该模块包含几部分:
- 调用训练好的并且已经保存的CNN模型(仅四层卷积层部分)
- 逐个读取tfrecords文件中的元素,并送入已训练好的CNN中,给每个图片提取128个特征
- 每首歌包含11个图片,即11*128个特征,将每首歌的11*128个特征之间进行余弦相似度计算
- 逐个歌曲计算,返回每个歌曲的最相似的三首歌歌名,以列表的形式
-
调用训练好的并且已经保存的CNN模型(仅四层卷积层部分)
- 定义CNN模型的参数
lr = tf.Variable(0.001, dtype=tf.float32)
x = tf.placeholder(tf.float32, [None, 256, 256, 1],name='x')
y_ = tf.placeholder(tf.float32, [None],name='y_')
keep_prob = tf.placeholder(tf.float32)
-
CNN模型结构定义
def weight_variable(shape,name):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial,name=name)
def bias_variable(shape,name):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial,name=name)
with tf.name_scope('conv2d'):
def conv2d(x, W):
# stride [1, x_movement, y_movement, 1]
# Must have strides[0] = strides[3] = 1
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
with tf.name_scope('max_pool_2x2'):
def max_pool_2x2(x):
# stride [1, x_movement, y_movement, 1]
return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
def max_pool_4x4(x):
# stride [1, x_movement, y_movement, 1]
return tf.nn.max_pool(x, ksize=[1,4,4,1], strides=[1,4,4,1], padding='SAME')
def define_predict_y(x):
with tf.variable_scope("conv1"):
## conv1 layer ##
W_conv1 = weight_variable([3,3, 1,64],'W_conv1') # patch 3x3, in size 1, out size 64
b_conv1 = bias_variable([64],'b_conv1')
h_conv1 = tf.nn.elu(conv2d(x, W_conv1) + b_conv1) # output size 28x28x32
h_pool1 = tf.nn.max_pool(h_conv1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME') # output size 14x14x32
with tf.variable_scope("conv2"):
## conv2 layer ##
W_conv2 = weight_variable([3,3, 64, 128],'W_conv2') # patch 5x5, in size 32, out size 64
b_conv2 = bias_variable([128],'b_conv2')
h_conv2 = tf.nn.elu(conv2d(h_pool1, W_conv2) + b_conv2) # output size 14x14x64
h_pool2 = max_pool_4x4(h_conv2)
with tf.variable_scope("conv3"):
## conv3 layer ##
W_conv3 = weight_variable([3,3, 128, 256],'W_conv3') # patch 5x5, in size 32, out size 64
b_conv3 = bias_variable([256],'b_conv3')
h_conv3 = tf.nn.elu(conv2d(h_pool2, W_conv3) + b_conv3) # output size 14x14x64
h_pool3 = max_pool_4x4(h_conv3)
with tf.variable_scope("conv4"):
## conv4 layer ##
W_conv4 = weight_variable([3,3, 256, 512],'W_conv4') # patch 5x5, in size 32, out size 64
b_conv4 = bias_variable([512],'b_conv4')
h_conv4 = tf.nn.elu(conv2d(h_pool3, W_conv4) + b_conv4) # output size 14x14x64
h_pool4 = max_pool_4x4(h_conv4)
with tf.variable_scope("fc1"):
## fc1 layer ##
W_fc1 = weight_variable([2*2*512, 128],'W_fc1')
b_fc1 = bias_variable([128],'b_fc1')
# [n_samples, 7, 7, 64] ->> [n_samples, 7*7*64]
h_pool4_flat = tf.reshape(h_pool4, [-1, 2*2*512])
h_fc1 = tf.nn.elu(tf.matmul(h_pool4_flat, W_fc1) + b_fc1)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# ## fc2 layer ##
# with tf.variable_scope("fc2"):
# W_fc2 = weight_variable([128, 10],'W_fc2')
# b_fc2 = bias_variable([10],'b_fc2')
# predict_y = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
return h_fc1_drop
prediction = define_predict_y(x)
# 用于保存和载入模型
new_saver=tf.train.Saver()
- 载入已经保存的模型参数
new_saver.restore(sess, tf.train.latest_checkpoint('C:/Users/Administrator/Desktop/ckpt/'))
print("导入参数成功!")
