将libFM模型变换成tensorflow可serving的形式

fm_model是libFM生成的模型

model.ckpt是可以tensorflow serving的模型结构

亲测输出正确。

代码：

 1 import tensorflow as tf
 2 
 3 # libFM model
 4 def load_fm_model(file_name):
 5     state = ''
 6     fid = 0
 7     max_fid = 0
 8     w0 = 0.0
 9     wj = {}
10     v = {}
11     k = 0
12     with open(file_name) as f:
13         for line in f:
14             line = line.rstrip()
15             if 'global bias W0' in line:
16                 state = 'w0'
17                 fid = 0
18                 continue
19             elif 'unary interactions Wj' in line:
20                 state = 'wj'
21                 fid = 0
22                 continue
23             elif 'pairwise interactions Vj,f' in line:
24                 state = 'v'
25                 fid = 0
26                 continue
27 
28             if state == 'w0':
29                 fv = float(line)
30                 w0 = fv
31             elif state == 'wj':
32                 fv = float(line)
33                 if fv != 0:
34                     wj[fid] = fv
35                 fid += 1
36                 max_fid = max(max_fid, fid)
37             elif state == 'v':
38                 fv = [float(_v) for _v in line.split(' ')]
39                 k = len(fv)
40                 if any([_v!=0 for _v in fv]):
41                     v[fid] = fv
42                 fid += 1
43                 max_fid = max(max_fid, fid)
44     return w0, wj, v, k, max_fid
45 
46 _w0, _wj, _v, _k, _max_fid = load_fm_model('libfm_model_file')
47 
48 # max feature_id
49 n = _max_fid
50 print 'n', n
51 
52 # vector dimension
53 k = _k
54 print 'k', k
55 
56 # write fm algorithm
57 w0 = tf.constant(_w0)
58 w1c = tf.constant([_wj.get(fid, 0) for fid in xrange(n)], shape=[n])
59 w1 = tf.Variable(w1c)
60 #print 'w1', w1
61 
62 vec = []
63 for fid in xrange(n):
64     vec.append(_v.get(fid, [0]*k))
65 w2c = tf.constant(vec, shape=[n,k])
66 w2 = tf.Variable(w2c)
67 print 'w2', w2
68 
69 # inputs
70 x = tf.placeholder(tf.string, [None])
71 batch = tf.shape(x)[0]
72 x_s = tf.string_split(x)
73 inds = tf.stack([tf.cast(x_s.indices[:,0], tf.int64), tf.string_to_number(x_s.values, tf.int64)], axis=1)
74 x_sparse = tf.sparse.SparseTensor(indices=inds, values=tf.ones([tf.shape(inds)[0]]), dense_shape=[batch,n])
75 x_ = tf.sparse.to_dense(x_sparse)
76 
77 w2_rep = tf.reshape(tf.tile(w2, [batch,1]), [-1,n,k])
78 print 'w2_rep', w2_rep
79 
80 x_rep = tf.reshape(tf.tile(tf.reshape(x_, [batch*n, 1]), [1,k]), [-1,n,k])
81 print 'x_rep', x_rep
82 x_rep2 = tf.square(x_rep)
83 
84 #print tf.multiply(w2_rep,x_rep)
85 #print tf.reduce_sum(tf.multiply(w2_rep,x_rep), axis=1)
86 q = tf.square(tf.reduce_sum(tf.multiply(w2_rep, x_rep), axis=1))
87 h = tf.reduce_sum(tf.multiply(tf.square(w2_rep), x_rep2), axis=1)
88 
89 y = w0 + tf.reduce_sum(tf.multiply(x_, w1), axis=1) +\
90     1.0/2 * tf.reduce_sum(q-h, axis=1)
91 
92 saver = tf.train.Saver()
93 with tf.Session() as sess:
94     sess.run(tf.global_variables_initializer())
95     #a = sess.run(y, feed_dict={x_:x_train,y_:y_train,batch:70})
96     #print a
97     save_path = "./model.ckpt"
98     tf.saved_model.simple_save(sess, save_path, inputs={"x": x}, outputs={"y": y})

参考：

https://blog.csdn.net/u010159842/article/details/78789355 （开头借鉴此文，但其有不少细节错误）

https://www.tensorflow.org/guide/saved_model

http://nowave.it/factorization-machines-with-tensorflow.html

posted on 2019-03-04 20:00 冰山上的博客阅读(521) 评论(0) 编辑收藏举报