completed machine learning toolbox

learning_curve.py

@@ -7,17 +7,32 @@
 from sklearn.linear_model import LogisticRegression
 
 data = load_digits()
-print data.DESCR
 num_trials = 10
 train_percentages = range(5,95,5)
 test_accuracies = numpy.zeros(len(train_percentages))
 
 # train a model with training percentages between 5 and 90 (see train_percentages) and evaluate
 # the resultant accuracy.
 # You should repeat each training percentage num_trials times to smooth out variability
-# for consistency with the previous example use model = LogisticRegression(C=10**-10) for your learner
+# for consistency with the previous example use model = LogisticRegression(C=10**-10) for your learneron
+
+for j in range(len(train_percentages)):
+
+	total_accuracy = 0
+
+	for i in range(num_trials):
+
+		t = train_percentages[j] / 100.0 #converts percentage into decimal 
+
+		X_train, X_test, y_train, y_test = train_test_split(data.data, data.target, train_size = t)
+		model = LogisticRegression(C=10**-10)
+		model.fit(X_train, y_train)
+		total_accuracy += model.score(X_test,y_test)
+
+	avg_accuracy = total_accuracy / num_trials
+
+	test_accuracies[j] = avg_accuracy
 
-# TODO: your code here
 
 fig = plt.figure()
 plt.plot(train_percentages, test_accuracies)

questions.txt

@@ -0,0 +1,4 @@
+1. The curve generally increases and is fairly linear.
+2. The beginning part of the curve generally seems to be a lot noisier than the end of the curve. This is probably due to the fact that when the computer is only learning from 5% of the data, it is basically guessing for the entirety of the testing set of numbers, which leads to inaccurate data.
+3. I got a fairly smooth curve with about 100 trials. 70 trials was still a bit bumpy. With 200 trials, there were still some bumps at the beginning of the curve.
+4. When I changed C to 1/10, the graph appeared to be more exponential than linear. The test accuracy numbers were much higher for all percentages. The graph was also fairly smooth. When I changed C to  10^(-15), the numbers were about the same as 10^(-10), but the graph was much bumpier.