def test_viterbi():
site = [0,1,2,3,0,1,2,3]
background = lambda n:[random.choice(range(4)) for i in range(n)]
obs = (site + background(1000) +
site + background(1000) +
site)
states = [0,1,2,3,4,5,6,7,8] # 0 is off, 1-8 are bs positions
start_p = [0.99,0.01,0,0,0,0,0,0,0]
trans_p = [[0.99,0.01,0,0,0,0,0,0,0],
[0,0,1,0,0,0,0,0,0],
[0,0,0,1,0,0,0,0,0],
[0,0,0,0,1,0,0,0,0],
[0,0,0,0,0,1,0,0,0],
[0,0,0,0,0,0,1,0,0],
[0,0,0,0,0,0,0,1,0],
[0,0,0,0,0,0,0,0,1],
[0.99,0.01,0,0,0,0,0,0,0],
]
emit_p = [[0.25,0.25,0.25,0.25],
[1,0,0,0],
[0,1,0,0],
[0,0,1,0],
[0,0,0,1],
[1,0,0,0],
[0,1,0,0],
[0,0,1,0],
[0,0,0,1]
]
return viterbi(obs, states, start_p, trans_p, emit_p)
def tagger(testFile,taggedFile,n):
dicOfWordTag,dicOfTag,dicOfTags=read_corpus("corpus/CoNLL2009-ST-English-train-pos.txt")
f=open(testFile,'r')
fw=open(taggedFile,'w')
totalNumber=0
noSolution=0
words=[]
lines=[]
for line in f:
line=line.strip()
if line=='':
if len(words)>n:
for row in lines:
for item in row:
print >> fw,item,
words=[]
lines=[]
continue
prob,path=viterbi(words,dicOfWordTag,dicOfTag,dicOfTags)
if prob==0:
#print >>fw,"no solution"
noSolution+=1
for row in lines:
for item in row:
print >> fw,item,
print >> fw
print >>fw
words=[]
lines=[]
continue
i=0
while i<len(path):
for item in lines[i]:
print >> fw,item,
print >> fw,path[i],path[i]
i+=1
print >> fw
words=[]
lines=[]
else:
line=line.split()
words.append(line[1])
lines.append(line)
totalNumber+=1
print 1-float(noSolution)/totalNumber
def main():
"""main function
"""
n = 2 # Bigram HMM
args = parse_arguments()
treebank = TaggedCorpusReader(
os.path.split(args.train_f)[0],
os.path.split(args.train_f)[1])
observation_space = [item[0] for item in treebank.sents()] # all words
state_space = [item[1] for item in treebank.sents()] # all pos tags
words = dict.fromkeys(observation_space)
tags = dict.fromkeys(state_space)
# HMM parameter estimation- initial, transition and emission probablity
start = time.time()
init_p = [item[1] for item in comp_initial(tags, treebank)]
trans_p = comp_transition(n, tags, state_space)
emission_p = comp_emission(words,
tags,
state_space,
treebank,
smoothing=args.smoothing)
end = time.time()
print("Runtime (training): %.3f s" % (end - start))
# Test your HMM-trained model
treebank = TaggedCorpusReader(
os.path.split(args.eval_f)[0],
os.path.split(args.eval_f)[1])
viterbi_tags = []
start = time.time()
for sentence in treebank.paras():
test_words = [item[0] for item in sentence]
O, S, Y, pi, A, B = pre_process(words, tags, test_words, init_p,
trans_p, emission_p)
# Computes Viterbi's most likely tags
if args.log_prob:
X = viterbi_log(O, S, Y, pi, A, B)
else:
X = viterbi(O, S, Y, pi, A, B)
viterbi_tags.append(X)
end = time.time()
print("Runtime (viterbi): %.3f s" % (end - start))
output_path = "./" + "de-tagger.tt"
post_processing(viterbi_tags, args.test_f, output_path)
def evaluate(testFile,n):
dicOfWordTag,dicOfTag,dicOfTags=read_corpus("corpus/CoNLL2009-ST-English-train-pos.txt")
f=open(testFile,'r')
rightNumber=0
denominator=0
totalNumber=0
noSolution=0
words=[]
right=[]
for line in f:
line=line.strip()
if line=='':
if len(words)>n:
#print "length> ",n
words=[]
right=[]
continue
prob,path=viterbi(words,dicOfWordTag,dicOfTag,dicOfTags)
if prob==0:
words=[]
right=[]
noSolution+=1
continue
i=0
while i<len(path):
if path[i]==right[i]:
rightNumber+=1
denominator+=1
i+=1
words=[]
right=[]
else:
line=line.split()
words.append(line[1])
right.append(line[4])
totalNumber+=1
precision=float(rightNumber)/denominator
recall=1-float(denominator)/totalNumber
return precision,recall
def __cut(sen):
prob, pos_list = viterbi(sen,
status,
start_P,
trans_P,
emit_P,
end_status='ES')
flag = 0
for num, pos in enumerate(pos_list):
if pos == 'E':
word = sen[flag:num + 1]
flag = num + 1
yield word
elif pos == 'S':
