Closed ghost closed 3 years ago
You need to use the same settings in networks.fermi_net
as you used in training. In particular, the envelope_type
and full_det
arguments should match cfg.network.envelope_type
and cfg.network.full_det
.
is envelope_type
and full_det
in networks.fermi_net
for the current Neon result the same as the current settings in cfg.network.envelope_type
and cfg.network.full_det
? If not, what should these values be?
Also, what are the param settings to reproduce the other elements? A list of params to reproduce the results would be helpful :)
By "current Neon result" do you mean from the Phys Rev Research paper? That was produced with the TF version (which is essentially deprecated). The isotropic
envelope setting was introduced in the NeurIPS workshop paper, which gives motivation for this and comparison to the full
envelope. All results published before then used the full
envelope setting.
The full_det
setting is experimental. Set it to False
to match the published results.
For neon, neither of these settings will make a substantial difference to the final energy within statistical errors. Note your batch size is very small and might limit the accuracy you achieve.
Please refer to our papers for the settings used in the models. The PRR paper used the same set of configuration options for all experiments, except where noted, and this broadly matches the current defaults (except for full_det
and envelope_type
).
Hi, I am trying to reproduce the results for Neon, I am running the following code with default base config and only changes to batch_size = 256, pretrain iterations = 100 and optim iterations = 100_000 (for now, will be increased if results not matched):
Training Code
Loading Model
At this step
loss_ = ploss(params, data)
, I am getting this error:I compared my
params
with the cloud files given and it seems mypi
andsigma
envelopes have different shapes.Any help on how to reproduce the pretrained results would be appreciated.