How to get the benchmark performance on FER2013?

[rrryan2016]

Thanks for your great job and kind sharing.

I am doing some work on FER, and intend to get the performance of some classic network (including Resnet34, Resnet152, etc.) as described in https://github.com/phamquiluan/ResidualMaskingNetwork#benchmarking-on-fer2013.

But I cannot get the accuracy of 70%+, but only around 60%, though I tried to refer to parameters in https://github.com/phamquiluan/ResidualMaskingNetwork/blob/master/configs/fer2013_config.json.

Any recommendation and suggestion please?

[phamquiluan]

@rrryan2016 use Imagenet pre-trained weights

[rrryan2016]

@rrryan2016 use Imagenet pre-trained weights

Yep, I used the pretrained model. :P

[phamquiluan]

exp1: no pre-trained
exp2: pre-trained

@rrryan2016 try to conduct two exp above, show me the results.

[rrryan2016]

exp1: no pre-trained
exp2: pre-trained

@rrryan2016 try to conduct two exp above, show me the results.

Thanks for your reply and patience.

Taking the accuracy on Public testset as example,

	Densenet121	Resnet152	VGG19_bn
Unpretrained	0.462803	0.373084	0.529117
Pretrained	0.622458	0.615770	0.651714

Envs: Python 3.7, CUDA 11.0, 3090Ti

[phamquiluan]

@rrryan2016 run pip list command and show me the results

[rrryan2016]

Package           Version
----------------- -------------------
backcall          0.2.0
cached-property   1.5.2
certifi           2020.12.5
cmake             3.20.2
cycler            0.10.0
decorator         4.4.2
h5py              3.2.1
imageio           2.9.0
intel-openmp      2021.2.0
ipython           7.24.1
ipython-genutils  0.2.0
jedi              0.18.0
kiwisolver        1.3.1
matplotlib        3.4.2
matplotlib-inline 0.1.2
mkl               2021.2.0
mkl-include       2021.2.0
networkx          2.5.1
ninja             1.10.0.post2
numpy             1.20.3
opencv-python     4.5.2.54
pandas            1.2.4
parso             0.8.2
pexpect           4.8.0
pickleshare       0.7.5
Pillow            8.2.0
pip               21.1.1
prompt-toolkit    3.0.18
protobuf          3.17.3
ptyprocess        0.7.0
Pygments          2.9.0
pyparsing         2.4.7
python-dateutil   2.8.1
pytz              2021.1
PyWavelets        1.1.1
PyYAML            5.4.1
scikit-image      0.18.1
scipy             1.6.3
setuptools        52.0.0.post20210125
six               1.16.0
tbb               2021.2.0
tensorboardX      2.2
tifffile          2021.6.6
torch             1.8.0+cu111
torchaudio        0.8.0
torchvision       0.9.0+cu111
traitlets         5.0.5
typing-extensions 3.10.0.0
wcwidth           0.2.5
wheel             0.36.2

[phamquiluan]

@rrryan2016 hmm, please try to degrade torch and torchvision version

v1: pip install torch==1.3.1 torchvision==0.4.0
v2: pip install torch==1.4.0 torchvision==0.5.0

[phamquiluan]

@rrryan2016 And run again the experiments 🙏

[phamquiluan]

@rrryan2016 any progress, sir?

[rrryan2016]

Thanks for the concerns. But sorry, I intend to give up the version switch.

1. I guess the torch version makes little difference to the accuracy.

2. The torch version may not be compatible with CUDA 11.0 that I am using. It can be very time-consuming if I focus on the switch.

[phamquiluan]

@rrryan2016
In this notebook, I show you how to get 73% with resnet34 only.

https://colab.research.google.com/drive/1LbDiAs5xOmhwaoKtJepaK_oVU_IkgLM8?usp=sharing

[rrryan2016]

@rrryan2016
In this notebook, I show you how to get 73% with resnet34 only.

https://colab.research.google.com/drive/1LbDiAs5xOmhwaoKtJepaK_oVU_IkgLM8?usp=sharing

Grateful, it is helpful!

[ranjiewwen]

FER2013 resolution is 48*48, why input model resize to (224, 224), this operate will get blur or damage image quality?

[Sarah-em]

@rrryan2016 In this notebook, I show you how to get 73% with resnet34 only.

https://colab.research.google.com/drive/1LbDiAs5xOmhwaoKtJepaK_oVU_IkgLM8?usp=sharing

the link is not working ...can you please reshare it again...thank you