Antonio Mirarchi

I'm not sure if there's a more efficient way to improve this implementation, if you have it just propose it. Moreover, if you believe that addressing this error, which arises...

It works like a dict so you can do: `data.hparams['load_model']= model.hparams['load_model']`

It's possible to use the `hparams_file` argument in `LNNP.load_from_checkpoint()` to directly specify the path to the YAML file containing the hyperparameters used for training. This ensures that the model (used...

The `dtype` arg is not a problem anymore, we fixed it in #208. The original idea was to update just some parameters, essentially the ones involved in the no-matching error....

It works fine. This is the time column form the metrics.csv: ``` 0 4.779309 1 6.532793 2 8.925880 3 10.662247 4 13.062045 5 14.766587 6 17.175591 7 18.913141 8 21.361235...

I think the Nan is due to reaching the max num epochs by the trainer, so it's not starting a new epoch but it's creating a row in the metrics...

I think the problem is more about when the input.yaml is [written](https://github.com/torchmd/torchmd-net/blob/d921666a89868f2d828894c4b0340bb4ff5f152c/torchmdnet/scripts/train.py#L142). Perhaps, it's enough to modify this [function](https://github.com/torchmd/torchmd-net/blob/d921666a89868f2d828894c4b0340bb4ff5f152c/torchmdnet/utils.py#L312-L327) excluding all the hparms that are not used by the model...