I'm using Tensorflow to process a simple matrix factorization algorithm. Every step went correct but at the last step, where I want to eval() a Tensor to store it, the program didn't work and only occupied more and more memory. So is there something wrong in my code? I'm a beginner in Tensorflow and I'm don't know where the problem is. Below is my code.
class model(object):
def __init__(self, D, Q, stepsize = 6e-7, max_iter = 200, inner_maxiter = 50, dim = 200, verbose = 5):
self.D = tf.constant(D, dtype = tf.float32)
self.Q = tf.constant(Q, dtype = tf.float32)
self.rank = dim
self.stepsize = stepsize
self.max_iter = max_iter
self.inner_maxiter = inner_maxiter
self.verbose = verbose
self.W = tf.Variable((np.random.rand(self.rank, sample_num)), dtype = tf.float32, name = 'W')
self.C = tf.Variable((np.random.rand(context_num, self.rank)), dtype = tf.float32, name = 'C')
def _run(self, sess):
Q = self.Q
D = self.D
W = self.W
C = self.C
for i in xrange(self.max_iter):
if (i + 1) % 2 == 1:
for j in xrange(self.inner_maxiter):
ED = tf.transpose(Q) * (1.0 / (1.0 + tf.exp(- tf.matmul(C, W))))
recons = D - ED
W_grad = tf.matmul(tf.transpose(C), recons)
W = W + self.stepsize * W_grad
else:
for j in xrange(self.inner_maxiter):
ED = tf.transpose(Q) * (1.0 / (1.0 + tf.exp(- tf.matmul(C, W))))
recons = D - ED
C_grad = tf.matmul(recons, tf.transpose(W))
C = C + self.stepsize * C_grad
print 'epoch: %d' % i
print W.eval()
print C.eval()
train_epoch = model(D, Q, args.step_size, \
args.max_iter, args.inner_maxiter, args.dim, args.verbose)
with tf.Session(config = config) as sess:
tf.initialize_all_variables().run()
train_epoch._run(sess)
The program stopped and occupying memory at the last two lines in _run(), which contains W.eval() and C.eval(). So what should I do to fix it? Can someone help?
Solved. You cannot add iteration of symbolic operations in Tensorflow. Instead, you should build the data flow first, which means you should define your operations in the initialize step.