-
逐个读取tfrecords文件中的元素,并送入已训练好的CNN中,给每个图片提取128个特征
1.逐个读取tfrecords文件中的元素
def _parse_record(example_proto):
features = {
'encoded': tf.FixedLenFeature((), tf.string),
'fname': tf.FixedLenFeature((), tf.string),
'width': tf.FixedLenFeature((), tf.int64),
'height': tf.FixedLenFeature((), tf.int64),
'label': tf.FixedLenFeature((), tf.int64),}
parsed_features = tf.parse_single_example(example_proto, features=features)
return parsed_features
###1.....
img_vec_list = [] #所有图片的向量,按顺序存的
def read_test(input_file):
# 用 dataset 读取 tfrecord 文件
dataset = tf.data.TFRecordDataset(input_file)
dataset = dataset.map(_parse_record)#解析tfrecord文件中的所有记录,使用dataset的map方法
#dataset = dataset.repeat(epochs).shuffle(buffer_size).batch(batch_size)
iterator = dataset.make_one_shot_iterator()
with tf.Session() as sess:
try:
i =0
while iterator.get_next():
i = i+1
print(i)
features = sess.run(iterator.get_next())
img_fname = features['fname']
img_fname = img_fname.decode()
img = tf.decode_raw(features['encoded'], tf.uint8)
img = tf.reshape(img, [256, 256, 1])
img = tf.cast(img, tf.float32) / 255.0 #将矩阵归一化0-1之间
label = tf.cast(features['label'], tf.int32)
one = [sess.run(img),img_fname,sess.run(label)]
print(one[1])
img_vec_list.append(one)
except tf.errors.OutOfRangeError:
print("..")
print("-------------",len(img_vec_list))
img_vec_list.sort(key = lambda x:x[1])
print("over..")
read_test('F:/data/test0.tfrecords')
read_test('F:/data/train0.tfrecords')
read_test('F:/data/test1.tfrecords')
read_test('F:/data/train1.tfrecords')
read_test('F:/data/test2.tfrecords')
read_test('F:/data/train2.tfrecords')
read_test('F:/data/test3.tfrecords')
read_test('F:/data/train3.tfrecords')
read_test('F:/data/test4.tfrecords')
read_test('F:/data/train4.tfrecords')
read_test('F:/data/test5.tfrecords')
read_test('F:/data/train5.tfrecords')
read_test('F:/data/test6.tfrecords')
read_test('F:/data/train6.tfrecords')
read_test('F:/data/test7.tfrecords')
read_test('F:/data/train7.tfrecords')
read_test('F:/data/test8.tfrecords')
read_test('F:/data/train8.tfrecords')
read_test('F:/data/test9.tfrecords')
read_test('F:/data/train9.tfrecords')
2.并送入已训练好的CNN中
vector_list = []
def get_vector():
with tf.Session() as sess:
print("there..")
# 如果是训练,初始化参数
sess.run(tf.global_variables_initializer())
print("222")
# 创建一个协调器,管理线程
coord = tf.train.Coordinator()
print("333")
# 启动QueueRunner,此时文件名队列已经进队
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
print("444")
new_saver.restore(sess, tf.train.latest_checkpoint('C:/Users/Administrator/Desktop/ckpt/'))
print("导入参数成功!")
for i in range(len(img_vec_list)):
vector = sess.run(prediction,feed_dict={x:np.expand_dims(img_vec_list[i][0],0),y_:np.expand_dims(img_vec_list[i][2],0),keep_prob:0.5})
vector_list.append(vector)
#print("vector is :",len(vector[0]))
get_vector()
-
每首歌包含11个图片,即11*128个特征,将每首歌的11*128个特征之间进行余弦相似度计算
def cos_sim(vector_a, vector_b):
"""
计算两个向量之间的余弦相似度
:param vector_a: 向量 a
:param vector_b: 向量 b
:return: sim
"""
vector_a = np.mat(vector_a)
vector_b = np.mat(vector_b)
num = float(vector_a * vector_b.T)
denom = np.linalg.norm(vector_a) * np.linalg.norm(vector_b)
cos = num / denom
sim = 0.5 + 0.5 * cos
return sim
##########3....
cos_list = []
def get_all_vec_cos():
for i in range(len(img_vec_list)):
max_cos = 0
max_index = i
for j in range(len(img_vec_list)):
if int(i/11) == int(j/11):
continue
else:
temp_cos = cos_sim(vector_list[i],vector_list[j])
if temp_cos>max_cos:
print("temp_cos:",temp_cos,"max_cos",max_cos)
max_cos = temp_cos
max_index = int(j/11)
cos_list.append([int(i/11),max_index,max_cos])
print("cos:",i," ",cos_list[i])
print("cos_list:",len(cos_list))
get_all_vec_cos()
-
逐个歌曲计算,返回每个歌曲的最相似的三首歌歌名,以列表的形式
most_video = []
#返回的是vidoe序号
def get_most_video():
#将cos_list分割,每份11个
#cos_list = [cos_list[i:i+11] for i in range(0,len(cos_list),11)]
print("cos_list:",cos_list)
split_cos_list = []
for j in range(0,len(cos_list),11):
split_cos_list.append(cos_list[j:j+11])
print("split_cos_list:",split_cos_list)
for i in range(len(split_cos_list)):
index = []
for item in split_cos_list[i]:
index.append(item[1])
most_index = Counter(index).most_common(3)
most_video.append(most_index)
#print("most_video:",len(most_video))
get_most_video()
#print(most_video)