def inference(self, X, w):
"""Run Viterbi inference.
This methods is a wrapper that converts the CRF weights into
different arrays of scores that represent transition and emission.
Then this method can call the general purpose Viterbi code in
viterbi.py to compute the best label sequence.
This function just returns the best sequence, y.
"""
from viterbi import run_viterbi
L = self.num_classes
N = len(X)
start_scores = np.zeros(L)
end_scores = np.zeros(L)
trans_scores = np.zeros((L, L))
emission_scores = np.zeros((N, L))
# fill the above arrays for the weight vector
for j in xrange(L):
start_scores[j] = w[0, self.get_start_trans_idx(j)]
end_scores[j] = w[0, self.get_end_trans_idx(j)]
# transition
for k in xrange(L):
trans_scores[j][k] = w[0, self.get_trans_idx(j, k)]
# emission
for i in xrange(N):
score = 0.0
for fidx in X[i]:
score += w[0, self.get_ftr_idx(fidx, j)]
emission_scores[i][j] = score
# now run the viterbi code!
(score, yhat) = run_viterbi(emission_scores, trans_scores,
start_scores, end_scores)
return yhat
def task(self):
data_filename = "robot_no_momentum.data"
hmm, d = train_hmm_from_data(data_filename)
err_full = run_viterbi(hmm, d)
data_filename_m = "robot_with_momentum.data"
hmm_m, d_m = train_hmm_from_data(data_filename_m)
err_full_m = run_viterbi(hmm_m, d_m)
listNames = ["Without momentum", "With momentum"]
listData = [1 - err_full, 1 - err_full_m]
chart = {"chart": {"defaultSeriesType": "column"},
"xAxis": {"categories": listNames},
"yAxis": {"title": {"text": "Fraction Correct"}},
"title": {"text": "HMM performance on"
" inferring robot location."},
"series": [{"name": "Test set performance",
"data": listData}]}
return chart
def main(): ''' Load test data ''' # Input: Testing, generate new windows, oversampling, viterbi training DATA_TYPE = "testing" GENERATE_NEW_WINDOWS = True OVERSAMPLING = False VITERBI = False data_set = get_data_set(DATA_TYPE, GENERATE_NEW_WINDOWS, OVERSAMPLING, VITERBI) ''' Create network ''' cnn = Convolutional_Neural_Network() cnn.set_data_set(data_set) cnn.load_model() '''''' actual = data_set._labels cnn_result = cnn.get_predictions() np.savetxt(V.VITERBI_PREDICTION_PATH_TESTING, cnn_result, delimiter=",") cnn_result = pd.read_csv(V.VITERBI_PREDICTION_PATH_TESTING, header=None, sep='\,',engine='python').as_matrix() viterbi_result = run_viterbi() np.savetxt(V.VITERBI_RESULT_TESTING, viterbi_result, delimiter=",") viterbi_result = pd.read_csv(V.VITERBI_RESULT_TESTING, header=None, sep='\,',engine='python').as_matrix() ''' Add results in array with actual label''' result = np.zeros((len(cnn_result), 3)) for i in range(0,len(cnn_result)): a = np.argmax(actual[i]) c = np.argmax(cnn_result[i]) v = viterbi_result[i]-1 result[i] = [a,c,v] # Remove activities labelled as -100 - activites such as shuffling, transition ... See data.py boolean_actual = np.invert(actual[:,0] == -100).T result = result[boolean_actual] np.savetxt(V.PREDICTION_RESULT_TESTING, result, delimiter=",") result = pd.read_csv(V.PREDICTION_RESULT_TESTING, header=None, sep='\,',engine='python').as_matrix() produce_statistics_json(result) visualize(result)
def inference(self, X, w):
"""Run Viterbi inference.
This methods is a wrapper that converts the CRF weights into
different arrays of scores that represent transition and emission.
Then this method can call the general purpose Viterbi code in
viterbi.py to compute the best label sequence.
This function just returns the best sequence, y.
"""
from viterbi import run_viterbi
L = self.num_classes
start_scores, end_scores, trans_scores, emission_scores = self.load_weights(
w, L, X)
# now run the viterbi code!
score, yhat = run_viterbi(emission_scores, trans_scores, start_scores,
end_scores)
return yhat
def main(): ''' Load test data ''' # Input: Testing, generate new windows, oversampling, viterbi training DATA_TYPE = "predicting" GENERATE_NEW_WINDOWS = True OVERSAMPLING = False VITERBI = False data_set = get_data_set(DATA_TYPE, GENERATE_NEW_WINDOWS, OVERSAMPLING, VITERBI) ''' Create network ''' cnn = Convolutional_Neural_Network() cnn.set_data_set(data_set) cnn.load_model() '''''' cnn_result = cnn.get_predictions() viterbi_result = run_viterbi() print 'Prediction saved at path', V.VITERBI_RESULT_PREDICTING
def run_viterbi_test():
"""A simple tester for Viterbi algorithm.
