Replies: 4 comments
-
Please take the logarithm of both x-axis and y-axis and plot a new figure. Thanks! |
Beta Was this translation helpful? Give feedback.
-
Here the log plot |
Beta Was this translation helpful? Give feedback.
-
The oscillating behavior should be due to the fact that the on-the-fly
training and testing errors are measured by the data randomly visited in
different systems. So the accuracy of the model is different for
different systems.
On the other hand, it seems that 1) the training error is large and 2)
there seems to be some overfitting since the testing accuracy is worse than
training.
For 1), could you also show the error of forces? I also find some places
that might improve the accuracy: "neuron": [128,64,32,16], could be changed
to, say, [20, 40, 80], since it is embedding a one-dimensional quantity (r)
onto a high-dimensional space; start_lr is typically 1e-3. Please see other
reference values at
https://github.com/deepmodeling/deepmd-kit/blob/master/examples/water/train/water_se_a.json
For 2), I feel that 60000 training data is a lot. Is there any reason why
overfitting happens?
…On Fri, Feb 7, 2020 at 1:19 AM Marius ***@***.***> wrote:
[image: log_training_error]
<https://user-images.githubusercontent.com/13686746/73992727-513a0b00-4993-11ea-93de-c92df3d5a0af.png>
Here the log plot
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<https://github.com/deepmodeling/deepmd-kit/issues/175?email_source=notifications&email_token=AEJ6DC2A3QPHGZ5PWHN7RW3RBSZKXA5CNFSM4KRGN2W2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOELBMKBQ#issuecomment-583189766>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEJ6DC7IUABZ2ZKBE6AWRBLRBSZKXANCNFSM4KRGN2WQ>
.
|
Beta Was this translation helpful? Give feedback.
-
I changed the setting following your suggestions, here are the force and energy errors. It did not really improve much on the oscillating behavior. I honestly don't see the reason for overfitting. I think it got better now but difficult to say, though. |
Beta Was this translation helpful? Give feedback.
-
Hi,
After the both training and test error decreased in the initial steps, both starts to oscillate later in the training (see attached figure). I have ~ 60000 training sets, which is split into 4 sets with different number of atoms if this makes a different. The input is
{
"_comment": " model parameters",
"model": {
"type_map": ["K"],
"descriptor" :{
"type": "se_ar",
"a" : {
"sel": [120],
"rcut_smth": 9.0,
"rcut": 10.0,
"neuron": [128,64,32,16],
"resnet_dt": false,
"axis_neuron": 16,
"seed": 1,
"_comment": " that's all"
},
"r": {
"sel": [120],
"rcut_smth": 10.0,
"rcut": 13.0,
"neuron": [128,64,32,16],
"resnet_dt": false,
"seed": 1,
"_comment": " that's all"
}
},
"fitting_net" : {
"neuron": [128,128,128,128],
"resnet_dt": false,
"seed": 1,
"_comment": " that's all"
},
"_comment": " that's all"
},
}
(base)
Best,
Marius
Beta Was this translation helpful? Give feedback.
All reactions