def _process_list_for_saving(l: list) -> list:
"""
The yaml.dump function can't save Tensors, ndarrays, or callables, so we cast them to types it can save.
:param l: list containing parameters to save
:return: list with values processable by yaml.dump
"""
copy = deepcopy(l) # do not mutate the input
for i, item in enumerate(copy):
# Check the values of the list
if isinstance(item, (to.Tensor, np.ndarray)):
# Save Tensors as lists
copy[i] = item.tolist()
elif isinstance(item, np.float64):
# PyYAML can not save numpy floats
copy[i] = float(item)
elif isinstance(item, nn.Module):
# Only save the class name as a sting
copy[i] = get_class_name(item)
elif callable(item):
# Only save function name as a sting
try:
copy[i] = str(item)
except AttributeError:
copy[i] = item.__name__
elif isinstance(item, dict):
# If the value is another dict, recursively go through this one
copy[i] = _process_dict_for_saving(item)
elif isinstance(item, list):
# If the value is a list, recursively go through this one
copy[i] = _process_list_for_saving(item)
elif item is None:
copy[i] = 'None'
return copy
def _process_dict_for_saving(d: dict) -> dict:
"""
The yaml.dump function can't save Tensors, ndarrays, or callables, so we cast them to types it can save.
:param d: dict containing parameters to save
:return: dict with values processable by yaml.dump
"""
copy = deepcopy(d) # do not mutate the input
for k, v in copy.items():
# Check the values of the dict
if isinstance(v, (to.Tensor, np.ndarray)):
# Save Tensors as lists
copy[k] = v.tolist()
elif isinstance(v, np.float64):
# PyYAML can not save numpy floats
copy[k] = float(v)
elif isinstance(v, nn.Module):
# Only save the class name as a sting
copy[k] = get_class_name(v)
elif callable(v):
# Only save function name as a sting
try:
copy[k] = str(v)
except AttributeError:
try:
copy[k] = get_class_name(v)
except Exception:
copy[k] = v.__name__
elif isinstance(v, dict):
# If the value is another dict, recursively go through this one
copy[k] = _process_dict_for_saving(v)
elif isinstance(v, list):
# If the value is a list, recursively go through this one
copy[k] = _process_list_for_saving(v)
elif v is None:
copy[k] = 'None'
return copy
def _process_list_for_saving(l: [list, tuple]) -> [list, tuple]:
"""
The yaml.dump function can't save PyTorch tensors, numpy arrays, or callables, so we cast them to types it can save.
:param l: list or tuple containing parameters to save
:return: list or tuple with values processable by yaml.dump
"""
# Do not mutate the input. Convert tuple to list to make elements mutable
copy = list(deepcopy(l))
for i, item in enumerate(copy):
# Check the values of the list
if isinstance(item, (to.Tensor, np.ndarray)):
# Save Tensors as lists
copy[i] = item.tolist()
elif isinstance(item, np.float64):
# PyYAML can not save numpy floats
copy[i] = float(item)
elif isinstance(item, nn.Module):
# Only save the class name as a sting
copy[i] = get_class_name(item)
elif callable(item):
# Only save function name as a sting
try:
copy[i] = str(item)
except AttributeError:
copy[i] = item.__name__
elif isinstance(item, dict):
# If the value is another dict, recursively go through this one
copy[i] = _process_dict_for_saving(item)
elif isinstance(item, (list, tuple)):
# If the value is a list or tuple, recursively go through this one
copy[i] = _process_list_for_saving(item)
elif item is None:
copy[i] = "None"
# The returned object should be of the same type as the input
return copy if isinstance(l, list) else tuple(copy)