This function generates a bunch of random emission and transition scores,
and computes the best sequence by performing a brute force search over all
possible sequences and scoring them. It then runs Viterbi code to see what
is the score and sequence returned by it.
Compares both the best sequence and its score to make sure Viterbi is correct.
"""
from viterbi import run_viterbi
from numpy import random
import numpy as np
from itertools import product
maxN = 7 # maximum length of a sentence (min is 1)
maxL = 4 # maximum number of labels (min is 2)
num_tests = 1000 # number of sentences to generate
random.seed(0)
tolerance = 1e-5 # how close do the scores have to be?
emission_var = 1.0 # variance of the gaussian generating emission scores
trans_var = 1.0 # variance of the gaussian generating transition scores
passed_y = 0 # how many times the correct sequence was predicted
passed_s = 0 # how many times the correct score was returned
for t in xrange(num_tests):
N = random.randint(1, maxN + 1)
L = random.randint(2, maxL + 1)
# Generate the scores
emission_scores = random.normal(0.0, emission_var, (N, L))
trans_scores = random.normal(0.0, trans_var, (L, L))
start_scores = random.normal(0.0, trans_var, L)
end_scores = random.normal(0.0, trans_var, L)
# run viterbi
(viterbi_s, viterbi_y) = run_viterbi(emission_scores, trans_scores,
start_scores, end_scores)
# print ("Viterbi", viterbi_s, viterbi_y)
# compute the best sequence and score
best_y = []
best_s = -np.inf
for y in product(range(L), repeat=N): # all possible ys
# compute its score
score = 0.0
score += start_scores[y[0]]
# print(y,'y')
# print(y)
# print(score,'1')
for i in xrange(N - 1):
score += trans_scores[y[i], y[i + 1]]
# print(score,'2',trans_scores[y[i], y[i+1]])
score += emission_scores[i, y[i]]
# print(score,'3',emission_scores[i,y[i]])
score += emission_scores[N - 1, y[N - 1]]
# print(score,'4',emission_scores[N-1,y[N-1]])
score += end_scores[y[N - 1]]
# print(score,'5',end_scores[y[N-1]])
# if 8.38435628640<score<8.38435628650:
# break
# if y[0]==1 and y[1]==2 and y[2]==2 and y[3]==0 and y[4]==0:
# break
# update the best
if score > best_s:
best_s = score
best_y = list(y)
# break
# print ("Brute", best_s, best_y)
# mismatch if any label prediction doesn't match
match_y = True
for i in xrange(len(best_y)):
if viterbi_y[i] != best_y[i]:
match_y = False
if match_y: passed_y += 1
# the scores should also be very close
if abs(viterbi_s - best_s) < tolerance:
passed_s += 1
print "Passed(y)", passed_y * 100.0 / num_tests
print "Passed(s)", passed_s * 100.0 / num_tests
assert passed_y == num_tests
assert passed_s == num_tests
def viterbi_tags(self, logits: torch.Tensor,
mask: torch.Tensor) -> List[Tuple[List[int], float]]:
"""
Uses viterbi algorithm to find most likely tags for the given inputs.
If constraints are applied, disallows all other transitions.
"""
_, max_seq_length, num_tags = logits.size()
# Get the tensors out of the variables
logits, mask = logits.data, mask.data
# Augment transitions matrix with start and end transitions
start_tag = num_tags
end_tag = num_tags + 1
transitions = torch.Tensor(num_tags + 2, num_tags + 2).fill_(-10000.)
# Apply transition constraints
constrained_transitions = (
self.transitions * self._constraint_mask[:num_tags, :num_tags] +
-10000.0 * (1 - self._constraint_mask[:num_tags, :num_tags]))
transitions[:num_tags, :num_tags] = constrained_transitions.data
if self.include_start_end_transitions:
transitions[start_tag, :num_tags] = (
self.start_transitions.detach() *
self._constraint_mask[start_tag, :num_tags].data + -10000.0 *
(1 - self._constraint_mask[start_tag, :num_tags].detach()))
transitions[:num_tags, end_tag] = (
self.end_transitions.detach() *
self._constraint_mask[:num_tags, end_tag].data + -10000.0 *
(1 - self._constraint_mask[:num_tags, end_tag].detach()))
else:
transitions[start_tag, :num_tags] = (
-10000.0 *
(1 - self._constraint_mask[start_tag, :num_tags].detach()))
transitions[:num_tags, end_tag] = -10000.0 * (
1 - self._constraint_mask[:num_tags, end_tag].detach())
transitions = transitions.cpu().numpy()
best_paths = []
# Pad the max sequence length by 2 to account for start_tag + end_tag.
tag_sequence = torch.Tensor(max_seq_length + 2, num_tags + 2)
for prediction, prediction_mask in zip(logits, mask):
sequence_length = torch.sum(prediction_mask)
# Start with everything totally unlikely
tag_sequence.fill_(-10000.)
