mtconv not improve acc compare with depthwise ,is the training data too small matter?
Closed this issue · 2 comments
l2009312042 commented
it is a great job ,thanks for sharing ! the following is my test result , from this ,the acc is not improve ,can you give me some advice how to improve the acc ?
data | model | acc | size |
---|---|---|---|
net12 | 96.6 | 1.3m | |
net6 | 96.7 | 785kb | |
net12_mt_noconvert | 96.56 | 2.41m | |
net12_mt_convert | 96.56 | 1.78m | |
net6_mt_noconvert | 96.75 | 1.31m | |
net6_mt_convert | 96.75 | 0.99m |
Interlagos commented
Thanks for your attention to our work.
We recommend training TENet with the default hyper-parameters in this repo, especially the weight decay is set to 4e-5 instead of 1e-4, which affects the result a lot.
We retrain the following models 10 times, respectively. Here is the result:
TENet12_MTConv | TENet12 | TENet6_MTConv | TENet6 | |
---|---|---|---|---|
no.1 | 96.7543 | 96.6569 | 96.7218 | 96.5920 |
no.2 | 96.9815 | 96.4946 | 96.7868 | 96.3648 |
no.3 | 97.1438 | 96.3973 | 96.6245 | 96.3973 |
no.4 | 96.8517 | 96.2025 | 96.2999 | 96.4946 |
no.5 | 96.8814 | 96.5271 | 96.2350 | 96.4622 |
no.6 | 96.5920 | 96.8192 | 96.6569 | 96.7868 |
no.7 | 96.5324 | 96.8841 | 97.1762 | 96.2674 |
no.8 | 96.6569 | 96.4622 | 96.4297 | 96.6245 |
no.9 | 97.0464 | 96.5271 | 96.5271 | 96.5271 |
no.10 | 96.6569 | 96.4946 | 96.7543 | 96.5271 |
mean | 96.8097 | 96.5466 | 96.6212 | 96.5044 |
Hope it can help you.
l2009312042 commented
thanks for your advice