a = tf.constant(3)
b = tf.constant(3,shape=[2,2]) 

w = tf.Variable(b)

tf.matmul(m,n)
tf.conv2d(...)

#根據(jù)給定的shape定義并初始化卷積核的權(quán)值變量
  def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)

  #根據(jù)shape初始化bias變量
  def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)

def conv2d(x,W):
  return tf.nn.cov2d(x,W,strides=[1,1,1,1],padding='VALID')

def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME')

  x = tf.placeholder("float", shape=[None, 96, 96, 1])
  y_ = tf.placeholder("float", shape=[None, 30])
  keep_prob = tf.placeholder("float")

  def model():
    W_conv1 = weight_variable([3, 3, 1, 32])
    b_conv1 = bias_variable([32])

    h_conv1 = tf.nn.relu(conv2d(x, W_conv1) + b_conv1)
    h_pool1 = max_pool_2x2(h_conv1)

    W_conv2 = weight_variable([2, 2, 32, 64])
    b_conv2 = bias_variable([64])

    h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
    h_pool2 = max_pool_2x2(h_conv2)

    W_conv3 = weight_variable([2, 2, 64, 128])
    b_conv3 = bias_variable([128])

    h_conv3 = tf.nn.relu(conv2d(h_pool2, W_conv3) + b_conv3)
    h_pool3 = max_pool_2x2(h_conv3)

    W_fc1 = weight_variable([11 * 11 * 128, 500])
    b_fc1 = bias_variable([500])

    h_pool3_flat = tf.reshape(h_pool3, [-1, 11 * 11 * 128])
    h_fc1 = tf.nn.relu(tf.matmul(h_pool3_flat, W_fc1) + b_fc1)

    W_fc2 = weight_variable([500, 500])
    b_fc2 = bias_variable([500])

    h_fc2 = tf.nn.relu(tf.matmul(h_fc1, W_fc2) + b_fc2)
    h_fc2_drop = tf.nn.dropout(h_fc2, keep_prob)

    W_fc3 = weight_variable([500, 30])
    b_fc3 = bias_variable([30])

    y_conv = tf.matmul(h_fc2_drop, W_fc3) + b_fc3
    rmse = tf.sqrt(tf.reduce_mean(tf.square(y_ - y_conv)))
    return y_conv, rmse

  import pandas as pd
  import numpy as np

  TRAIN_FILE = 'training.csv'
  TEST_FILE = 'test.csv'
  SAVE_PATH = 'model'


  VALIDATION_SIZE = 100  #驗(yàn)證集大小
  EPOCHS = 100       #迭代次數(shù)
  BATCH_SIZE = 64     #每個(gè)batch大小，稍微大一點(diǎn)的batch會(huì)更穩(wěn)定
  EARLY_STOP_PATIENCE = 10 #控制early stopping的參數(shù)


  def input_data(test=False):
    file_name = TEST_FILE if test else TRAIN_FILE
    df = pd.read_csv(file_name)
    cols = df.columns[:-1]

    #dropna()是丟棄有缺失數(shù)據(jù)的樣本，這樣最后7000多個(gè)樣本只剩2140個(gè)可用的。
    df = df.dropna()  
    df['Image'] = df['Image'].apply(lambda img: np.fromstring(img, sep=' ') / 255.0)

    X = np.vstack(df['Image'])
    X = X.reshape((-1,96,96,1))

    if test:
      y = None
    else:
      y = df[cols].values / 96.0    #將y值縮放到[0,1]區(qū)間

