释放双眼，带上耳机，听听看~！
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
2
3
4
5
6
2x = tf.placeholder(tf.float32, [None, 256, 256, 1],name='x')
3y_ = tf.placeholder(tf.float32, [None],name='y_')
4keep_prob = tf.placeholder(tf.float32)
5
6
-
CNN模型结构定义
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
2 initial = tf.truncated_normal(shape, stddev=0.1)
3 return tf.Variable(initial,name=name)
4
5
6def bias_variable(shape,name):
7 initial = tf.constant(0.1, shape=shape)
8 return tf.Variable(initial,name=name)
9
10with tf.name_scope('conv2d'):
11 def conv2d(x, W):
12 # stride [1, x_movement, y_movement, 1]
13 # Must have strides[0] = strides[3] = 1
14 return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
15
16with tf.name_scope('max_pool_2x2'):
17 def max_pool_2x2(x):
18 # stride [1, x_movement, y_movement, 1]
19 return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
20 def max_pool_4x4(x):
21 # stride [1, x_movement, y_movement, 1]
22 return tf.nn.max_pool(x, ksize=[1,4,4,1], strides=[1,4,4,1], padding='SAME')
23
24
25def define_predict_y(x):
26 with tf.variable_scope("conv1"):
27 ## conv1 layer ##
28 W_conv1 = weight_variable([3,3, 1,64],'W_conv1') # patch 3x3, in size 1, out size 64
29 b_conv1 = bias_variable([64],'b_conv1')
30 h_conv1 = tf.nn.elu(conv2d(x, W_conv1) + b_conv1) # output size 28x28x32
31 h_pool1 = tf.nn.max_pool(h_conv1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME') # output size 14x14x32
32 with tf.variable_scope("conv2"):
33 ## conv2 layer ##
34 W_conv2 = weight_variable([3,3, 64, 128],'W_conv2') # patch 5x5, in size 32, out size 64
35 b_conv2 = bias_variable([128],'b_conv2')
36 h_conv2 = tf.nn.elu(conv2d(h_pool1, W_conv2) + b_conv2) # output size 14x14x64
37 h_pool2 = max_pool_4x4(h_conv2)
38 with tf.variable_scope("conv3"):
39 ## conv3 layer ##
40 W_conv3 = weight_variable([3,3, 128, 256],'W_conv3') # patch 5x5, in size 32, out size 64
41 b_conv3 = bias_variable([256],'b_conv3')
42 h_conv3 = tf.nn.elu(conv2d(h_pool2, W_conv3) + b_conv3) # output size 14x14x64
43 h_pool3 = max_pool_4x4(h_conv3)
44 with tf.variable_scope("conv4"):
45 ## conv4 layer ##
46 W_conv4 = weight_variable([3,3, 256, 512],'W_conv4') # patch 5x5, in size 32, out size 64
47 b_conv4 = bias_variable([512],'b_conv4')
48 h_conv4 = tf.nn.elu(conv2d(h_pool3, W_conv4) + b_conv4) # output size 14x14x64
49 h_pool4 = max_pool_4x4(h_conv4)
50
51 with tf.variable_scope("fc1"):
52 ## fc1 layer ##
53 W_fc1 = weight_variable([2*2*512, 128],'W_fc1')
54 b_fc1 = bias_variable([128],'b_fc1')
55 # [n_samples, 7, 7, 64] ->> [n_samples, 7*7*64]
56 h_pool4_flat = tf.reshape(h_pool4, [-1, 2*2*512])
57 h_fc1 = tf.nn.elu(tf.matmul(h_pool4_flat, W_fc1) + b_fc1)
58 h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
59
60 # ## fc2 layer ##
61 # with tf.variable_scope("fc2"):
62 # W_fc2 = weight_variable([128, 10],'W_fc2')
63 # b_fc2 = bias_variable([10],'b_fc2')
64 # predict_y = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
65
66 return h_fc1_drop
67
68prediction = define_predict_y(x)
69# 用于保存和载入模型
70new_saver=tf.train.Saver()
71
72
-
载入已经保存的模型参数
2
3
2 print("导入参数成功！")
3
-
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
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
2 features = {
3 'encoded': tf.FixedLenFeature((), tf.string),
4 'fname': tf.FixedLenFeature((), tf.string),
5 'width': tf.FixedLenFeature((), tf.int64),
6 'height': tf.FixedLenFeature((), tf.int64),
7 'label': tf.FixedLenFeature((), tf.int64),}
8 parsed_features = tf.parse_single_example(example_proto, features=features)
9 return parsed_features
10
11
12
13###1.....
