Scikit-Learn（类型错误：ufunc 'subtract' 不包含签名匹配类型 dtype('

Question

I created a small MLPRegressor code for the Iris dataset that I had downloaded as a csv. Although I can use the load_iris() module for my code, my uni wants me to do it the old fashioned way.

from sklearn.neural_network import MLPRegressor
import pandas as pd

data = pd.read_csv("iris.csv", header=None)
# data = load_iris()

attributes = data.iloc[:,:-1].values.tolist()
outputs = data.iloc[:,-1].values.tolist()
clf = MLPRegressor(
    # solver="sgd", 
    # activation="logistic", 
    # learning_rate_init=0.01,
    # max_iter=10000,
    # verbose=True,
    hidden_layer_sizes=[4]
)

clf.fit(attributes, outputs)

print(clf.predict([[5.4,3.9,1.3,0.4],[6.5,3.2,5.1,2.0]]))

I get an error like this when I try to run:

Traceback (most recent call last):
  File "prog4_scikit_learn.py", line 28, in <module>
    clf.fit(attributes, outputs)
  File "/usr/local/lib/python3.6/dist-packages/sklearn/neural_network/multilayer_perceptron.py", line 620, in fit
    return self._fit(X, y, incremental=False)
  File "/usr/local/lib/python3.6/dist-packages/sklearn/neural_network/multilayer_perceptron.py", line 378, in _fit
    intercept_grads, layer_units, incremental)
  File "/usr/local/lib/python3.6/dist-packages/sklearn/neural_network/multilayer_perceptron.py", line 516, in _fit_stochastic
    coef_grads, intercept_grads)
  File "/usr/local/lib/python3.6/dist-packages/sklearn/neural_network/multilayer_perceptron.py", line 230, in _backprop
    loss = LOSS_FUNCTIONS[loss_func_name](y, activations[-1])
  File "/usr/local/lib/python3.6/dist-packages/sklearn/neural_network/_base.py", line 194, in squared_loss
    return ((y_true - y_pred) ** 2).mean() / 2
TypeError: ufunc 'subtract' did not contain a loop with signature matching types dtype('<U32') dtype('<U32') dtype('<U32')

Although I created a code using MLPClassifier for the same dataset and it works like a charm

from sklearn.neural_network import MLPRegressor
import pandas as pd

data = pd.read_csv("iris.csv", header=None)

attributes = data.iloc[:,:-1].values.tolist()
outputs = data.iloc[:,-1].values.tolist()

clf = MLPClassifier(solver="sgd", 
activation="logistic", 
learning_rate_init=0.01,
max_iter=10000,
# verbose=True,
hidden_layer_sizes=[4])
clf.fit(attributes, outputs)

print(clf.predict([[5.4,3.9,1.3,0.4],[6.5,3.2,5.1,2.0]]))

Can someone help me out

Answer 1

Check your y . In this case: 'outputs'

In a typical 'iris.csv' file, the last column denotes the species of the samples and will have values as strings

• 'setosa'
• 'versicolor'
• 'virginica'

In the classifier, that is a valid input, so it works correctly.

But for a regressor, its not. Regression is for finding a continuous numerical value as target, which you dont have in this case. Agreed that the error message should be more clear, but hope that its clear to you now.

There is no need to call .tolist() on the arrays as they can be handled directly by the scikit-learn estimators, and in any case if you use tolist() , they will be again converted to numpy arrays internally.

I think if you have not used tolist() then the error might be different and easier to understand.

Now coming to load_iris() . load_data().target contains numerical values 0,1,2 as representation of the classes. Since it is numbers, its working in regression as well as regression. Hope this makes sense.