word = sen[flag:num + 1]
flag = num + 1
yield word
def main():
outFileName = sys.argv[1]
n = int(sys.argv[2])
outFile = open(outFileName,'w')
data = genDieNumber(n)
prediction = viterbi(data[0])
for i in range(0,len(prediction)//50):
outFile.write("Rolls ")
for j in range(0,50):
outFile.write(str(data[0][j]))
outFile.write("\n")
outFile.write("Die ")
for j in range(0,50):
outFile.write(str(data[1][j]))
outFile.write("\n")
outFile.write("Viterbi ")
for j in range(0,50):
outFile.write(str(prediction[j]))
outFile.write("\n")
def test(trainfile,testfile):
word_dict,pos_dict,tag_list,emission_probabilities,transition_probabilities,prefixp,suffixp = learnFromTraining(trainfile)
file = open(testfile,"r")
unknown = count_unknown(file)
file = open(testfile,"r")
words = []
for line in file:
words.append(line.rstrip())
pos=[]
new_sentence=[]
for i in range(0,len(words)):
if words[i]=='' or words[i]=='\n':
v,p = viterbi(new_sentence,emission_probabilities,transition_probabilities,pos_dict,prefixp,suffixp,unknown)
tags = bestPath(new_sentence,v,p,transition_probabilities.keys())
tags.append('')
tags=tags[1:]
pos.extend(tags)
new_sentence=[]
else:
new_sentence.append(words[i])
return pos,words
def evaluate():
seqLen = []
avg_accuracy = []
avg_MCC = []
for n in range(1000, 10100, 100):
for i in range(0, 10):
accuracy = []
MCC = []
data = genDieNumber(n)
true_result = data[1]
prediction = viterbi(data[0])
TP = 0
TN = 0
FP = 0
FN = 0
for i in range(0, len(data[0])):
if true_result[i] == "F":
if prediction[i] == "F":
TP += 1
else:
FN += 1
else:
if prediction[i] == "F":
FP += 1
else:
TN += 1
accuracy.append((TP + TN) / (TP + TN + FP + FN))
if (TP + TN) * (TP + FP) * (TN + FP) * (TN + FN) != 0:
MCC.append((TP * TN - FP * FN) / math.sqrt(
(TP + TN) * (TP + FP) * (TN + FP) * (TN + FN)))
else:
MCC.append((TP * TN - FP * FN) / 0.000001)
seqLen.append(n)
avg_accuracy.append(sum(accuracy) / len(accuracy))
avg_MCC.append(sum(MCC) / len(MCC))
return [seqLen, avg_accuracy, avg_MCC]
def baum_welch(obs,L):
"""Given sequence and bs length L, approximate MLE parameters for
emission probabilities,transition rate a01 (background->site).
TODO: non-uniform background frequencies"""
states = range(L+1)
a01 = random.random()
start_p = make_start_p(a01)
trans_p = make_trans_p(a01)
emit_p = [simplex_sample(4) for state in states]
hidden_states = [random.choice(states) for ob in obs]
iterations = 0
while True:
# compute hidden states, given probs
prob,hidden_states_new = viterbi(obs, states, start_p, trans_p, emit_p)
# compute probs, given hidden states
# first compute a01
a01_new = estimate_a01(hidden_states_new)
start_p_new = make_start_p(a01_new)
trans_p_new = make_trans_p(a01_new)
emit_p_new = estimate_emit_p(obs,hidden_states_new,states)
if (start_p_new == start_p and
trans_p_new == trans_p and
emit_p_new == emit_p and
hidden_states_new == hidden_states):
break
else:
print iterations,a01,l2(start_p,start_p_new),
print l2(concat(trans_p),concat(trans_p_new)),
print l2((hidden_states),hidden_states_new)
a01 = a01_new
start_p = start_p_new
trans_p = trans_p_new
emit_p = emit_p_new
hidden_states = hidden_states_new
iterations += 1
return start_p,trans_p,emit_p,hidden_states
def memm(classifier, test_file, all_pos, start, transition, w):
test = open(test_file, "r")
tokens = []
index = []
pos = []
counter = 0
#used for validation only
correct = 0
seen = set()
greedy_predictions = []
memm_predictions = []
memm_lex = defaultdict(lambda: defaultdict(float))
for line in test:
predictions = []
if (counter % 3 == 0):
tokens = line.split()
for t in tokens:
seen.add(t)
#print ("tokens: " + "".join(tokens))
elif (counter % 3 == 1):
pos = line.split()
else:
index = line.split()
for i in range(len(tokens)):
dict = {}
#feature one = is this token captialized
#print ("word is " + tokens[i])
