Different results with AlexNet on SHHB

[MagazzuGaetano]

I was trying to reproduce the results reported in the repo for the alexnet experiment on the SHHB dataset.

I trained alexnet using the recommended parameters in "results_reports/AlexNet/SHHB/config.py" "results_reports/AlexNet/SHHB/setting.py" and stopped at 400 epochs instead of 406 epochs which is reported in the "results_reports/AlexNet/SHHB/readme.md",

I initialized the weights of the network with those of pre-trained on ImageNet, but the results do not seem to be correct.

########################### results obtained: mse: 20.41, mae: 13.22 (validated on the test set) repo results: mse: 21.7, mae: 13.6 ###########################

In addition to the final [mse / mae] results which are different, I note that the first iterations already have different [mse / mae / loss] values, other than those reported in the github repo

I was wondering if in the experiment reported in the repo has been considered the following observations:

• The RandomCrop in "./datasets/SHHB/loading_data.py"?
• The use of a pre-trained model on GCC instead of on ImageNet?

I report the log of my experiment: [Log_exp1.txt] (https://github.com/gjy3035/C-3-Framework/files/4667995/log_exp1.txt)

[gjy3035]

1. Slight difference is normal
2. ImageNet

[MagazzuGaetano]

thanks