def softmax(*args, hardness: float = 1.0):
"""
An element-wise softmax between two or more arrays. Also referred to as the logsumexp() function.
Useful for optimization because it's differentiable and preserves convexity!
Great writeup by John D Cook here:
https://www.johndcook.com/soft_maximum.pdf
Args:
Provide any number of arguments as values to take the softmax of.
hardness: Hardness parameter. Higher values make this closer to max(x1, x2).
Returns:
Soft maximum of the supplied values.
"""
if _np.any(hardness <= 0):
raise ValueError("The value of `hardness` must be positive.")
if len(args) <= 1:
raise ValueError(
"You must call softmax with the value of two or more arrays that you'd like to take the "
"element-wise softmax of.")
### Scale the args by hardness
args = [arg * hardness for arg in args]
### Find the element-wise max and min of the arrays:
min = args[0]
max = args[0]
for arg in args[1:]:
min = _np.fmin(min, arg)
max = _np.fmax(max, arg)
out = max + _np.log(sum([_np.exp(array - max) for array in args]))
out = out / hardness
return out
def variable(self,
init_guess: Union[float, np.ndarray],
n_vars: int = None,
scale: float = None,
freeze: bool = False,
log_transform: bool = False,
category: str = "Uncategorized",
lower_bound: float = None,
upper_bound: float = None,
) -> cas.MX:
"""
Initializes a new decision variable (or vector of decision variables). You must pass an initial guess (
`init_guess`) upon defining a new variable. Dimensionality is inferred from this initial guess, but it can be
overridden; see below for syntax.
It is highly, highly recommended that you provide a scale (`scale`) for each variable, especially for
nonconvex problems, although this is not strictly required.
Args:
init_guess: Initial guess for the optimal value of the variable being initialized. This is where in the
design space the optimizer will start looking.
This can be either a float or a NumPy ndarray; the dimension of the variable (i.e. scalar,
vector) that is created will be automatically inferred from the shape of the initial guess you
provide here. (Although it can also be overridden using the `n_vars` parameter; see below.)
For scalar variables, your initial guess should be a float:
>>> opti = asb.Opti()
>>> scalar_var = opti.variable(init_guess=5) # Initializes a scalar variable at a value of 5
For vector variables, your initial guess should be either:
* a float, in which case you must pass the length of the vector as `n_vars`, otherwise a scalar
variable will be created:
>>> opti = asb.Opti()
>>> vector_var = opti.variable(init_guess=5, n_vars=10) # Initializes a vector variable of length
>>> # 10, with all 10 elements set to an initial guess of 5.
* a NumPy ndarray, in which case each element will be initialized to the corresponding value in
the given array:
>>> opti = asb.Opti()
>>> vector_var = opti.variable(init_guess=np.linspace(0, 5, 10)) # Initializes a vector variable of
>>> # length 10, with all 10 elements initialized to linearly vary between 0 and 5.
In the case where the variable is to be log-transformed (see `log_transform`), the initial guess
should not be log-transformed as well - just supply the initial guess as usual. (Log-transform of the
initial guess happens under the hood.) The initial guess must, of course, be a positive number in
this case.
n_vars: [Optional] Used to manually override the dimensionality of the variable to create; if not
provided, the dimensionality of the variable is inferred from the initial guess `init_guess`.
The only real case where you need to use this argument would be if you are initializing a vector
variable to a scalar value, but you don't feel like using `init_guess=value * np.ones(n_vars)`.
For example:
>>> opti = asb.Opti()
>>> vector_var = opti.variable(init_guess=5, n_vars=10) # Initializes a vector variable of length
>>> # 10, with all 10 elements set to an initial guess of 5.
scale: [Optional] Approximate scale of the variable.
For example, if you're optimizing the design of a automobile and setting the tire diameter as an
optimization variable, you might choose `scale=0.5`, corresponding to 0.5 meters.
Properly scaling your variables can have a huge impact on solution speed (or even if the optimizer
converges at all). Although most modern second-order optimizers (such as IPOPT, used here) are
theoretically scale-invariant, numerical precision issues due to floating-point arithmetic can make
solving poorly-scaled problems really difficult or impossible. See here for more info:
https://web.casadi.org/blog/nlp-scaling/
If not specified, the code will try to pick a sensible value by defaulting to the `init_guess`.
freeze: [Optional] This boolean tells the optimizer to "freeze" the variable at a specific value. In
order to select the determine to freeze the variable at, the optimizer will use the following logic:
* If you initialize a new variable with the parameter `freeze=True`: the optimizer will freeze
the variable at the value of initial guess.