if tokens[i][0].isupper():
#print ("upper")
dict["caps"] = 1
else:
#print ("lower")
dict["caps"] = 0
dict["pos"] = pos[i]
if len(predictions) == 0:
dict["prevBIO"] = "<s>"
else:
dict["prevBIO"] = predictions[-1]
#feature 4 = does prev word start in caps
if i == 0:
dict["prevCaps"] = 0
else:
if tokens[i - 1][0].isupper():
dict["prevCaps"] = 1
else:
dict["prevCaps"] = 0
# feature 5 = previous pos
# if pos[i - 1] == "NNP":
# dict["prevNNP"] = 1
# else:
# dict["prevNNP"] = 0
if i == 0:
dict["prevPOS"] = "<s>"
else:
dict["prevPOS"] = pos[i - 1]
# feature 6 = previous word
if i == 0:
dict["prevWord"] = "<s>"
else:
dict["prevWord"] = tokens[i - 1]
# feature 7 = next word
if i == len(tokens) - 1:
dict["nextWord"] = "</s>"
else:
dict["nextWord"] = tokens[i + 1]
# feature 8 = next word pos
if i == len(tokens) - 1:
dict["nextPOS"] = "</s>"
else:
dict["nextPOS"] = pos[i + 1]
# feature 9 = does this word appear in training
dict["training"] = "nan"
# w[word] = tag
if tokens[i] in w:
dict["training"] = w[tokens[i]]
#print (dict)
probs = classifier.prob_classify(dict)
#greedy predictions
maxscore = 0
best_tag = ""
for tag in TAGS:
score = probs.prob(tag)
memm_lex[tag][tokens[i]] = score
if score > maxscore:
maxscore = score
best_tag = tag
#print(best_tag)
predictions.append(best_tag)
memm_pre = viterbi(tokens, start, transition, memm_lex, seen)
memm_predictions.append(memm_pre)
greedy_predictions.append(predictions)
counter += 1
test.close()
return greedy_predictions, memm_predictions
def execute():
# If there is no sentence input, asks for a sentence input
if len(sentence_input.get()) == 0:
canvas.delete("all")
canvas.create_text(375, 50, text="Please Enter a Sentence")
# If there is an illegal word in the input, tells the user
elif len(sentence_in_corpus(word_tokenize(sentence_input.get()))) > 0:
canvas.delete("all")
canvas.create_text(375, 50, text="Found an Illegal Word")
# If the user has selected "Forward Algorithm", it shows the results of running the Forward
# Algorithm on the entire sentence with only the most likely part of speech shown
elif v.get() == "1":
canvas.delete("all")
new_sentence = word_tokenize(sentence_input.get())
non_normalized_result = forward(set(pos_tags), new_sentence,
transition_matrix, emission_matrix)
result = normalize(non_normalized_result)
top_results = {}
for i in range(len(new_sentence)):
highest = 0.0
best_pos = None
for pos in set(pos_tags):
if result[i][pos] >= highest:
best_pos = pos
highest = result[i][pos]
top_results[new_sentence[i]] = best_pos, highest
new_len = 100 * len(result[0])
new_height = 60 * len(result) + 100
if new_len > 750 or new_height > 750:
canvas.config(height=new_height, width=new_len)
for i in range(len(result)):
canvas.create_rectangle(10 + 100 * i,
25,
100 + 100 * i,
75,
fill="white")
canvas.create_text(55 + 100 * i, 50, text=new_sentence[i])
canvas.create_rectangle(10 + 100 * i,
90,
100 + 100 * i,
135,
fill="lightblue")
val = top_results[new_sentence[i]][0] + ": " + "{0:.4f}".format(
top_results[new_sentence[i]][1])
canvas.create_text(55 + 100 * i, 112, text=val)
# If the user has selected "Viterbi Algorithm", it shows the results of running the Viterbi
# Algorithm on the entire sentence
elif v.get() == "2":
canvas.delete("all")
new_sentence = word_tokenize(sentence_input.get())
result = viterbi(set(pos_tags), new_sentence, transition_matrix,
emission_matrix)
new_len = 100 * len(result["predicted_tags"])
new_height = 750
if new_len > 750 or new_height > 750:
canvas.config(height=new_height, width=new_len)
for i in range(len(new_sentence)):
canvas.create_rectangle(10 + 100 * i,
25,
100 + 100 * i,
75,
fill="white")
canvas.create_text(55 + 100 * i, 50, text=new_sentence[i])
canvas.create_rectangle(10 + 100 * i,
90,
100 + 100 * i,
135,
fill="lightblue")
canvas.create_text(55 + 100 * i,
112,
