finish lec5 and lec4

Author: tmp-coder

10601ML-CMU/hw3/ClassificationError.m

@@ -0,0 +1,4 @@
+function [error] = ClassificationError(yHat, yTruth)
+  error = sum(yHat != yTruth) / length(yTruth);
+end
+

10601ML-CMU/hw3/HW3Data.mat

@@ -0,0 +1,19 @@
+function [yHat] = NB_Classify(D, p, XTest)
+  m = size(XTest, 1);
+  yHat = zeros(m, 1);
+
+  for i = 1:m
+    econo_probs = D(1,:) .* XTest(i,:) + (1 - D(1,:)) .* (1 - XTest(i,:));
+    onion_probs = D(2,:) .* XTest(i,:) + (1 - D(2,:)) .* (1 - XTest(i,:));
+
+    econo_score = logProd([log(econo_probs), log(p)]);
+    onion_score = logProd([log(onion_probs), log(1-p)]);
+    
+    if econo_score > onion_score
+      yHat(i) = 1;
+    else
+      yHat(i) = 2;
+    end
+  end
+end
+

10601ML-CMU/hw3/NB_Classify.m

@@ -0,0 +1,8 @@
+function [D] = NB_XGivenY(XTrain, yTrain)
+  EconoRows = yTrain == 1;
+  OnionRows = yTrain == 2;
+
+  D = [(sum(XTrain(EconoRows,:), 1) .+ 1) / (sum(EconoRows) + 1) ;
+       (sum(XTrain(OnionRows,:), 1) .+ 1) / (sum(OnionRows) + 1)];    
+end
+

10601ML-CMU/hw3/NB_XGivenY.m

@@ -0,0 +1,4 @@
+function [p] = NB_YPrior(yTrain)
+  p = sum(yTrain == 1) / length(yTrain);
+end
+

10601ML-CMU/hw3/NB_YPrior.m

@@ -0,0 +1,3 @@
+from scipy.io import loadmat
+import numpy as np
+

10601ML-CMU/hw3/dataprocess.py

@@ -0,0 +1,5 @@
+
+hw3 终结，这是一个简单的 bag of words 的naive bayes 判别模型实践
+
+- [python code](./hw3.py)
+- [hw3 solution](./homework3_sol.pdf)

10601ML-CMU/hw3/homework3_sol.pdf

@@ -0,0 +1,63 @@
+from scipy.io import loadmat
+import numpy as np
+
+mat_file_name = "HW3Data.mat"
+data_name = ['Vocabulary', 'XTrain', 'yTrain', 'XTest', 'yTest', 'XTrainSmall','yTrainSmall']
+eps = 1e-9
+def data_preprocess(file_name):
+    data = loadmat(file_name)
+    return [data[e] for e in data_name]
+
+Vocabulary, XTrain, yTrain, XTest, yTest, XTrainSmall,yTrainSmall = data_preprocess(mat_file_name)
+
+def logProd(x):
+    return x.sum(1)
+
+def XgivenY(XTrain,YTrain):
+    """
+    return : a 2*V matrix represent P(X_w=1|Y=y) with a prior beta(beta[i]) distribution
+    """
+    yTrain = YTrain.squeeze()
+    Erows = yTrain==1
+    Orows = yTrain==2
+    return np.row_stack(((XTrain[Erows].sum(0)+1)/(Erows.sum()+1),
+    (XTrain[Orows].sum(0)+1)/(Orows.sum()+1)))
+
+def YPrior_Eco(YTrain):
+    """
+    return P(Y==1) with MLE
+    """
+    yTrain = YTrain.squeeze()
+    return np.sum(YTrain==1) / yTrain.size
+
+def classify(D,p,XTest):
+    D = np.asarray(D)
+    XTest = XTest.toarray()
+    pos_prob = D[0,:]*XTest + (1-D[0,:]) * (1-XTest)
+    neg_prob = D[1,:] * XTest+(1-D[1,:]) * ( 1 - XTest)
+    pos_prob = logProd(np.log(pos_prob+eps)) + np.log(p+eps)
+    neg_prob = logProd(np.log(neg_prob + eps)) + np.log(1-p + eps)
+    return np.argmax(np.column_stack((pos_prob,neg_prob)),axis=1) + 1
+
+def classificationErr(y_true,y_hat):
+    y_true = y_true.squeeze()
+    y_hat = y_hat.squeeze()
+    return 1 - np.sum(y_true == y_hat) / y_hat.size
+
+def model_err(XTrain,YTrain):
+    D = XgivenY(XTrain,YTrain)
+    p = YPrior_Eco(YTrain)
+    yhat_train = classify(D,p,XTrain)
+    yhat_test = classify(D,p,XTest)
+    train_err = classificationErr(YTrain,yhat_train)
+    test_err = classificationErr(yhat_test,yTest)
+    print("train err = ",train_err,"test err = ",test_err)
+
+
+def problem_g():
+    model_err(XTrain,yTrain)
+
+def problem_h():
+    model_err(XTrainSmall,yTrainSmall)
+
+    

10601ML-CMU/hw3/hw3-overview.md

@@ -0,0 +1,4 @@
+function [log_product] = logProd(x)
+  log_product = sum(x);
+end
+

10601ML-CMU/hw3/hw3.py

@@ -13,7 +13,7 @@ $$
 
     [code](./hw2.py)
 
-    ![Ltheta](img/pb2.png)
+    ![Ltheta](./img/pb2.png)
 
 3. <br>
 
@@ -27,12 +27,12 @@ $$
 
 4. <br>
 
-   ![pb2-3.png](img/pb2-3.png)
+   ![pb2-3.png](./img/pb2-3.png)
 
 # problem 3
 
 
-![pb3-1](img/pb3-1.png)
+![pb3-1](./img/pb3-1.png)
 
 **note**: