def _fit_top_2_special(self, predictions):
### Fitting across top-2 points (done differently than top-k, k=2).
output_colourings = {}
for output, probability in predictions.items():
self.mapped_output(output)
colour = self.output_colour(output)
output_colourings[output] = (colour, probability)
interpolation_points = {}
for output, colour_probability in sorted(output_colourings.items(),
key=lambda item: item[1][1],
reverse=True)[:2]:
colour, probability = colour_probability
interpolation_points[colour] = (output, probability)
if len(interpolation_points) == 1:
return "rgb(%d, %d, %d)" % next(iter(interpolation_points.keys()))
else:
point_a, point_b = [item for item in interpolation_points.items()]
distance = geometry.distance(point_a[0], point_b[0])
# Not a typo: we want to invert their probabilities so that the most likely prediction gets the smallest distance, and visa versa.
# v v v v
pdist = mlbase.regmax({
point_a[1][0]: point_b[1][1],
point_b[1][0]: point_a[1][1]
})
fit = geometry.fit_proportion(
(point_a[0], point_b[0]),
(pdist[point_a[1][0]], pdist[point_b[1][0]]))
return "rgb(%d, %d, %d)" % tuple([round(i) for i in fit])
def test_fit_point2_vline(self):
reference_points = [[0, -2], [0, 4]]
target_distances = [2, 4]
expected = [0, 0]
point, t = fit_point(reference_points, target_distances)
self.assertTrue(
math.isclose(distance(point, expected), 0, abs_tol=0.0001), point)
self.assertLess(t, SMALL_MAX_T)
def test_fit_point1_d0(self):
reference_point = [5, 3]
target_distance = 0
point, t = fit_point([reference_point], [target_distance])
self.assertTrue(
math.isclose(distance(point, reference_point), 0, abs_tol=.0001),
point)
self.assertLess(t, SMALL_MAX_T)
def test_fit_point3(self):
reference_points = [[math.sqrt(8), math.sqrt(8)], [0, -4], [-4, 0]]
target_distances = [4, 4, 4]
expected = [0, 0]
point, t = fit_point(reference_points,
target_distances,
visualize=True)
self.assertTrue(
math.isclose(distance(point, expected), 0, abs_tol=0.0001), point)
self.assertLess(t, SMALL_MAX_T)
def test_fit_point2_hline(self):
reference_points = [[-1, 0], [5, 0]]
target_distances = [2, 4]
expected = [1, 0]
point, t = fit_point(reference_points,
target_distances,
visualize=True)
self.assertTrue(
math.isclose(distance(point, expected), 0, abs_tol=0.0001), point)
self.assertLess(t, SMALL_MAX_T)
def test_fit_proportion(self):
point_a = [0, 0]
point_b = [1, 1]
self.assertEqual(fit_proportion([point_a, point_b], [.5, .5]),
[.5, .5])
point = fit_proportion([point_a, point_b], [.2, .8])
self.assertTrue(
math.isclose(distance(point, [.2, .2]), 0, abs_tol=.0001), point)
point = fit_proportion([point_a, point_b], [.9, .1])
self.assertTrue(
math.isclose(distance(point, [.9, .9]), 0, abs_tol=.0001), point)
# Test 3d as well
point = fit_proportion([(0, 0, 0), (1, -1, 2)], [.9, .1])
self.assertTrue(
math.isclose(distance(point, [.9, -.9, 1.8]), 0, abs_tol=.0001),
point)
point = fit_proportion([(1, 2, 3), (3, 0, -3)], [.5, .5])
self.assertTrue(
math.isclose(distance(point, [2, 1, 0]), 0, abs_tol=.0001), point)
def find_closest(query_part, query_layer, query):
assert len(query) > 0, "empty query - would simply return everything!"
global activation_data
result = []
minimum_distance = None
maximum_distance = None
for candidate in activation_data:
if query_part == candidate.part and query_layer == candidate.layer:
sub_point = [candidate.point[axis] for axis, _ in query]
target_point = [target for axis, target in query]
distance = geometry.distance(sub_point, target_point)
result += [(distance, candidate)]
if minimum_distance is None or distance < minimum_distance:
minimum_distance = distance
if maximum_distance is None or distance > maximum_distance:
maximum_distance = distance
if len(result) == 0:
return result
q50 = maximum_distance * .5
q25 = maximum_distance * .25
q10 = maximum_distance * .1
print("distance stats: [%.4f, %.4f] q10: %.4f q25: %.4f q50: %.4f" %
(minimum_distance, maximum_distance, q10, q25, q50))
sorted_result = sorted(result)
cut10 = int(len(result) * 0.1)
cut25 = int(len(result) * 0.25)
cut50 = int(len(result) * 0.5)
histogram_q10, q10 = build_histogram(sorted_result[:cut10])
histogram_q25, q25 = build_histogram(sorted_result[:cut25])
histogram_q50, q50 = build_histogram(sorted_result[:cut50])
print("q10: %.4f q25: %.4f q50: %.4f" % (q10, q25, q50))
print("histograms\n q10: %s\n q25: %s\n q50: %s" %
(adjutant.dict_as_str(histogram_q10),
adjutant.dict_as_str(histogram_q25),
adjutant.dict_as_str(histogram_q50)))
return sorted_result, cut10, cut25, cut50
def _fit_top_k(self, predictions):
## Fitting across top-k points.
colour_probabilities = {}
for output, probability in predictions.items():
self.mapped_output(output)
colour = self.output_colour(output)
if colour not in colour_probabilities:
colour_probabilities[colour] = 0
colour_probabilities[colour] += probability
colour, probability = max(colour_probabilities.items(),
key=lambda item: item[1])
total = sum(colour_probabilities.values())
fit = geometry.fit_proportion((colour, [255, 255, 255]),
(total - probability, probability))
return "rgb(%d, %d, %d)" % tuple([round(i) for i in fit])
maximum_distance = None
for pair in itertools.combinations(
[colour for colour in colour_probabilities.keys()], 2):
distance = geometry.distance(pair[0], pair[1])
if maximum_distance is None or distance > maximum_distance:
maximum_distance = distance
inverted = [
item for item in mlbase.regmax(
{c: 1.0 - p
for c, p in colour_probabilities.items()}).items()
]
fit, _ = geometry.fit_point(
[item[0] for item in inverted],
[item[1] * maximum_distance for item in inverted])
return "rgb(%d, %d, %d)" % tuple([round(i) for i in fit])
def test_fit_point(self):
reference_points_1 = [[0, -2, 3], [1, 0, -6], [2, 4, 9], [3, 6, -12]]
#reference_points_1 = [[0, -2, 3], [1, 0, -6], [0, 4, 9], [3, 6, -12]]
reference_points_2 = [[3, -2, 3], [2, 0, -6], [1, 4, 9], [0, 6, -12]]
target_distances_1 = [5, 4, 3, 2]
target_distances_2 = [1, 2, 3, 4]
#target_distances_2 = [1, 1, 4, 3]
point_11, t_11 = fit_point(reference_points_1,
target_distances_1,
visualize=True)
self.assertTrue(all([not math.isnan(p) for p in point_11]), point_11)
self.assertLess(t_11, BIG_MAX_T)
point_12, t_12 = fit_point(reference_points_1,
target_distances_2,
visualize=True)
self.assertTrue(all([not math.isnan(p) for p in point_12]), point_12)
self.assertLess(t_12, BIG_MAX_T)
point_21, t_21 = fit_point(reference_points_2,
target_distances_1,
visualize=True)
self.assertTrue(all([not math.isnan(p) for p in point_21]), point_21)
self.assertLess(t_21, BIG_MAX_T)
point_22, t_22 = fit_point(reference_points_2,
target_distances_2,
visualize=True)
self.assertTrue(all([not math.isnan(p) for p in point_22]), point_22)
self.assertLess(t_22, BIG_MAX_T)
self.assertGreater(distance(point_11, point_12), 1)
self.assertGreater(distance(point_11, point_21), 1)
self.assertGreater(distance(point_11, point_22), 1)