>>> opti = Opti()
>>> my_var = opti.variable(init_guess=5, freeze=True) # This will freeze my_var at a value of 5.
* If the Opti instance is associated with a cache file, and you told it to freeze a specific
category(s) of variables that your variable is a member of, and you didn't manually specify to
freeze the variable: the variable will be frozen based on the value in the cache file (and ignore
the `init_guess`). Example:
>>> opti = Opti(cache_filename="my_file.json", variable_categories_to_freeze=["Wheel Sizing"])
>>> # Assume, for example, that `my_file.json` was from a previous run where my_var=10.
>>> my_var = opti.variable(init_guess=5, category="Wheel Sizing")
>>> # This will freeze my_var at a value of 10 (from the cache file, not the init_guess)
* If the Opti instance is associated with a cache file, and you told it to freeze a specific
category(s) of variables that your variable is a member of, but you then manually specified that
the variable should be frozen: the variable will once again be frozen at the value of `init_guess`:
>>> opti = Opti(cache_filename="my_file.json", variable_categories_to_freeze=["Wheel Sizing"])
>>> # Assume, for example, that `my_file.json` was from a previous run where my_var=10.
>>> my_var = opti.variable(init_guess=5, category="Wheel Sizing", freeze=True)
>>> # This will freeze my_var at a value of 5 (`freeze` overrides category loading.)
Motivation for freezing variables:
The ability to freeze variables is exceptionally useful when designing engineering systems. Let's say
we're designing an airplane. In the beginning of the design process, we're doing "clean-sheet" design
- any variable is up for grabs for us to optimize on, because the airplane doesn't exist yet!
However, the farther we get into the design process, the more things get "locked in" - we may have
ordered jigs, settled on a wingspan, chosen an engine, et cetera. So, if something changes later (
let's say that we discover that one of our assumptions was too optimistic halfway through the design
process), we have to make up for that lost margin using only the variables that are still free. To do
this, we would freeze the variables that are already decided on.
By categorizing variables, you can also freeze entire categories of variables. For example,
you can freeze all of the wing design variables for an airplane but leave all of the fuselage
variables free.
This idea of freezing variables can also be used to look at off-design performance - freeze a
design, but change the operating conditions.
log_transform: [Optional] Advanced use only. A flag of whether to internally-log-transform this variable
before passing it to the optimizer. Good for known positive engineering quantities that become nonsensical
if negative (e.g. mass). Log-transforming these variables can also help maintain convexity.
category: [Optional] What category of variables does this belong to?
Usage notes:
When using vector variables, individual components of this vector of variables can be accessed via normal
indexing. Example:
>>> opti = asb.Opti()
>>> my_var = opti.variable(n_vars = 5)
>>> opti.subject_to(my_var[3] >= my_var[2]) # This is a valid way of indexing
>>> my_sum = asb.sum(my_var) # This will sum up all elements of `my_var`
Returns:
The variable itself as a symbolic CasADi variable (MX type).
"""
### Set defaults
if n_vars is None: # Infer dimensionality from init_guess if it is not provided
n_vars = np.length(init_guess)
if scale is None: # Infer a scale from init_guess if it is not provided
if log_transform:
scale = 1
else:
scale = np.mean(np.fabs(init_guess)) # Initialize the scale to a heuristic based on the init_guess
if scale == 0: # If that heuristic leads to a scale of 0, use a scale of 1 instead.
scale = 1
# scale = np.fabs(
# np.where(
# init_guess != 0,
# init_guess,
# 1
# ))
# Validate the inputs
if log_transform:
if np.any(init_guess <= 0):
raise ValueError(
"If you are initializing a log-transformed variable, the initial guess(es) must all be positive.")
if np.any(scale <= 0):
raise ValueError("The 'scale' argument must be a positive number.")