14img_vec_list = [] #所有图片的向量，按顺序存的
15
16
17
18def read_test(input_file):
19
20 # 用 dataset 读取 tfrecord 文件
21 dataset = tf.data.TFRecordDataset(input_file)
22 dataset = dataset.map(_parse_record)#解析tfrecord文件中的所有记录，使用dataset的map方法
23 #dataset = dataset.repeat(epochs).shuffle(buffer_size).batch(batch_size)
24 iterator = dataset.make_one_shot_iterator()
25
26 with tf.Session() as sess:
27 try:
28 i =0
29 while iterator.get_next():
30 i = i+1
31 print(i)
32 features = sess.run(iterator.get_next())
33 img_fname = features['fname']
34 img_fname = img_fname.decode()
35 img = tf.decode_raw(features['encoded'], tf.uint8)
36 img = tf.reshape(img, [256, 256, 1])
37 img = tf.cast(img, tf.float32) / 255.0 #将矩阵归一化0-1之间
38 label = tf.cast(features['label'], tf.int32)
39
40 one = [sess.run(img),img_fname,sess.run(label)]
41 print(one[1])
42 img_vec_list.append(one)
43 except tf.errors.OutOfRangeError:
44 print("..")
45 print("-------------",len(img_vec_list))
46 img_vec_list.sort(key = lambda x:x[1])
47 print("over..")
48read_test('F:/data/test0.tfrecords')
49read_test('F:/data/train0.tfrecords')
50read_test('F:/data/test1.tfrecords')
51read_test('F:/data/train1.tfrecords')
52read_test('F:/data/test2.tfrecords')
53read_test('F:/data/train2.tfrecords')
54read_test('F:/data/test3.tfrecords')
55read_test('F:/data/train3.tfrecords')
56read_test('F:/data/test4.tfrecords')
57read_test('F:/data/train4.tfrecords')
58read_test('F:/data/test5.tfrecords')
59read_test('F:/data/train5.tfrecords')
60read_test('F:/data/test6.tfrecords')
61read_test('F:/data/train6.tfrecords')
62read_test('F:/data/test7.tfrecords')
63read_test('F:/data/train7.tfrecords')
64read_test('F:/data/test8.tfrecords')
65read_test('F:/data/train8.tfrecords')
66read_test('F:/data/test9.tfrecords')
67read_test('F:/data/train9.tfrecords')
68
2.并送入已训练好的CNN中
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
2
3def get_vector():
4 with tf.Session() as sess:
5 print("there..")
6 # 如果是训练，初始化参数
7 sess.run(tf.global_variables_initializer())
8 print("222")
9 # 创建一个协调器，管理线程
10 coord = tf.train.Coordinator()
11 print("333")
12 # 启动QueueRunner,此时文件名队列已经进队
13 threads = tf.train.start_queue_runners(sess=sess, coord=coord)
14 print("444")
15
16 new_saver.restore(sess, tf.train.latest_checkpoint('C:/Users/Administrator/Desktop/ckpt/'))
17 print("导入参数成功！")
18
19 for i in range(len(img_vec_list)):
20 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})
21 vector_list.append(vector)
22 #print("vector is :",len(vector[0]))
23
24get_vector()
25
-
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
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
2 """
3 计算两个向量之间的余弦相似度
4 :param vector_a: 向量 a
5 :param vector_b: 向量 b
6 :return: sim
7 """
8 vector_a = np.mat(vector_a)
9 vector_b = np.mat(vector_b)
10 num = float(vector_a * vector_b.T)
11 denom = np.linalg.norm(vector_a) * np.linalg.norm(vector_b)
12 cos = num / denom
13 sim = 0.5 + 0.5 * cos
14 return sim
15
16##########3....
17cos_list = []
18
19def get_all_vec_cos():
20 for i in range(len(img_vec_list)):
21 max_cos = 0
22 max_index = i
23 for j in range(len(img_vec_list)):
24 if int(i/11) == int(j/11):
25 continue
26 else:
27 temp_cos = cos_sim(vector_list[i],vector_list[j])
28
29 if temp_cos>max_cos:
30 print("temp_cos:",temp_cos,"max_cos",max_cos)
31 max_cos = temp_cos
32 max_index = int(j/11)
33 cos_list.append([int(i/11),max_index,max_cos])
34 print("cos:",i," ",cos_list[i])
35 print("cos_list:",len(cos_list))
36
37get_all_vec_cos()
38
39
-
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
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
2
3#返回的是vidoe序号
4def get_most_video():
5 #将cos_list分割,每份11个
6 #cos_list = [cos_list[i:i+11] for i in range(0,len(cos_list),11)]
7 print("cos_list:",cos_list)
8 split_cos_list = []
9 for j in range(0,len(cos_list),11):
10 split_cos_list.append(cos_list[j:j+11])
11 print("split_cos_list:",split_cos_list)
12 for i in range(len(split_cos_list)):
13 index = []
14 for item in split_cos_list[i]:
15 index.append(item[1])
16 most_index = Counter(index).most_common(3)
17 most_video.append(most_index)
18 #print("most_video:",len(most_video))
19
20get_most_video()
21#print(most_video)
22
23
