def update(self, experiment, seed, step):
"""updates the scatter plot.
This method is called by the model picker """
self.experiment = experiment
self.seed = seed
self.step = step
self.failures_indices = DataReader.get_episodes_with_outcome(
experiment, seed, step, 0)
features = DimensionalityReduction.get_model_failing_features(
experiment, seed, step)
costs = DataReader.get_model_states(experiment, seed, step)
print('costs shape', len(costs))
print('max failure indices', max(self.failures_indices))
res = self.DimensionalityReduction.transform(features)
colors = ['gray'] * res.shape[0]
opacities = [0.3] * res.shape[0]
classes = self.DimensionalityReduction.cluster(features)
category = bq.colorschemes.CATEGORY10[2:]
for f in range(len(features)):
if f - 1 < len(colors): # TODO: wtf?
if f < len(classes):
colors[f] = category[classes[f]]
else:
colors[f] = 'red'
opacities[f] = 0.8
self.scatter.x = res[:, 0]
self.scatter.y = res[:, 1]
self.scatter.colors = colors
self.scatter.opacity = opacities
def update2(self, experiment, seed, step):
# used in developement, not currently used
self.experiment = experiment
self.seed = seed
self.step = step
features = DimensionalityReduction.get_model_failing_features(
experiment, seed, step)
self.failures_indices = DataReader.get_episodes_with_outcome(
experiment, seed, step, 0)
failure_features = features[np.array(failures[:-1]) - 1]
res = self.DimensionalityReduction.transform(features)
colors = ['gray'] * res.shape[0]
opacities = [0.3] * res.shape[0]
classes = self.DimensionalityReduction.cluster(failure_features)
category = bq.colorschemes.CATEGORY20[2:]
for i, f in enumerate(failures):
if f - 1 < len(colors): # TODO: wtf?
if i < len(classes):
colors[f - 1] = category[classes[i]]
else:
colors[f - 1] = 'red'
opacities[f - 1] = 0.8
self.scatter.x = res[:, 0]
self.scatter.y = res[:, 1]
self.scatter.colors = colors
self.scatter.opacity = opacities
def get_model_failing_features(experiment, seed, checkpoint):
""" Get features for one model
This is used for dimensionality reduction, which is later used for
scatter plotting.
"""
speeds = DataReader.get_model_speeds(experiment, seed, checkpoint)
costs = DataReader.get_model_costs(experiment, seed, checkpoint)
states = DataReader.get_model_states(experiment, seed, checkpoint)
failing = DataReader.get_episodes_with_outcome(experiment, seed,
checkpoint, 0)
data = []
history_size = 10
for fail in failing:
columns_to_save = ["lane_cost", "pixel_proximity_cost"]
features = [speeds[fail - 1][-history_size:]]
for c in columns_to_save:
features.append(costs[fail - 1][c][-history_size:])
for i in range(len(features)):
features[i] = np.pad(
features[i],
(history_size - features[i].shape[0], 0),
"constant",
)
features = np.stack(features)
features = features.flatten()
length = costs[fail - 1]["collisions_per_frame"].shape[0]
features = np.append(
features,
[costs[fail - 1]["collisions_per_frame"][length - 1]])
features = np.append(
features, [costs[fail - 1]["arrived_to_dst"][length - 1]])
features = np.append(features, [states[fail - 1][length - 1][0]])
features = np.append(features, [states[fail - 1][length - 1][1]])
features = np.append(
features,
DataReader.get_last_gradient(experiment, seed, checkpoint,
fail),
)
data.append(features)
data = np.stack(data)
data = sklearn.preprocessing.scale(data)
return data
