add log(0, 3) + z-score normalization

hw2/best.py

@@ -26,18 +26,19 @@ def sigmoid(z):
 cubic = [0, 1, 3, 4, 5]
 X_train = np.concatenate((X_train,
  X_train[:, square]**2,
- X_train[:, cubic]**3), axis=1)
+ X_train[:, cubic]**3,
+ np.log(X_train[:, [0, 3]] + 1e-100)), axis=1)
 
 # scaling: only on features, not label
-maxnum = np.max(X_train, axis=0) #shape: (106,)
-minnum = np.min(X_train, axis=0) #shape: (106,)
-X_train = (X_train - minnum) / (maxnum - minnum + 1e-100)
+mean = np.mean(X_train, axis=0) #shape: (106,)
+std = np.std(X_train, axis=0) #shape: (106,)
+X_train = (X_train - mean) / (std + 1e-100)
 
 # initialize
 b = 0.0
 w = np.ones(X_train.shape[1])
 lr = 5e-1
-epoch = 10000
+epoch = 3000
 b_lr = 0.0
 w_lr = np.zeros(X_train.shape[1])
 
@@ -64,10 +65,8 @@ def sigmoid(z):
  if (e+1) % 100 == 0:
  f[f >= 0.5] = 1
  f[f < 0.5] = 0
- acc = Y_train - f #shape: (32561,)
- acc[acc == 0] = 2
- acc[acc != 2] = 0
- print('epoch:{}\n Loss:{}\n Accuracy:{}%\n'.format(e+1, loss, np.sum(acc) * 50 / acc.shape[0]))
+ acc = Y_train == f #shape: (32561,)
+ print('epoch:{}\n Loss:{}\n Accuracy:{}%\n'.format(e+1, loss, np.sum(acc) * 100 / acc.shape[0]))
 
 
 # Test
@@ -87,11 +86,12 @@ def sigmoid(z):
  # X_test select
  X_test = np.concatenate((X_test,
  X_test[:, square]**2,
- X_test[:, cubic]**3), axis=1)
+ X_test[:, cubic]**3,
+ np.log(X_test[:, [0, 3]] + 1e-100)), axis=1)
  ans = []
  for i in range(X_test.shape[0]):
  Z = X_test[i]
- Z = (Z - minnum) / (maxnum - minnum + 1e-100)
+ Z = (Z - mean) / (std + 1e-100)
  z = np.dot(w, Z) + b
  if (z >= 0):
  ans.append('{},{}'.format(i+1, 1))