# If the variable is in a category to be frozen, fix the variable at the initial guess.
is_manually_frozen = freeze
if category in self.variable_categories_to_freeze:
freeze = True
# If the variable is to be frozen, return the initial guess. Otherwise, define the variable using CasADi symbolics.
if freeze:
var = self.parameter(n_params=n_vars, value=init_guess)
else:
if not log_transform:
var = scale * super().variable(n_vars)
self.set_initial(var, init_guess)
else:
log_scale = scale / init_guess
log_var = log_scale * super().variable(n_vars)
var = np.exp(log_var)
self.set_initial(log_var, np.log(init_guess))
# Track the variable
if category not in self.variables_categorized: # Add a category if it does not exist
self.variables_categorized[category] = []
self.variables_categorized[category].append(var)
var.is_manually_frozen = is_manually_frozen
# Apply bounds
if lower_bound is not None:
self.subject_to(var >= lower_bound)
if upper_bound is not None:
self.subject_to(var <= upper_bound)
return var
def repanel(
self,
n_points_per_side: int = 100,
) -> 'Airfoil':
"""
Returns a repaneled version of the airfoil with cosine-spaced coordinates on the upper and lower surfaces.
:param n_points_per_side: Number of points per side (upper and lower) of the airfoil [int]
Notes: The number of points defining the final airfoil will be n_points_per_side*2-1,
since one point (the leading edge point) is shared by both the upper and lower surfaces.
:return: Returns the new airfoil.
"""
upper_original_coors = self.upper_coordinates(
) # Note: includes leading edge point, be careful about duplicates
lower_original_coors = self.lower_coordinates(
) # Note: includes leading edge point, be careful about duplicates
# Find distances between coordinates, assuming linear interpolation
upper_distances_between_points = (
(upper_original_coors[:-1, 0] - upper_original_coors[1:, 0])**2 +
(upper_original_coors[:-1, 1] - upper_original_coors[1:, 1])**
2)**0.5
lower_distances_between_points = (
(lower_original_coors[:-1, 0] - lower_original_coors[1:, 0])**2 +
(lower_original_coors[:-1, 1] - lower_original_coors[1:, 1])**
2)**0.5
upper_distances_from_TE = np.hstack(
(0, np.cumsum(upper_distances_between_points)))
lower_distances_from_LE = np.hstack(
(0, np.cumsum(lower_distances_between_points)))
upper_distances_from_TE_normalized = upper_distances_from_TE / upper_distances_from_TE[
-1]
lower_distances_from_LE_normalized = lower_distances_from_LE / lower_distances_from_LE[
-1]
distances_from_TE_normalized = np.hstack(
(upper_distances_from_TE_normalized,
1 + lower_distances_from_LE_normalized[1:]))
# Generate a cosine-spaced list of points from 0 to 1
cosspaced_points = np.cosspace(0, 1, n_points_per_side)
s = np.hstack((
cosspaced_points,
1 + cosspaced_points[1:],
))
# Check that there are no duplicate points in the airfoil.
if np.any(np.diff(distances_from_TE_normalized) == 0):
raise ValueError(
"This airfoil has a duplicated point (i.e. two adjacent points with the same (x, y) coordinates), so you can't repanel it!"
)
x = interp1d(
distances_from_TE_normalized,
self.x(),
kind="cubic",
)(s)
y = interp1d(
distances_from_TE_normalized,
self.y(),
kind="cubic",
)(s)
return Airfoil(name=self.name, coordinates=stack_coordinates(x, y))
def __init__(
self,
model: Callable[
[Union[np.ndarray,
Dict[str, np.ndarray]], Dict[str, float]], np.ndarray],
x_data: Union[np.ndarray, Dict[str, np.ndarray]],
y_data: np.ndarray,
parameter_guesses: Dict[str, float],
parameter_bounds: Dict[str, tuple] = None,
residual_norm_type: str = "L2",
fit_type: str = "best",
weights: np.ndarray = None,
put_residuals_in_logspace: bool = False,
verbose=True,
):
"""
Fits an analytical model to n-dimensional unstructured data using an automatic-differentiable optimization approach.
Args:
model: The model that you want to fit your dataset to. This is a callable with syntax f(x, p) where:
* x is a dict of dependent variables. Same format as x_data [dict of 1D ndarrays of length n].
* If the model is one-dimensional (e.g. f(x1) instead of f(x1, x2, x3...)), you can instead interpret x
as a 1D ndarray. (If you do this, just give `x_data` as an array.)
* p is a dict of parameters. Same format as param_guesses [dict with syntax param_name:param_value].
Model should return a 1D ndarray of length n.
Basically, if you've done it right:
>>> model(x_data, parameter_guesses)
should evaluate to a 1D ndarray where each x_data is mapped to something analogous to y_data. (The fit
will likely be bad at this point, because we haven't yet optimized on param_guesses - but the types
should be happy.)
