imagenet-100 result cannot reproduce

[zhaoedf]

for cifar and imagenet-1000, i can get similar results as what paper reported.
how ever, for imangenet-100 only, i cannot get the same result, with 3~4% deviation.

the followings are my running scripts and experiment results.

scipts

bash train.sh 6,7
--options options/data/imagenet100_10-10.yaml options/data/imagenet100_order1.yaml options/model/imagenet_dytox.yaml \
--name dytox \
--data-path ~/ilsvrc2012 \
--output-basedir ./checkpoints

results (i ran two times on different machines)

i focus on both avg and last acc.

[arthurdouillard]

Hello, I've saw your issue. I'll first look into #5, which may (or not) explain your issue and #2

I'll get back to you asap. Sorry for the inconvenience.

[zhaoedf]

Hello, I've saw your issue. I'll first look into #5, which may (or not) explain your issue and #2

I'll get back to you asap. Sorry for the inconvenience.

ok, i totally understand.

[GengDavid]

Hi @arthurdouillard
I also tried ImageNet-100 and got a similar result as @zhaoedf.

BTW, I wonder that is it possible to share your trained checkpoints with us?

[GengDavid]

Hi all, I also tried ImageNet-1000 but got a lower result:(
I am using Pytorch 1.10.1, CUDA 10.1

I'll try to use a different Pytorch version and try again.

[arthurdouillard]

Please see erratum.