first refine

hw1/linreg.py

@@ -9,13 +9,16 @@ def ensure_dir(file_path):
  if not os.path.exists(directory):
  os.makedirs(directory)
 
-def extract_feature(M, features):
+def extract_feature(M, features, squares):
  x_data = []
  y_data = []
  for month in range(M.shape[0]):
  for i in range(M.shape[2]-10+1):
- x_data.append(M[month, features, i:i+9].flatten().astype("float"))
- y_data.append(float(M[month, 9, i+9]))
+ X = M[month, features, i:i+9].flatten()
+ Y = M[month, squares, i:i+9].flatten()
+ Z = np.concatenate((X, Y), axis=0)
+ x_data.append(Z)
+ y_data.append(M[month, 9, i+9])
  return np.array(x_data), np.array(y_data)
 
 # Start Program
@@ -26,27 +29,28 @@ def extract_feature(M, features):
 M = M[:, 3:] #shape: (4320, 24)
 M = np.reshape(M, (12, -1, 18, 24)) #shape: (12, 20, 18, 24)
 M = M.swapaxes(1, 2).reshape(12, 18, -1) #shape: (12, 18, 480)
-## map 'NR' -> '-1'
-for month in range(12):
- M[month, 10, :] = np.array(list(map(lambda i: i if i != 'NR' else '-1', M[month, 10, :])))
+M[M == 'NR'] = '0.0'
+M = M.astype(float)
 
 
 # extract feature into x_data <shape:(5652, 9*18)>, y_data <shape:(5652,)>
-feature_sieve = [i for i in range(0, 18)]
-x_data, y_data = extract_feature(M, feature_sieve)
+feature_sieve = [i for i in range(18)]
+square_sieve = []
+length = len(feature_sieve) + len(square_sieve)
+x_data, y_data = extract_feature(M, feature_sieve, square_sieve)
 
 # ydata = b + w * xdata
 b = 0.0
-w = np.zeros(len(feature_sieve)*9)
-lr = 5e-3
+w = np.zeros(length*9)
+lr = 0.5
 epoch = 200000
 b_lr = 0.0
-w_lr = np.zeros(len(feature_sieve)*9)
+w_lr = np.zeros(length*9)
 
 prev_loss = 1e10
 for e in range(epoch):
  b_grad = 0.0
- w_grad = np.zeros(len(feature_sieve)*9)
+ w_grad = np.zeros(length*9)
  loss = 0.0
 
  # Calculate the value of the loss function
@@ -64,10 +68,10 @@ def extract_feature(M, features):
  w = w - lr/np.sqrt(w_lr) * w_grad
 
  # Print loss
- if e % 100 == 0:
- print('epoch:{}\n Loss:{}'.format(e, loss))
- if prev_loss - loss < 1e-8: break
- prev_loss = loss
+ if (e+1) % 1000 == 0:
+ print('epoch:{}\n Loss:{}'.format(e+1, np.sqrt(loss)))
+ # if prev_loss - loss < 1e-8: break
+ # prev_loss = loss
 
 
 # Test
@@ -82,16 +86,18 @@ def extract_feature(M, features):
  f.write('{}\n'.format(','.join(list(map(lambda x: str(x), w.flatten())))))
 
 with open(outfile, 'w+') as f:
- f.write('id,value\n')
  M = pd.read_csv(infile2, header=None, encoding='big5').as_matrix()
+ M = M[:, 2:] #shape: (4320, 9)
+ M = M.reshape(-1, 18, 9) #shape: (240, 18, 9)
+ M[M == 'NR'] = '0.0'
+ M = M.astype(float)
 
- i = 0
- while i < M.shape[0]:
- ## map 'NR' -> '-1'
- M[i+10, :] = np.array(list(map(lambda x: x if x != 'NR' else '-1', M[i+10, :])))
+ selected = [i for i in feature_sieve]
+ square_selected = [i for i in square_sieve]
 
- modified_sieve = [n+i for n in feature_sieve]
- X = M[modified_sieve, 2:].flatten().astype("float")
- y = M[i, 0]
- f.write('{},{}\n'.format(y, b + np.dot(w, X)))
- i += 18
+ f.write('id,value\n')
+ for i in range(M.shape[0]):
+ X = M[i, selected, :].flatten()
+ Y = M[i, square_selected, :].flatten()
+ Z = np.concatenate((X, Y), axis=0)
+ f.write('id_{},{}\n'.format(i, b + np.dot(w, Z)))