Model should use aerosandbox.numpy operators.
The model is not allowed to make any in-place changes to the input `x`. The most common way this
manifests itself is if someone writes something to the effect of `x += 3` or similar. Instead, write `x =
x + 3`.
x_data: Values of the dependent variable(s) in the dataset to be fitted. This is a dictionary; syntax is {
var_name:var_data}.
* If the model is one-dimensional (e.g. f(x1) instead of f(x1, x2, x3...)), you can instead supply x_data
as a 1D ndarray. (If you do this, just treat `x` as an array in your model, not a dict.)
y_data: Values of the independent variable in the dataset to be fitted. [1D ndarray of length n]
parameter_guesses: a dict of fit parameters. Syntax is {param_name:param_initial_guess}.
* Parameters will be initialized to the values set here; all parameters need an initial guess.
* param_initial_guess is a float; note that only scalar parameters are allowed.
parameter_bounds: Optional: a dict of bounds on fit parameters. Syntax is {"param_name":(min, max)}.
* May contain only a subset of param_guesses if desired.
* Use None to represent one-sided constraints (i.e. (None, 5)).
residual_norm_type: What error norm should we minimize to optimize the fit parameters? Options:
* "L1": minimize the L1 norm or sum(abs(error)). Less sensitive to outliers.
* "L2": minimize the L2 norm, also known as the Euclidian norm, or sqrt(sum(error ** 2)). The default.
* "Linf": minimize the L_infinty norm or max(abs(error)). More sensitive to outliers.
fit_type: Should we find the model of best fit (i.e. the model that minimizes the specified residual norm),
or should we look for a model that represents an upper/lower bound on the data (useful for robust surrogate
modeling, so that you can put bounds on modeling error):
* "best": finds the model of best fit. Usually, this is what you want.
* "upper bound": finds a model that represents an upper bound on the data (while still trying to minimize
the specified residual norm).
* "lower bound": finds a model that represents a lower bound on the data (while still trying to minimize
the specified residual norm).
weights: Optional: weights for data points. If not supplied, weights are assumed to be uniform.
* Weights are automatically normalized. [1D ndarray of length n]
put_residuals_in_logspace: Whether to optimize using the logarithmic error as opposed to the absolute error
(useful for minimizing percent error).
Note: If any model outputs or data are negative, this will raise an error!
verbose: Should the progress of the optimization solve that is part of the fitting be displayed? See
`aerosandbox.Opti.solve(verbose=)` syntax for more details.
Returns: A model in the form of a FittedModel object. Some things you can do:
>>> y = FittedModel(x) # evaluate the FittedModel at new x points
>>> FittedModel.parameters # directly examine the optimal values of the parameters that were found
>>> FittedModel.plot() # plot the fit
"""
super().__init__()
##### Prepare all inputs, check types/sizes.
### Flatten all inputs
def flatten(input):
return np.array(input).flatten()
try:
x_data = {k: flatten(v) for k, v in x_data.items()}
x_data_is_dict = True
except AttributeError: # If it's not a dict or dict-like, assume it's a 1D ndarray dataset
x_data = flatten(x_data)
x_data_is_dict = False
y_data = flatten(y_data)
n_datapoints = np.length(y_data)
### Handle weighting
if weights is None:
weights = np.ones(n_datapoints)
else:
weights = flatten(weights)
sum_weights = np.sum(weights)
if sum_weights <= 0:
raise ValueError("The weights must sum to a positive number!")
if np.any(weights < 0):
raise ValueError(
"No entries of the weights vector are allowed to be negative!")
weights = weights / np.sum(
weights) # Normalize weights so that they sum to 1.
### Check format of parameter_bounds input
if parameter_bounds is None:
parameter_bounds = {}
for param_name, v in parameter_bounds.items():
if param_name not in parameter_guesses.keys():
raise ValueError(
f"A parameter name (key = \"{param_name}\") in parameter_bounds was not found in parameter_guesses."
)
if not np.length(v) == 2:
raise ValueError(
"Every value in parameter_bounds must be a tuple in the format (lower_bound, upper_bound). "
"For one-sided bounds, use None for the unbounded side.")
### If putting residuals in logspace, check positivity
if put_residuals_in_logspace:
if not np.all(y_data > 0):
raise ValueError(
"You can't fit a model with residuals in logspace if y_data is not entirely positive!"
