在 TensorFlow 中计算图上的梯度不会产生类型提取错误
Posted
技术标签:
【中文标题】在 TensorFlow 中计算图上的梯度不会产生类型提取错误【英文标题】：Computing gradients on a graph in TensorFlow gives none type fetching error 【发布时间】：2018-02-06 09:40:45 【问题描述】：我正在尝试在下图中计算梯度（这是一个类方法）：
def __define_likelihood_computation(self):
self.__lik_graph = tf.Graph()
lik_graph = self.__lik_graph
r = self.__C(self.__th).shape[1]
m = self.__H(self.__th).shape[0]
n = self.__F(self.__th).shape[1]
p = self.__G(self.__th).shape[1]
x0_mean = self.__x0_mean
x0_cov = self.__x0_cov
with lik_graph.as_default():
# FIXME: Don't Repeat Yourself (in simulation and here)
th = tf.placeholder(tf.float64, shape=[None], name='th')
u = tf.placeholder(tf.float64, shape=[r, None], name='u')
t = tf.placeholder(tf.float64, shape=[None], name='t')
y = tf.placeholder(tf.float64, shape=[m, None], name='y')
N = tf.stack([tf.shape(t)[0]])
N = tf.reshape(N, ())
F = tf.py_func(self.__F, [th], tf.float64, name='F')
F.set_shape([n, n])
C = tf.py_func(self.__C, [th], tf.float64, name='C')
C.set_shape([n, r])
G = tf.py_func(self.__G, [th], tf.float64, name='G')
G.set_shape([n, p])
H = tf.py_func(self.__H, [th], tf.float64, name='H')
H.set_shape([m, n])
x0_mean = tf.py_func(x0_mean, [th], tf.float64, name='x0_mean')
x0_mean.set_shape([n, 1])
P_0 = tf.py_func(x0_cov, [th], tf.float64, name='x0_cov')
P_0.set_shape([n, n])
Q = tf.py_func(self.__w_cov, [th], tf.float64, name='w_cov')
Q.set_shape([p, p])
R = tf.py_func(self.__v_cov, [th], tf.float64, name='v_cov')
R.set_shape([m, m])
I = tf.eye(n, n, dtype=tf.float64)
def lik_loop_cond(k, P, S, t, u, x, y):
return tf.less(k, N-1)
def lik_loop_body(k, P, S, t, u, x, y):
# TODO: this should be function of time
u_t_k = tf.slice(u, [0, k], [r, 1])
# k+1, cause zeroth measurement should not be taken into account
y_k = tf.slice(y, [0, k+1], [m, 1])
t_k = tf.slice(t, [k], [2], 't_k')
# TODO: extract Kalman filter to a separate class
def state_predict(x, t):
Fx = tf.matmul(F, x, name='Fx')
Cu = tf.matmul(C, u_t_k, name='Cu')
dx = Fx + Cu
return dx
def covariance_predict(P, t):
GQtG = tf.matmul(G @ Q, G, transpose_b=True)
PtF = tf.matmul(P, F, transpose_b=True)
dP = tf.matmul(F, P) + PtF + GQtG
return dP
x = tf.contrib.integrate.odeint(state_predict, x, t_k,
name='state_predict')
x = x[-1]
P = tf.contrib.integrate.odeint(covariance_predict, P, t_k,
name='covariance_predict')
P = P[-1]
E = y_k - tf.matmul(H, x)
B = tf.matmul(H @ P, H, transpose_b=True) + R
invB = tf.matrix_inverse(B)
K = tf.matmul(P, H, transpose_b=True) @ invB
S_k = tf.matmul(E, invB @ E, transpose_a=True)
S_k = 0.5 * (S_k + tf.log(tf.matrix_determinant(B)))
S = S + S_k
# state update
x = x + tf.matmul(K, E)
# covariance update
P = (I - K @ H) @ P
k = k + 1
return k, P, S, t, u, x, y
k = tf.constant(0, name='k')
P = P_0
S = tf.constant(0.0, dtype=tf.float64, shape=[1, 1], name='S')
x = x0_mean
# TODO: make a named tuple of named list
lik_loop = tf.while_loop(lik_loop_cond, lik_loop_body,
[k, P, S, t, u, x, y], name='lik_loop')
dS = tf.gradients(lik_loop[2], th)
self.__lik_loop_op = lik_loop
self.__dS = dS
评估本身如下：
def dL(self, t, u, y, th=None):
if th is None:
th = self.__th
self.__validate(th)
g = self.__lik_graph
if t.shape[0] != u.shape[1]:
raise Exception('''t.shape[0] != u.shape[1]''')
# run lik graph
with tf.Session(graph=g) as sess:
t_ph = g.get_tensor_by_name('t:0')
th_ph = g.get_tensor_by_name('th:0')
u_ph = g.get_tensor_by_name('u:0')
y_ph = g.get_tensor_by_name('y:0')
rez = sess.run(self.__dS, th_ph: th, t_ph: t, u_ph: u, y_ph: y)
return rez
似然计算确实有效，如下：
def lik(self, t, u, y, th=None):
if th is None:
th = self.__th
self.__validate(th)
g = self.__lik_graph
if t.shape[0] != u.shape[1]:
raise Exception('''t.shape[0] != u.shape[1]''')
# run lik graph
with tf.Session(graph=g) as sess:
t_ph = g.get_tensor_by_name('t:0')
th_ph = g.get_tensor_by_name('th:0')
u_ph = g.get_tensor_by_name('u:0')
y_ph = g.get_tensor_by_name('y:0')
rez = sess.run(self.__lik_loop_op, th_ph: th, t_ph: t, u_ph: u,
y_ph: y)
N = len(t)
m = y.shape[0]
S = rez[2]
S = S + N*m * 0.5 + np.log(2*math.pi)
return S
当我尝试计算梯度时（调用dL），我收到以下回溯错误：
可能是什么原因？
抱歉发了这么久。
【问题讨论】：
【参考方案1】：我明白了。原因是我的图表中有tf.py_func()，这是在th 上执行的第一个操作，我尝试计算梯度。 tf.py_func() 似乎还有一个限制，没有记录在案 - 可能被认为是显而易见的。
也许我应该报告错误（或提出功能请求）并临时解决问题。
【讨论】：
是的，py_func 不会自动获取渐变。以上是关于在 TensorFlow 中计算图上的梯度不会产生类型提取错误的主要内容，如果未能解决你的问题，请参考以下文章