the sparse convolution to onnx inference output does not match the torch output

[Xiao-Hu-Z]

After training with my own data, the sparse convolution to onnx inference output does not match the torch output. It is found that the sparse convolution weight is too large and exceeds the fp16 precision range. What may be the cause?

[Xiao-Hu-Z]

稀疏卷积各层的权重，请问下有什么训练策略可以降低参数数值大小