Trained weights for ensemble method are missing

[igorRadic]

Can you add "resnet50_pretrained_vgg_rot30_2019Nov13_08.20" and "resnet18_rot30_2019Nov05_17.44" trained weights on your shared google drive folder for ensemble method, they are missing and I can't get 76.82% accuracy without them. Thanks!

[phamquiluan]

Thanks, @igorRadic or your interest! My ex-university disable my google drive, and my ex-machine just broke last month. So I technically lost the weights 😞

I will try to find it in another old machine of mine. Please wait.

[igorRadic]

Ok, thank you for your effort! 🙂

[igorRadic]

@phamquiluan have you perhaps found trained weights?

[phamquiluan]

@igorRadic I can't so far. The weights are not in the old machine. Hmm, I will keep an eye on this matter and return if I find it.

[igorRadic]

@phamquiluan Maybe I can recreate those two models whose weights are missing, but for recreating "resnet50_pretrained_vgg_rot30_2019Nov13_08.20" I need "resnet50_scratch_dims_2048.pth" weights or not? That's what I assumed because in "resnet50_pretrained_vgg_rot30_2019Nov13_08.20" there is "pretrained" label and in models/resnet50_scratch_dims_2048.py there is commented code for loading weights:

569      class Resnet50_pretrained_vgg(Resnet50_scratch):
570          def __init__(self):
571              super(Resnet50_pretrained_vgg, self).__init__()
572              # state_dict = torch.load('./saved/pretrained/resnet50_scratch_dims_2048.pth')
573              # self.load_state_dict(state_dict)

also i don't understand why is in this model name 'vgg', can you please explain how to recreate "resnet50_pretrained_vgg_rot30_2019Nov13_08.20" model? Thanks!

[phamquiluan]

Wow, you truly impressed me with your determination, @igorRadic. I conducted this research for a long time and forgot almost everything detail about the naming and conventions of using which model and not.

I assume that you are a researcher. Please read this email, the screenshot I attached. I hope this help!

p/s: I tried so much to keep the reproducibility when I started this research years ago... I'm mad at myself so much when this training code can't be reproduced anymore and the school delete my drive :(

[igorRadic]

@phamquiluan thank you for your quick response and honesty, I will definitely try to reproduce your results and If I make it I will let you know.

[chn-lee-yumi]

Hi @igorRadic , have you reproduced the result successfully?

Without these two weights, I got 75.64% test accuracy.

[igorRadic]

Hi @chn-lee-yumi , I got 76.34 % test accuracy without these two weights.

[chn-lee-yumi]

I got 76.34 % test accuracy without these two weights.

@igorRadic That's weird. Theoretically, we should get the same result. Did you use gen_results.py and gen_ensemble.py to get the result?

[igorRadic]

@chn-lee-yumi Maybe that is because I made a few code fixes, try to reproduce 76.34 % test accuracy with my fork of this repository. And yes, I used gen_results.py and gen_ensemble.py scripts for reproducing this result. Also, remove missing weights names and their paths from model_dict dictionary in gen_results.py and gen_ensemble.py scripts and accordingly to these changes change model_dict_proba list in gen_ensemble.py script.

[chn-lee-yumi]

@igorRadic Thanks, I will try your fork. By the way, how did you deal with the missing test_targets.npy? I wrote a simple script read from the json file and converted it.

[igorRadic]

@chn-lee-yumi Problem with test_targets.npy is also fixed in my fork, you can simply download the wights and run gen_results.py and gen_ensemble.py scripts.

[chn-lee-yumi]

@igorRadic Hi, did you archive 74.14% using fer2013_ResMaskingNet_74.14_test_acc_config.json? I only got 68.849%.