search

xingyul / flownet3d

FlowNet3D: Learning Scene Flow in 3D Point Clouds (CVPR 2019)
MIT License
359 stars 83 forks source link

test get NaN loss and unreasonable prediction #27

Closed Guptajakala closed 4 years ago

Guptajakala commented 4 years ago

Hi, I downloaded your preprocessed Flyingthings3d dataset and when I run the evaluate command I get these unreasonable numbers for amost every batch:

loss_val nan
pred_val [[[ 7.80575696e+36  3.97343718e+36  2.06021729e+36]
  [ 5.75067476e+36  1.49935495e+36  1.30909517e+36]
  [ 1.14683589e+37  7.62623228e+36  2.77971755e+36]
  ...
  [ 1.69982565e+37 -2.35347431e+36 -7.84758246e+36]
  [ 1.00178770e+37  7.96418159e+36  3.90681105e+36]
  [ 1.12353090e+37  1.07022519e+36  7.07184686e+34]]

 [[ 4.45941417e+35 -2.48141516e+35  9.25996419e+33]
  [ 3.36659517e+35 -4.07464617e+35 -1.70080173e+34]
  [-6.47037686e+34 -1.89484747e+35 -7.36334955e+34]
  ...
  [ 3.55201482e+35 -4.03794649e+35 -1.92975676e+34]
  [ 4.68557136e+35  5.35420753e+34  1.08498641e+35]
  [ 2.90908917e+35 -8.76533249e+34  3.15325586e+34]]

 [[-8.20998982e+35 -7.27648726e+34  2.14235447e+35]
  [-1.01996078e+36 -2.44069148e+35 -5.02296449e+35]
  [-2.28599505e+36  5.09481255e+35 -3.73197565e+35]
  ...
  [-1.46082059e+36 -2.59530524e+35 -8.53456066e+35]
  [-1.59142710e+36  2.68669968e+35 -2.88856551e+35]
  [-1.02024886e+36 -2.46612155e+35 -5.03460509e+35]]

Is there anything I'm doing wrong?

Yang-L1 commented 4 years ago

Did you solved this issue?