    return X, y

  #最后生成提交結(jié)果的時(shí)候要用到
  keypoint_index = {
    'left_eye_center_x':0,
    'left_eye_center_y':1,
    'right_eye_center_x':2,
    'right_eye_center_y':3,
    'left_eye_inner_corner_x':4,
    'left_eye_inner_corner_y':5,
    'left_eye_outer_corner_x':6,
    'left_eye_outer_corner_y':7,
    'right_eye_inner_corner_x':8,
    'right_eye_inner_corner_y':9,
    'right_eye_outer_corner_x':10,
    'right_eye_outer_corner_y':11,
    'left_eyebrow_inner_end_x':12,
    'left_eyebrow_inner_end_y':13,
    'left_eyebrow_outer_end_x':14,
    'left_eyebrow_outer_end_y':15,
    'right_eyebrow_inner_end_x':16,
    'right_eyebrow_inner_end_y':17,
    'right_eyebrow_outer_end_x':18,
    'right_eyebrow_outer_end_y':19,
    'nose_tip_x':20,
    'nose_tip_y':21,
    'mouth_left_corner_x':22,
    'mouth_left_corner_y':23,
    'mouth_right_corner_x':24,
    'mouth_right_corner_y':25,
    'mouth_center_top_lip_x':26,
    'mouth_center_top_lip_y':27,
    'mouth_center_bottom_lip_x':28,
    'mouth_center_bottom_lip_y':29
  }

  def save_model(saver,sess,save_path):
    path = saver.save(sess, save_path)
    print 'model save in :{0}'.format(path)

  if __name__ == '__main__':
    sess = tf.InteractiveSession()
    y_conv, rmse = model()
    train_step = tf.train.AdamOptimizer(1e-3).minimize(rmse)

    #變量都要初始化 
    sess.run(tf.initialize_all_variables())
    X,y = input_data()
    X_valid, y_valid = X[:VALIDATION_SIZE], y[:VALIDATION_SIZE]
    X_train, y_train = X[VALIDATION_SIZE:], y[VALIDATION_SIZE:]

    best_validation_loss = 1000000.0
    current_epoch = 0
    TRAIN_SIZE = X_train.shape[0]
    train_index = range(TRAIN_SIZE)
    random.shuffle(train_index)
    X_train, y_train = X_train[train_index], y_train[train_index]

    saver = tf.train.Saver()

    print 'begin training..., train dataset size:{0}'.format(TRAIN_SIZE)
    for i in xrange(EPOCHS):
      random.shuffle(train_index) #每個(gè)epoch都shuffle一下效果更好
      X_train, y_train = X_train[train_index], y_train[train_index]

      for j in xrange(0,TRAIN_SIZE,BATCH_SIZE):
        print 'epoch {0}, train {1} samples done...'.format(i,j)

        train_step.run(feed_dict={x:X_train[j:j+BATCH_SIZE], 
          y_:y_train[j:j+BATCH_SIZE], keep_prob:0.5})

      #電腦太渣，用所有訓(xùn)練樣本計(jì)算train_loss居然死機(jī)，只好注釋了。
      #train_loss = rmse.eval(feed_dict={x:X_train, y_:y_train, keep_prob: 1.0})
      validation_loss = rmse.eval(feed_dict={x:X_valid, y_:y_valid, keep_prob: 1.0})

      print 'epoch {0} done! validation loss:{1}'.format(i, validation_loss*96.0)
      if validation_loss < best_validation_loss:
        best_validation_loss = validation_loss
        current_epoch = i
        save_model(saver,sess,SAVE_PATH)  #即時(shí)保存最好的結(jié)果
      elif (i - current_epoch) >= EARLY_STOP_PATIENCE:
        print 'early stopping'
        break

  X,y = input_data(test=True)
  y_pred = []

  TEST_SIZE = X.shape[0]
  for j in xrange(0,TEST_SIZE,BATCH_SIZE):
    y_batch = y_conv.eval(feed_dict={x:X[j:j+BATCH_SIZE], keep_prob:1.0})
    y_pred.extend(y_batch)

  print 'predict test image done!'

  output_file = open('submit.csv','w')
  output_file.write('RowId,Location\n')

  IdLookupTable = open('IdLookupTable.csv')
  IdLookupTable.readline()

  for line in IdLookupTable:
    RowId,ImageId,FeatureName = line.rstrip().split(',')
    image_index = int(ImageId) - 1
    feature_index = keypoint_index[FeatureName]
    feature_location = y_pred[image_index][feature_index] * 96
    output_file.write('{0},{1}\n'.format(RowId,feature_location))

  output_file.close()
  IdLookupTable.close()