# At timestep 0 we must have the START_TAG
tag_sequence[0, start_tag] = 0.
# At steps 1, ..., sequence_length we just use the incoming prediction
tag_sequence[1:(sequence_length +
1), :num_tags] = prediction[:sequence_length]
# And at the last timestep we must have the END_TAG
tag_sequence[sequence_length + 1, end_tag] = 0.
# We pass the tags and the transitions to ``run_viterbi``.
target_tag_sequence = tag_sequence[:(sequence_length +
2)].cpu().numpy()
viterbi_score, viterbi_path =\
viterbi.run_viterbi(target_tag_sequence[1:-1, :num_tags], transitions[:num_tags, :num_tags], transitions[start_tag, :num_tags], transitions[:num_tags, end_tag])
best_paths.append((viterbi_path, viterbi_score))
return best_paths
def run_viterbi_test():
"""A simple tester for Viterbi algorithm.
This function generates a bunch of random emission and transition scores,
and computes the best sequence by performing a brute force search over all
possible sequences and scoring them. It then runs Viterbi code to see what
is the score and sequence returned by it.
Compares both the best sequence and its score to make sure Viterbi is correct.
"""
from viterbi import run_viterbi
from numpy import random
from itertools import product
maxN = 7 # maximum length of a sentence (min is 1)
maxL = 4 # maximum number of labels (min is 2)
num_tests = 1 # number of sentences to generate
random.seed(0)
tolerance = 1e-5 # how close do the scores have to be?
emission_var = 1.0 # variance of the gaussian generating emission scores
trans_var = 1.0 # variance of the gaussian generating transition scores
passed_y = 0 # how many times the correct sequence was predicted
passed_s = 0 # how many times the correct score was returned
for t in xrange(num_tests):
N = 2
L = 3
# Generate the scores
# emission_scores = random.normal(0.0, emission_var, (N,L))
# trans_scores = random.normal(0.0, trans_var, (L,L))
# start_scores = random.normal(0.0, trans_var, L)
# end_scores = random.normal(0.0, trans_var, L)
#print start_scores
emission_scores = np.array([[0.1, 0.3, 0.25], [0.2, 0.45, 0.31]])
trans_scores = np.array([[0.3, 0.2, 0.5], [0.12, 0.4, 0.3],
[0.1, 0.6, 0.5]])
start_scores = np.array([0.2, 0.5, 0.7])
end_scores = np.array([0.5, 0.4, 0.1])
# run viterbi
(viterbi_s, viterbi_y) = run_viterbi(emission_scores, trans_scores,
start_scores, end_scores)
print "Viterbi", viterbi_s, viterbi_y
# compute the best sequence and score
best_y = []
best_s = -np.inf
for y in product(range(L), repeat=N): # all possible ys
# compute its score
score = 0.0
score += start_scores[y[0]]
for i in xrange(N - 1):
score += trans_scores[y[i], y[i + 1]]
score += emission_scores[i, y[i]]
score += emission_scores[N - 1, y[N - 1]]
score += end_scores[y[N - 1]]
# update the best
if score > best_s:
best_s = score
best_y = list(y)
print "Brute", best_s, best_y
# mismatch if any label prediction doesn't match
match_y = True
for i in xrange(len(best_y)):
if viterbi_y[i] != best_y[i]:
match_y = False
if match_y: passed_y += 1
# the scores should also be very close
print "scores: "
print viterbi_s, best_s
if abs(viterbi_s - best_s) < tolerance:
passed_s += 1
print "Passed(y)", passed_y * 100.0 / num_tests
print "Passed(s)", passed_s * 100.0 / num_tests
assert passed_y == num_tests
assert passed_s == num_tests
def test_small_robot_dataset(self):
data_filename = "robot_small.data"
data_filename = normalize_filename(data_filename)
hmm, d = train_hmm_from_data(data_filename)
err_full = run_viterbi(hmm, d, True)
self.assertAlmostEqual(err_full, 2.0 / 9)
