no problem with the following code, so I think there was something wrong with using np.array .

 import numpy as np import matplotlib.pyplot as plt t = np.arange(0.01, 0.99, 0.01) y = np.log(t) plt.plot(t, y) plt.show()

Any idea for this problem? Thank you very much.

Answer 1

You can't create a batch of matrices e from the variable t using the construct

 e = np.array([[1-t, t], [0.03, 0.97]])

as this would create a ragged array due to [1-t, t] and [0.03, 0.97] having different shapes. Instead, you can create e by repeating [0.03, 0.97] to match the shape of [1-t, t] , then stack them together as follows.

 t = np.arange(.01,.99,.01) # shape (98,) _t = np.stack([t, 1-t], axis=1) # shape (98, 2) e = np.array([[.03,.97]]) # shape (1, 2) e = np.repeat(e, len(ts), axis=0) # shape (98, 2) e = np.stack([_t, e], axis=1) # shape (98, 2, 2)

After this, e will be a batch of 2x2 matrices

array([[[0.01, 0.99], [0.03, 0.97]], [[0.02, 0.98], [0.03, 0.97]], [[0.03, 0.97], [0.03, 0.97]], [[0.04, 0.96], [0.03, 0.97]],...

Finally, expand other variables in the batch dimension to take advantage of numpy broadcast to batch the calculation

f = np.array([[0.35, 0.65]])[None,:] # shape (1,1,2) r = np.array([[0.95, 0.05], [0.06, 0.94]])[None,:] # shape (1,2,2) d = np.array([[0.99, 0.01], [0.08, 0.92]])[None,:] # shape (1,2,2) c = np.array([[0, 1], [1, 0]])[None,:] # shape (1,2,2)

and only sum across the last axis to get per-matrix result.

 D = np.sum(f@(e@r@d*c), axis=-1) # shape (98, 1) u = f@e I = np.sum(f@(e*np.log(e/u)), axis=-1) # shape (98, 1)