search

Swall0w / torchstat

Model analyzer in PyTorch
MIT License
1.46k stars 144 forks source link

‘Total MAdd’ and 'Total Flops' results should exchange with each other #12

Open Lukzin opened 5 years ago

Lukzin commented 5 years ago

MAdd means that the multiply and add operations are counted as 1 operation while FLOPs count them as 2. Should the results of 'Total MAdd' and 'Total Flops' exchange with each other?

andyhahaha commented 5 years ago

I have the same question. And why compute_flops need to multiply batch_size but compuute_MAdd didn't. I think there is no need to multiply batch_size, because we only consider one image situation. Is that right?

Sixkplus commented 5 years ago

same question

DocF commented 4 years ago

same question

clamcy commented 4 years ago

+1

supratikbanerjee commented 4 years ago

+1

qiuhaining commented 4 years ago

+1

schutera commented 4 years ago

+1

minjs commented 4 years ago

+1

BebDong commented 4 years ago

Same question. I reimplement the SwiftNetRN-18 with the MXNet framework, resulting in over 200 GFLOPs of input size 1024*2048 that is similar to (May be) BUGs in paper: wrong Gflops #14

zhoumenghan commented 2 years ago

+1!!!! I think ‘Total MAdd’ and 'Total Flops' results should exchange with each other.

hychiang-git commented 2 years ago

Same question

qiuhaining commented 2 years ago

收到了！谢谢！