Pytorch expert to check why the train MAE is 0.02 but test MAE is 0.22 so much overfit in deep learning graph study -- 2
Budget: $10 – $50 USD
I need to figure out why using the file I provided to replace the core.model.py file in the source code: https://github.com/XiaoxinHe/Graph-MLPMixer/ overfit so much. Now the train MAE is 0.03 but test MAE is 0.21. I need to make the test MAE from 0.21 become 0.06 to 0.07 (now train MAE is around 0.03 to 0.04), without other data augmentation skills, (I tried dropout, early stopping already but no effects). Please do not do hyperparameter tuning, as basically the model file I provided compared with baseline (baseline train MAE is 0.02-0.03 and test MAE is 0.07) is just model architecture difference, all other setting same. So overall, I need make the test MAE from 0.21 become 0.06 to 0.07 (below 0.075) without hyperparameter changing or data augmention
I have the code, running graph regression task using pytorch and pytorch geometric, for deep learning. The train MAE is 0.02 but the test MAE is 0.22, which is clearly overfit, I want to know why so much difference in train and test. I want to solve the problem and make test MAE reasonably improve to 0.07.
I have code which can run, but the train MAE is 0.408 while test MAE is 0.202, which is clearly overfit and even use dropout, regularisation still, do not know why, the data was used from pytorch geometric library.
mostly the code is based on : https://github.com/XiaoxinHe/Graph-MLPMixer for the ZINC dataset. only the core.model.py has bit difference, replace it with this file:
this it, you can see result: 'Epoch: 1999, Train perf: 0.0408, Train Loss: -0.0408, Val: -0.2401, Test: -0.2024, Seconds: 23.4236'
train MAE is 0.04 but test is remaining 0.2024
when you think can solve the problem? Basically I want to know why there is such big overfit, and successfully make the test MAE become 0.07 approximately .
all the code is : https://github.com/XiaoxinHe/Graph-MLPMixer/, just the model.py in the core folder replace with the file I sent to you.
just make the test MAE become 0.07 , with train MAE around 0.02-0.04, use the file I sent to you as the core.model.py
Just to be sure you understand my requirement, I need to figure out why overfit so much and make the test MAE from 0.21 become 0.06 to 0.07 (now train MAE is around 0.03 to 0.04), without other data augmentation, basically the model file I provided compared with baseline is just model architecture difference, all other setting same.
basicaly just want to see from the model architecture level why overfit and make it 0.07 test MAE.
thanks.
I have the code, running graph regression task using pytorch and pytorch geometric, for deep learning. The train MAE is 0.02 but the test MAE is 0.22, which is clearly overfit, I want to know why so much difference in train and test. I want to solve the problem and make test MAE reasonably improve to 0.07.
I have code which can run, but the train MAE is 0.408 while test MAE is 0.202, which is clearly overfit and even use dropout, regularisation still, do not know why, the data was used from pytorch geometric library.
mostly the code is based on : https://github.com/XiaoxinHe/Graph-MLPMixer for the ZINC dataset. only the core.model.py has bit difference, replace it with this file:
this it, you can see result: 'Epoch: 1999, Train perf: 0.0408, Train Loss: -0.0408, Val: -0.2401, Test: -0.2024, Seconds: 23.4236'
train MAE is 0.04 but test is remaining 0.2024
when you think can solve the problem? Basically I want to know why there is such big overfit, and successfully make the test MAE become 0.07 approximately .
all the code is : https://github.com/XiaoxinHe/Graph-MLPMixer/, just the model.py in the core folder replace with the file I sent to you.
just make the test MAE become 0.07 , with train MAE around 0.02-0.04, use the file I sent to you as the core.model.py
Just to be sure you understand my requirement, I need to figure out why overfit so much and make the test MAE from 0.21 become 0.06 to 0.07 (now train MAE is around 0.03 to 0.04), without other data augmentation, basically the model file I provided compared with baseline is just model architecture difference, all other setting same.
basicaly just want to see from the model architecture level why overfit and make it 0.07 test MAE.
thanks.