get_displacements (x)
Returns the displacements of trajectory x [dim, length].
split_tensor (t, indices)
Splits input tensor t according to indices in the first dimension.
lengths_from_cps (cps, length=200)
Returns segment lengths determined by cps and a total length length.
fit_segments (pred, pen=1.0, return_cps=False, kernel='linear', min_size=2, jump=1, params=None)
Fit piecewise constant segments to input signal pred.
fit_segments is mainly intended to process predictions of continuous values. However, the following functinos are mainly intended to post-process discrete predictions. Notably, post_process_prediction takes a prediction of discrete categories over a trajectory and extracts the most likely changepoints and segments, minimizing the impact of spurious mistakes along the predicted segment (see the example below).
post_process_prediction (pred, n_change_points=1)
Segmentation prediction post-processing to find change points and classes.
abundance (val, t)
Abundance of value val in tensor t.
majority_vote (t)
Returns majoritary value from t.
get_split_classes (splits)
Returns majority class of each split.
change_points_from_splits (splits)
Returns change point position from split tensor.
get_splits (t)
Splits tensor t into chunks with the same value.
find_change_points (t)
Finds points in tensor t where the value changes.
mean_relative_error (pred, true, base=10)
Mean relative error assuming pred and true in log_base.
mean_absolute_error (pred, true)
Mean absolute error between pred and true.
jaccard_index (true_positive, false_positive, false_negative)
Computes the Jaccard index a.k.a. Tanimoto index.
assign_changepoints (true, pred)
Matches predicted and true changepoints solving a linear sum assignment problem.
evaluate_cp_prediction (true, pred, changepoint_threshold=5)
Evaluates the change point prediction.
Since the changepoint detection algorithm can provide an arbitrary number of change points, we solve a linear sum assignment problem to perform the matching between the ground truth and the predicted changepoints.
Then, we consider a valid prediction, i.e., a true positive (TP), those changepoints that lie within a trheshold of their corresponding ground truth. This way, all the predicted change points that are not TP are false positives (FP). Finally, the ground truth change points that do not have a predicted counterpart within the threshold are false negatives (FN).
To evaluate the change point detection, we use the Jaccard index, which is a function of the TP, FP and FN: \[JI = \frac{TP}{TP + FP + FN}\,.\]
validate_andi_3_alpha (m, dim=1, task=3, bs=128, pct=1, path=None)
Validates model on the AnDi test set for task 3 (segmentation) predicting anomalous exponents.
validate_andi_3_models (m, dim=1, task=3, bs=128, pct=1, path=None)
Validates model on the AnDi test set for task 3 (segmentation) predicting diffusion models.
validate_andi_1 (m, dim=1, bs=1, pct=1, task=1, path=None)
Validates model on the AnDi test set for task 1 (anomalous exponent).
eval_andi_metrics (dls, model)
Evaluates model in validation set in order to obtain AnDi challenge metrics.
Here, we define colors and colormaps for our plots.
color_order = ['blue', 'orange', 'yellow', 'purple', 'green']
color_dict = {
'blue': {'dark': (0.2745098, 0.4, 0.6),
'medium': (0.39607843, 0.5254902, 0.71764706),
'light': (0.65098039, 0.79215686, 0.94117647)},
'orange': {'dark': (0.71764706, 0.36470588, 0.24313725),
'medium': (0.88627451, 0.4627451, 0.34901961),
'light': (1.0, 0.63921569, 0.44705882)},
'yellow': {'dark': (0.85882353, 0.58431373, 0.18039216),
'medium': (0.89803922, 0.68235294, 0.39607843),
'light': (0.96470588, 0.84705882, 0.52941176)},
'purple': {'dark': (0.6627451, 0.16078431, 0.30980392),
'medium': (0.7372549, 0.39607843, 0.55294118),
'light': (0.89019608, 0.38823529, 0.52941176)},
'green': {'dark': (0.22352941, 0.46666667, 0.4549019607843137),
'medium': (0.29803922, 0.60784314, 0.58431373),
'light': (0.50980392, 0.76862745, 0.76470588)}
}
colors = [color_dict[k]['medium'] for k in color_order]
colors_light = [color_dict[k]['light'] for k in color_order]
colors_dark = [color_dict[k]['dark'] for k in color_order]
cmap_hist1 = clr.LinearSegmentedColormap.from_list(
'custom cm', ['w',
color_dict['blue']['light'],
color_dict['blue']['dark']],
N=256
)
cmap_hist2 = clr.LinearSegmentedColormap.from_list(
'custom cm', ['w',
color_dict['orange']['light'],
color_dict['orange']['dark']],
N=256
)
cmap_points = clr.LinearSegmentedColormap.from_list(
'custom cm', [color_dict['yellow']['light'],
color_dict['purple']['light'],
color_dict['blue']['medium']],
N=256
)
fig_size = 4
linewidth = 2
alpha_grid = 0.2
scatter_size = 12
D_units = "($\mu$m$^2$/s)"