add any multiple of 16 to the output of NerfNetwork::padded_output_width() for instance: return std::max(m_rgb_network->padded_output_width(), (uint32_t)4) + 16;
build and run
What happens :
the initialization works fine.
the second step is seemingly also fine.
from the third training step onward, the loss become nan.
What should happen :
From what I understand as padding is only added to to match the size of the tensor cores, changing its size should not impact in anyway the behaviour of the program. Am I wrong ?
Steps to reproduce :
NerfNetwork::padded_output_width()
for instance:return std::max(m_rgb_network->padded_output_width(), (uint32_t)4) + 16;
What happens :
nan
.What should happen : From what I understand as padding is only added to to match the size of the tensor cores, changing its size should not impact in anyway the behaviour of the program. Am I wrong ?