self.assertGreater(distance(point_12, point_21), 1)
self.assertGreater(distance(point_12, point_22), .75)
self.assertGreater(distance(point_21, point_22), 1)
def test_distance(self):
self.assertEqual(distance([0, 1, 2], [0, 1, 2]), 0.0)
self.assertEqual(distance([0, 1, 2], [-1, 1, 4]),
math.sqrt(1**2 + 2**2))
}
#for word, probability in predictions[key].items():
# colour = self.colour_embeddings[word]
#
# if colour not in interpolation_points:
# interpolation_points[colour] = (word, probability)
# else:
# if interpolation_points[colour][1] < probability:
# interpolation_points[colour] = (word, probability)
#
if len(interpolation_points) == 1:
colours[key] = "rgb(%d, %d, %d)" % next(iter(interpolation_points.keys()))
else:
maximum_distance = None
for pair in itertools.combinations([colour for colour in interpolation_points.keys()], 2):
distance = geometry.distance(pair[0], pair[1])
if maximum_distance is None or distance > maximum_distance:
maximum_distance = distance
lowest_probability = min([p for w, p in interpolation_points.values()])
highest_probability = max([p for w, p in interpolation_points.values()])
maximum_domain = highest_probability + lowest_probability
prediction_distances = [(w, maximum_distance + (-p * maximum_distance / maximum_domain)) for w, p in interpolation_points.values()]
fit, _ = geometry.fit_point([colour_embeddings[item[0]] for item in prediction_distances], [item[1] for item in prediction_distances], epsilon=0.1, visualize=False)
print("rgb(%d, %d, %d)" % tuple([round(i) for i in fit]))
def measure(lstm, data_dir, kind, keys):
sequence_changes = {}
global_minimum = {key: (None, None, None) for key in keys}
global_maximum = {key: (None, None, None) for key in keys}
for j, xy in enumerate(data.stream_data(data_dir, kind)):
if j % 100 == 0:
logging.debug("At the %d instance." % (j))
sequence = tuple([item[0] for item in xy.x]) + (xy.y[-1][0], )
stepwise_rnn = lstm.stepwise(handle_unknown=True)
change_distances = {key: [] for key in keys}
previous_states = {}
minimum = {key: (None, None) for key in keys}
maximum = {key: (None, None) for key in keys}
for i, word_pos in enumerate(xy.x):
result, instruments = stepwise_rnn.step(word_pos[0],
rnn.LSTM_INSTRUMENTS)
for part, layer in lstm.part_layers():
key = lstm.encode_key(part, layer)
if key in keys:
current_state = instruments[part][layer]
if key in previous_states:
distance = geometry.distance(previous_states[key],
current_state)
else:
distance = geometry.hypotenuse(current_state)
change_distances[key] += [distance]
previous_states[key] = current_state
if minimum[key] == (None,
None) or distance < minimum[key][0]:
minimum[key] = (distance, i)
if maximum[key] == (None,
None) or distance > maximum[key][0]:
maximum[key] = (distance, i)
for key in keys:
if global_minimum[key] == (
None, None,
None) or minimum[key][0] < global_minimum[key][0]:
global_minimum[key] = minimum[key] + (sequence, )
# Only keeping track of the more notable sequence changes
sequence_changes[sequence] = change_distances
sequence_str, changes_str = stringify(
sequence, sequence_changes[sequence][key])
logging.debug(
"Noting minimum for %s of %.4f @%d:\n %s\n %s" %
(key, minimum[key][0], minimum[key][1], sequence_str,
changes_str))
if global_maximum[key] == (
None, None,
None) or maximum[key][0] > global_maximum[key][0]:
global_maximum[key] = maximum[key] + (sequence, )
# Only keeping track of the more notable sequence changes
sequence_changes[sequence] = change_distances
sequence_str, changes_str = stringify(
sequence, sequence_changes[sequence][key])
logging.debug(
"Noting maximum for %s of %.4f @%d:\n %s\n %s" %
(key, maximum[key][0], maximum[key][1], sequence_str,
changes_str))
return global_minimum, global_maximum, sequence_changes