text=result["predicted_tags"][i + 1])
# If the user has selected "Forward with All Parts of Speech", then it shows a trellis diagram for
# all the parts of speech
elif v.get() == "3":
canvas.delete("all")
new_sentence = word_tokenize(sentence_input.get())
non_normalized_result = forward(set(pos_tags), new_sentence,
transition_matrix, emission_matrix)
result = normalize(non_normalized_result)
print(result)
new_len = 100 * len(result)
new_height = 60 * len(result[0]) + 100
if new_len > 750 or new_height > 750:
canvas.config(height=new_height, width=new_len)
for i in range(len(result)):
canvas.create_rectangle(10 + 100 * i,
25,
100 + 100 * i,
75,
fill="white")
canvas.create_text(55 + 100 * i, 50, text=new_sentence[i])
j = 0
for pos in result[i]:
val = str(pos) + ": " + "{0:.3f}".format(result[i][pos])
canvas.create_rectangle(10 + 100 * i,
90 + 50 * j,
100 + 100 * i,
135 + 50 * j,
fill="lightblue")
canvas.create_text(55 + 100 * i, 112 + 50 * j, text=val)
j += 1
# If the user has selected "Show All Progressions for Forward", it shows the step progression after each word
# is added
elif v.get() == "4":
canvas.delete("all")
new_sentence = word_tokenize(sentence_input.get())
for i in range(1, len(new_sentence) + 1):
non_normalized_result = forward(set(pos_tags), new_sentence[:i],
transition_matrix, emission_matrix)
result = normalize(non_normalized_result)
top_results = {}
for j in range(len(new_sentence[:i])):
highest = 0.0
best_pos = None
for pos in set(pos_tags):
if result[j][pos] >= highest:
best_pos = pos
highest = result[j][pos]
top_results[new_sentence[j]] = best_pos, highest
new_len = 100 * len(result[0])
new_height = 60 * len(result) + 100
if new_len > 750 or new_height > 750:
canvas.config(height=new_height, width=new_len)
for k in range(len(result)):
canvas.create_rectangle(10 + 100 * k,
25,
100 + 100 * k,
75,
fill="white")
canvas.create_text(55 + 100 * k, 50, text=new_sentence[k])
canvas.create_rectangle(10 + 100 * k,
30 + 60 * i,
100 + 100 * k,
75 + 60 * i,
fill="lightblue")
val = top_results[
new_sentence[k]][0] + ": " + "{0:.4f}".format(
top_results[new_sentence[k]][1])
canvas.create_text(55 + 100 * k, 52 + 60 * i, text=val)
# If the user has selected "Show All Progressions for Viterbi", it shows the step progression after each word
# is added to the algorithm
elif v.get() == "5":
canvas.delete("all")
new_sentence = word_tokenize(sentence_input.get())
for i in range(1, len(new_sentence) + 1):
result = viterbi(set(pos_tags), new_sentence[:i],
transition_matrix, emission_matrix)
new_len = 100 * len(result["predicted_tags"])
new_height = 750
if new_len > 750 or new_height > 750:
canvas.config(height=new_len, width=new_len)
for j in range(len(new_sentence[:i])):
canvas.create_rectangle(10 + 100 * j,
25,
100 + 100 * j,
75,
fill="white")
canvas.create_text(55 + 100 * j, 50, text=new_sentence[j])
canvas.create_rectangle(10 + 100 * j,
30 + 60 * i,
100 + 100 * j,
75 + 60 * i,
fill="lightblue")
canvas.create_text(55 + 100 * j,
52 + 60 * i,
text=result["predicted_tags"][j + 1])
print(sum(B[1, :]))
# ## Question 4
# In[12]:
tags_true = []
tags_pred = []
scores = []
for sent in tqdm(testing):
word_list = to_ids(V, [word for word, _ in sent])
tag_list = to_ids(Q, [tag for _, tag in sent])
tags_true.append(tag_list)
predicted, score = viterbi((Pi, A, B), word_list)
tags_pred.append(predicted)
scores.append(score)
# In[13]:
predicted_set = set(flatten(tags_pred))
reference_set = set(flatten(tags_true))
print('Precision :', precision(predicted_set, reference_set))
print('Recall :', recall(predicted_set, reference_set))
print('F1-score :', f_measure(predicted_set, reference_set))
# ## Question 5
# We only conserve the pairs of tags which appears at least once in the whole set.
f.close()
if (function == 'viterbi'):
f = open(outfile, 'w')
print("generating tables..")
