将libFM模型变换成tensorflow可serving的形式

import tensorflow as tf

def load_fm_model(file_name):
    state = ‘‘
    fid = 0
    max_fid = 0
    w0 = 0.0
    wj = {}
    v = {}
    k = 0
    with open(file_name) as f:
        for line in f:
            line = line.rstrip()
            if ‘global bias W0‘ in line:
                state = ‘w0‘
                fid = 0
                continue
            elif ‘unary interactions Wj‘ in line:
                state = ‘wj‘
                fid = 0
                continue
            elif ‘pairwise interactions Vj,f‘ in line:
                state = ‘v‘
                fid = 0
                continue

            if state == ‘w0‘:
                fv = float(line)
                w0 = fv
            elif state == ‘wj‘:
                fv = float(line)
                if fv != 0:
                    wj[fid] = fv
                fid += 1
                max_fid = max(max_fid, fid)
            elif state == ‘v‘:
                fv = [float(_v) for _v in line.split(‘ ‘)]
                k = len(fv)
                if any([_v!=0 for _v in fv]):
                    v[fid] = fv
                fid += 1
                max_fid = max(max_fid, fid)
    return w0, wj, v, k, max_fid

_w0, _wj, _v, _k, _max_fid = load_fm_model(‘fm_model‘)

n=_max_fid
print ‘n‘, n
k=_k
print ‘k‘, k

#write fm algorithm
w0=tf.Variable(_w0)
w1=tf.Variable(tf.truncated_normal([n]))
print ‘w1‘, w1
w1_st = tf.SparseTensor(indices=[[a] for a in _wj.keys()], values=_wj.values(), dense_shape=[n])
tf.assign(w1, tf.sparse_tensor_to_dense(w1_st))
print ‘w1‘, w1
w2=tf.Variable(tf.truncated_normal([n,k]))
print ‘w2‘, w2
inds = []
vals = []
for fid, fv in _v.items():
    for i, v in enumerate(fv):
        if v != 0:
            inds.append([fid, i])
            vals.append(v)
w2_st = tf.SparseTensor(indices=inds, values=vals, dense_shape=[n,k])
tf.assign(w2, tf.sparse_tensor_to_dense(w2_st))
print ‘w2‘, w2

x_=tf.placeholder(tf.float32,[None,n])
#y_=tf.placeholder(tf.float32,[None])
batch=tf.placeholder(tf.int32)

w2_new=tf.reshape(tf.tile(w2,[batch,1]),[-1,n,k])
print ‘w2_new‘, w2_new

board_x=tf.reshape(tf.tile(x_,[1,k]),[-1,n,k])
print ‘board_x‘, board_x
board_x2=tf.square(board_x)

#print tf.multiply(w2_new,board_x)
#print tf.reduce_sum(tf.multiply(w2_new,board_x),axis=1)
q=tf.square(tf.reduce_sum(tf.multiply(w2_new,board_x),axis=1))
h=tf.reduce_sum(tf.multiply(tf.square(w2_new),board_x2),axis=1)

y_fm=w0+tf.reduce_sum(tf.multiply(x_,w1),axis=1)+    1/2*tf.reduce_sum(q-h,axis=1)

saver = tf.train.Saver()
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    #a = sess.run(y_fm, feed_dict={x_:x_train,y_:y_train,batch:70})
    #print a
    save_path = "./model.ckpt"
    #saver.save(sess, save_path)
    tf.saved_model.simple_save(sess, save_path, inputs={"x": x_, "batch":batch}, outputs={"y_fm": y_fm})