在数值积分期间处理异常的异常

Question

我在TensorFlow中进行基本的轨道力学模拟。当'planet'太靠近'sun'时（当x，y接近（0,0）时），TensorFlow在分割期间会获得异常（这可能有意义）。不知何故，它在异常期间返回异常，导致它完全失败。

我已经尝试使用tf.where有条件地用NaN替换这些除以零，然而，它然后遇到有效的相同错误。我也尝试使用tf.div_no_nan来获得零而不是NaN，但是得到完全相同的错误。

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

def gravity(state, t):
    print(len(tf.unstack(state)))
    x, y, vx, vy = tf.unstack(state)
    # Error is related to next two lines
    fx = -x/tf.pow(tf.reduce_sum(tf.square([x,y]),axis=0),3/2)
    fy = -y/tf.pow(tf.reduce_sum(tf.square([x,y]),axis=0),3/2)
    dvx = fx
    dvy = fy
    return tf.stack([vx, vy, dvx, dvy])

# Num simulations
size = 100

# Initialize at same position with varying y-velocity
init_state = tf.stack([tf.constant(-1.0,shape=(size,)),tf.zeros((size)),tf.zeros((size)),tf.range(0,10,.1)])

t = np.linspace(0, 10, num=5000)
tensor_state, tensor_info = tf.contrib.integrate.odeint(
    gravity, init_state, t, full_output=True)

init = tf.global_variables_initializer()
with tf.Session() as sess:   
    state, info = sess.run([tensor_state, tensor_info])
    state = tf.transpose(state, perm=[1,2,0]).eval()

x, y, vx, vy = state
for i in range(10):
    plt.figure()
    plt.plot(x[i], y[i])
    plt.scatter([0],[0])

我真的得到了

...
InvalidArgumentError: assertion failed: [underflow in dt] [9.0294095248318226e-17]
...
During handling of the above exception, another exception occurred:
...
InvalidArgumentError: assertion failed: [underflow in dt] [9.0294095248318226e-17]
...

我希望这个鸿沟能够产生NaNor Infinity，然后通常会像人们期望的那样传播数值积分。

Answer 1

你可以试试这个

with tf.Session() as sess:
    sess.run(init)
    try:
        state, info = sess.run([tensor_state, tensor_info])
    except tf.errors.InvalidArgumentError:
        state = #Whatever values/shape you need

我不知道你的情况是否合适，但也许你可以添加一些小常数来避免除以零。