)
### Check dimensionality of inputs to fitting algorithm
relevant_inputs = {
"y_data": y_data,
"weights": weights,
}
try:
relevant_inputs.update(x_data)
except TypeError:
relevant_inputs.update({"x_data": x_data})
for key, value in relevant_inputs.items():
# Check that the length of the inputs are consistent
series_length = np.length(value)
if not series_length == n_datapoints:
raise ValueError(
f"The supplied data series \"{key}\" has length {series_length}, but y_data has length {n_datapoints}."
)
##### Formulate and solve the fitting optimization problem
### Initialize an optimization environment
opti = Opti()
### Initialize the parameters as optimization variables
params = {}
for param_name, param_initial_guess in parameter_guesses.items():
if param_name in parameter_bounds:
params[param_name] = opti.variable(
init_guess=param_initial_guess,
lower_bound=parameter_bounds[param_name][0],
upper_bound=parameter_bounds[param_name][1],
)
else:
params[param_name] = opti.variable(
init_guess=param_initial_guess, )
### Evaluate the model at the data points you're trying to fit
x_data_original = copy.deepcopy(
x_data
) # Make a copy of x_data so that you can determine if the model did in-place operations on x and tattle on the user.
try:
y_model = model(x_data, params) # Evaluate the model
except Exception:
raise Exception("""
There was an error when evaluating the model you supplied with the x_data you supplied.
Likely possible causes:
* Your model() does not have the call syntax model(x, p), where x is the x_data and p are parameters.
* Your model should take in p as a dict of parameters, but it does not.
* Your model assumes x is an array-like but you provided x_data as a dict, or vice versa.
See the docstring of FittedModel() if you have other usage questions or would like to see examples.
""")
try: ### If the model did in-place operations on x_data, throw an error
x_data_is_unchanged = np.all(x_data == x_data_original)
except ValueError:
x_data_is_unchanged = np.all([
x_series == x_series_original
for x_series, x_series_original in zip(x_data, x_data_original)
])
if not x_data_is_unchanged:
raise TypeError(
"model(x_data, parameter_guesses) did in-place operations on x, which is not allowed!"
)
if y_model is None: # Make sure that y_model actually returned something sensible
raise TypeError(
"model(x_data, parameter_guesses) returned None, when it should've returned a 1D ndarray."
)
### Compute how far off you are (error)
if not put_residuals_in_logspace:
error = y_model - y_data
else:
y_model = np.fmax(
y_model, 1e-300
) # Keep y_model very slightly always positive, so that log() doesn't NaN.
error = np.log(y_model) - np.log(y_data)
### Set up the optimization problem to minimize some norm(error), which looks different depending on the norm used:
if residual_norm_type.lower() == "l1": # Minimize the L1 norm
abs_error = opti.variable(init_guess=0, n_vars=np.length(
y_data)) # Make the abs() of each error entry an opt. var.
opti.subject_to([
abs_error >= error,
abs_error >= -error,
])
opti.minimize(np.sum(weights * abs_error))
elif residual_norm_type.lower() == "l2": # Minimize the L2 norm
opti.minimize(np.sum(weights * error**2))
elif residual_norm_type.lower(
) == "linf": # Minimize the L-infinity norm
linf_value = opti.variable(
init_guess=0
) # Make the value of the L-infinity norm an optimization variable
opti.subject_to([
linf_value >= weights * error, linf_value >= -weights * error
])
opti.minimize(linf_value)
else:
raise ValueError("Bad input for the 'residual_type' parameter.")
### Add in the constraints specified by fit_type, which force the model to stay above / below the data points.
if fit_type == "best":
pass
elif fit_type == "upper bound":
opti.subject_to(y_model >= y_data)
elif fit_type == "lower bound":
opti.subject_to(y_model <= y_data)
else:
raise ValueError("Bad input for the 'fit_type' parameter.")
### Solve
sol = opti.solve(verbose=verbose)
##### Construct a FittedModel
### Create a vector of solved parameters
params_solved = {}
for param_name in params:
try:
params_solved[param_name] = sol.value(params[param_name])
except:
params_solved[param_name] = np.NaN
### Store all the data and inputs
self.model = model
self.x_data = x_data
self.y_data = y_data
self.parameters = params_solved
self.parameter_guesses = parameter_guesses
self.parameter_bounds = parameter_bounds
self.residual_norm_type = residual_norm_type
self.fit_type = fit_type
self.weights = weights
self.put_residuals_in_logspace = put_residuals_in_logspace
