import torch
import numpy as np
[docs]def compute_rmse(outputs, targets, atoms_per_image=None):
"""Compute RMSE
Useful when using futures.
Parameters
----------
outputs : list
List of outputs.
targets : list
List if targets.
atoms_per_image : list
List of atoms per image.
Returns
-------
rmse : float
Root-mean squared error.
"""
# Concatenate outputs and targets if they come as list of tensors
if isinstance(outputs, list):
try:
outputs = torch.cat(outputs)
numpy = False
except TypeError:
outputs = np.array(outputs)
numpy = True
if isinstance(targets, list):
try:
targets = torch.cat(targets)
numpy = False
except TypeError:
targets = np.array(targets)
numpy = True
# When doing atomistic models then atoms_per_image exists.
if atoms_per_image is not None:
# Dimensions do not match
outputs = outputs / atoms_per_image
targets = targets / atoms_per_image
if numpy:
rmse = np.sqrt((np.square(outputs - targets)).mean())
else:
rmse = torch.sqrt(torch.mean((outputs - targets).pow(2))).item()
return rmse
[docs]def compute_mse(outputs, targets, atoms_per_image=None):
"""Compute MSE
Useful when using futures.
Parameters
----------
outputs : list
List of outputs.
targets : list
List if targets.
atoms_per_image : list
List of atoms per image.
Returns
-------
mse : float
Mean squared error.
"""
# Concatenate outputs and targets if they come as list of tensors
if isinstance(outputs, list):
try:
outputs = torch.cat(outputs)
numpy = False
except TypeError:
outputs = np.array(outputs)
numpy = True
if isinstance(targets, list):
try:
targets = torch.cat(targets)
numpy = False
except TypeError:
targets = np.array(targets)
numpy = True
# When doing atomistic models then atoms_per_image exists.
if atoms_per_image is not None:
# Dimensions do not match
outputs = outputs / atoms_per_image
targets = targets / atoms_per_image
if numpy:
mse = (np.square(outputs - targets)).mean()
else:
mse = torch.mean((outputs - targets).pow(2)).item()
return mse
[docs]def compute_mae(outputs, targets, atoms_per_image=None):
"""Compute MAE
Useful when using futures.
Parameters
----------
outputs : list
List of outputs.
targets : list
List if targets.
atoms_per_image : list
List of atoms per image.
Returns
-------
mae : float
Mean absolute error.
"""
# Concatenate outputs and targets if they come as list of tensors
if isinstance(outputs, list):
try:
outputs = torch.cat(outputs)
numpy = False
except TypeError:
outputs = np.array(outputs)
numpy = True
if isinstance(targets, list):
try:
targets = torch.cat(targets)
numpy = False
except TypeError:
targets = np.array(targets)
numpy = True
# When doing atomistic models then atoms_per_image exists.
if atoms_per_image is not None:
# Dimensions do not match
outputs = outputs / atoms_per_image
targets = targets / atoms_per_image
if numpy:
mae = (np.abs(outputs - targets)).mean()
else:
mae = torch.mean(torch.abs(outputs - targets)).item()
return mae