import numpy as np
from matplotlib import cm
from mindpark.stats.figure import Figure
from mindpark.plot import Scalar, Histogram
from mpl_toolkits.axes_grid1 import make_axes_locatable
class Metrics(Figure):
"""
Generate figures showing multiple metrics as plots. Choose the plot type
automatically, based on the data.
"""
def __init__(self):
self._plot_scalar = Scalar()
self._plot_counts = Histogram()
self._plot_distribution = Histogram(normalize=True)
def __call__(self, metrics, title, filepath):
self._validate_input(metrics)
fig, ax = self._create_subplots(2, len(metrics))
fig.suptitle(title, fontsize=16)
names, metrics = zip(*metrics)
self._label_columns(ax, names)
self._label_rows(ax, ['Training', 'Evaluation'])
for index, metric in enumerate(metrics):
train = metric[metric.training == 1]
test = metric[metric.training == 0]
test.epoch -= 1
self._process_metric(ax[0, index], train)
self._process_metric(ax[1, index], test)
self._save(fig, filepath)
def _validate_input(self, metrics):
assert all(len(x) == 2 for x in metrics)
for name, metric in metrics:
assert isinstance(name, str)
assert all(isinstance(x, np.ndarray) for x in metric.values())
def _process_metric(self, ax, metric):
if not metric.data.size:
ax.tick_params(colors=(0, 0, 0, 0))
ax.set_axis_bgcolor(cm.get_cmap('viridis')(0))
divider = make_axes_locatable(ax)
divider.append_axes('right', size='7%', pad=0.1).axis('off')
return
domain = self._domain(metric)
categorical = self._is_categorical(metric.data)
if metric.data.shape[1] == 1 and not categorical:
self._plot_scalar(ax, domain, metric.data[:, 0])
elif metric.data.shape[1] == 1:
indices = metric.data[:, 0].astype(int)
min_, max_ = indices.min(), indices.max()
count = np.eye(max_ - min_ + 1)[indices - min_]
self._plot_distribution(ax, domain, count)
elif metric.data.shape[1] > 1:
self._plot_counts(ax, domain, metric.data)
def _is_categorical(self, data):
if not np.allclose(data, data.astype(int)):
return False
if np.unique(data.astype(int)).size > 8:
return False
return True
