def _plot_deviation(self,
true_values: np.ndarray,
estim_values: np.ndarray,
save=None,
show=True,
title=None,
return_axs=False):
"""
Plot estimated coefficients against reference (true) coefficients.
:param true_values:
:param estim_values:
:param save: Path+file name stem to save plots to.
File will be save+"_genes.png". Does not save if save is None.
:param show: Whether to display plot.
:param return_axs: Whether to return axis objects.
:return: Matplotlib axis objects.
"""
import seaborn as sns
import matplotlib.pyplot as plt
if isinstance(true_values, dask.array.core.Array):
true_values = true_values.compute()
if isinstance(estim_values, dask.array.core.Array):
estim_values = estim_values.compute()
plt.ioff()
n_par = true_values.shape[0]
summary_fit = pd.concat([
pd.DataFrame({
"deviation":
estim_values[i, :] - true_values[i, :],
"coefficient":
pd.Series(
["coef_" + str(i) for x in range(estim_values.shape[1])],
dtype="category")
}) for i in range(n_par)
])
summary_fit['coefficient'] = summary_fit['coefficient'].astype(
"category")
fig, ax = plt.subplots()
sns.violinplot(x=summary_fit["coefficient"],
y=summary_fit["deviation"],
ax=ax)
if title is not None:
ax.set_title(title)
# Save, show and return figure.
if save is not None:
plt.savefig(save + '_deviation_violin.png')
if show:
plt.show()
plt.close(fig)
plt.ion()
if return_axs:
return ax
else:
return
def groupwise_solve_lm(dmat, apply_fun: callable, constraints: np.ndarray):
r"""
Solve GLMs by estimating the distribution parameters of each unique group of observations independently and
solving then for the design matrix `dmat`.
Idea:
$$
\theta &= f(x) \\
\Rightarrow f^{-1}(\theta) &= x \\
&= (D \cdot D^{+}) \cdot x \\
&= D \cdot (D^{+} \cdot x) \\
&= D \cdot x' = f^{-1}(\theta)
$$
:param dmat: design matrix which should be solved for
:param apply_fun: some callable function taking one xr.DataArray argument.
Should compute a group-wise parameter solution.
Example method calculating group-wise means:
::
def apply_fun(grouping):
groupwise_means = data.groupby(grouping).mean(dim="observations").values
return np.log(groupwise_means)
The `data` argument provided to `apply_fun` is the same xr.DataArray provided to this
:param constraints: tensor (all parameters x dependent parameters)
Tensor that encodes how complete parameter set which includes dependent
parameters arises from indepedent parameters: all = <constraints, indep>.
This form of constraints is used in vector generalized linear models (VGLMs).
:return: tuple of (apply_fun(grouping), x_prime, rmsd, rank, s) where x_prime is the parameter matrix solved for
`dmat`.
"""
# Get unqiue rows of design matrix and vector with group assignments:
if isinstance(dmat, dask.array.core.Array
): # axis argument not supported by dask in .unique()
unique_design, inverse_idx = np.unique(dmat.compute(),
axis=0,
return_inverse=True)
unique_design = dask.array.from_array(unique_design,
chunks=unique_design.shape)
else:
unique_design, inverse_idx = np.unique(dmat,
axis=0,
return_inverse=True)
if unique_design.shape[0] > 500:
raise ValueError(
"large least-square problem in init, likely defined a numeric predictor as categorical"
)
full_rank = constraints.shape[1]
if isinstance(dmat,
dask.array.core.Array): # matrix_rank not supported by dask
rank = np.linalg.matrix_rank(
np.matmul(unique_design.compute(), constraints.compute()))
else:
rank = np.linalg.matrix_rank(np.matmul(unique_design, constraints))
if full_rank > rank:
logger.error("model is not full rank!")
# Get group-wise means in linker space based on group assignments
# based on unique rows of design matrix:
params = apply_fun(inverse_idx)
# Use least-squares solver to compute model parameterization
# accounting for dependent parameters, ie. degrees of freedom
# of the model which appear as groups in the design matrix
# and are not accounted for by parameters but which are
# accounted for by constraints:
# <X, <theta, H> = means -> <X, theta>, H> = means -> lstsqs for theta
# (This is faster and more accurate than using matrix inversion.)
logger.debug(" ** Solve lstsq problem")
if np.any(np.isnan(params)):
raise Warning(
"entries of params were nan which will throw error in lstsq")
x_prime, rmsd, rank, s = np.linalg.lstsq(
np.matmul(unique_design, constraints), params)
return params, x_prime, rmsd, rank, s
def _plot_coef_vs_ref(self,
true_values: np.ndarray,
estim_values: np.ndarray,
size=1,
log=False,
save=None,
show=True,
ncols=5,
row_gap=0.3,
col_gap=0.25,
title=None,
return_axs=False):
"""
Plot estimated coefficients against reference (true) coefficients.
:param true_values:
:param estim_values:
:param size: Point size.
:param save: Path+file name stem to save plots to.
File will be save+"_genes.png". Does not save if save is None.
:param show: Whether to display plot.
:param ncols: Number of columns in plot grid if multiple genes are plotted.
:param row_gap: Vertical gap between panel rows relative to panel height.
:param col_gap: Horizontal gap between panel columns relative to panel width.
:param title: Plot title.
:param return_axs: Whether to return axis objects.
:return: Matplotlib axis objects.
"""
import seaborn as sns
import matplotlib.pyplot as plt
from matplotlib import gridspec
from matplotlib import rcParams
if isinstance(true_values, dask.array.core.Array):
true_values = true_values.compute()
if isinstance(estim_values, dask.array.core.Array):
estim_values = estim_values.compute()
plt.ioff()
n_par = true_values.shape[0]
ncols = ncols if n_par > ncols else n_par
nrows = n_par // ncols + (n_par - (n_par // ncols) * ncols)
gs = gridspec.GridSpec(nrows=nrows,
ncols=ncols,
hspace=row_gap,
wspace=col_gap)
fig = plt.figure(figsize=(
ncols * rcParams['figure.figsize'][0], # width in inches
nrows * rcParams['figure.figsize'][1] *
(1 + row_gap) # height in inches
))
if title is None:
title = "parameter"
# Build axis objects in loop.
axs = []
for i in range(n_par):
ax = plt.subplot(gs[i])
axs.append(ax)
x = true_values[i, :]
y = estim_values[i, :]
if log:
x = np.log(x + 1)
y = np.log(y + 1)
sns.scatterplot(x=x, y=y, size=size, ax=ax, legend=False)
sns.lineplot(x=np.array([
np.min([np.min(x), np.min(y)]),
np.max([np.max(x), np.max(y)])
]),
y=np.array([
np.min([np.min(x), np.min(y)]),
np.max([np.max(x), np.max(y)])
]),
ax=ax)
title_i = title + "_" + str(i)
# Add correlation into title:
title_i = title_i + " (R=" + str(
np.round(np.corrcoef(x, y)[0, 1], 3)) + ")"
ax.set_title(title_i)
ax.set_xlabel("true parameter")
ax.set_ylabel("estimated parameter")
# Save, show and return figure.
if save is not None:
plt.savefig(save + '_parameter_scatter.png')
if show:
plt.show()
plt.close(fig)
plt.ion()
if return_axs:
return axs
else:
return