EMISSION = emissionTable(train_X, train_Y, test_X)
print("emission done")
TRANSITION = transitionTable(train_Y)
print("transition done")
unique_tags = getUniqueY(train_Y)
print("unique tags gotten from text")
print("All pre-requisites done, now running viterbi")
for i in range(0, len(test_X)):
# print(test_X[i])
print("Writing one sentence, " + str(len(test_X) - i) + " to go.")
viterbi_sentence = viterbi(test_X[i], len(test_X[i]), TRANSITION,
EMISSION, unique_tags)
for j in range(0, len(test_X[i])):
towrite = str(test_X[i][j]) + " " + str(viterbi_sentence[j])
f.write(towrite + '\n')
f.write('\n')
f.close()
if (function == 'viterbi_topk'):
f = open(outfile, 'w')
print("generating tables..")
EMISSION = emissionTable(train_X, train_Y, test_X)
print("emission done")
TRANSITION = transitionTable(train_Y)
print("transition done")
unique_tags = getUniqueY(train_Y)
print("unique tags gotten from text")
def main(
param=0.2,
PATH_LOAD_FILE='/home/keums/Melody/dataset/adc2004_full_set/file/pop4.wav',
PATH_SAVE_FILE='./SAVE_RESULTS/pop4.txt'):
# PATH_LOAD_FILE = sys.argv[1]
# PATH_SAVE_FILE = sys.argv[2]
#==================================
# Feature Extraction
# .wav --> spectrogram
#==================================
x_test_log = myFeatureExtraction(PATH_LOAD_FILE) #path ??
#==================================
# making multi column spectrogram
# for trainging
#==================================
x_test_SF = making_multi_frame(x_test_log, num_frames=1)
x_test_MF = making_multi_frame(x_test_log, num_frames=11)
select_res_1st = 1
select_res_2nd = 2
select_res_3rd = 4
pitch_range = np.arange(min_pitch, max_pitch + 1.0 / select_res_3rd,
1.0 / select_res_3rd)
#==================================
# Melody extraction
# using DNN
#==================================
y_predict_1st = MelodyExtraction_SCDNN(x_test_MF, select_res_1st)
y_predict_2nd = MelodyExtraction_SCDNN(x_test_MF, select_res_2nd)
y_predict_3rd = MelodyExtraction_SCDNN(x_test_MF, select_res_3rd)
#==================================
# merge SCDNN
#==================================
# print 'Merging....'
ratio_res_1_3 = select_res_3rd / select_res_1st
ratio_res_2_3 = select_res_3rd / select_res_2nd
y_predict_tmp_1_3 = np.zeros(y_predict_3rd.shape)
y_predict_tmp_2_3 = np.zeros(y_predict_3rd.shape)
for i in range(y_predict_3rd.shape[0]):
for j in range(y_predict_1st.shape[1] - 1):
y_predict_tmp_1_3[i, j * ratio_res_1_3:j * ratio_res_1_3 +
ratio_res_1_3] = y_predict_1st[i, j]
y_predict_tmp_1_3[i, -1] = y_predict_1st[i, -1]
for i in range(y_predict_3rd.shape[0]):
for j in range(y_predict_2nd.shape[1] - 1):
y_predict_tmp_2_3[i, j * ratio_res_2_3:j * ratio_res_2_3 +
ratio_res_2_3] = y_predict_2nd[i, j]
y_predict_tmp_2_3[i, -1] = y_predict_2nd[i, -1]
# y_predict = (y_predict_tmp_1_3+0.0000001) *(y_predict_tmp_2_3+0.0000001) * (y_predict_3rd +0.0000001)
y_predict = 10**(np.log10(y_predict_tmp_1_3) +
np.log10(y_predict_tmp_2_3) + np.log10(y_predict_3rd))
del y_predict_tmp_1_3
del y_predict_tmp_2_3
#==================================
# singing voice detection
#==================================
voice_frame_vad = VAD_DNN(x_test_SF, y_predict_1st, param=0.2)
#==================================
# viterbi algorithm
#==================================
path_viterbi = './viterbi/'
path_prior_matrix_file = path_viterbi + 'prior_' + str(
select_res_3rd) + '.npy'
path_transition_matrix_file = path_viterbi + 'transition_matrix_' + str(
select_res_3rd) + '.npy'
prior = np.load(path_prior_matrix_file)
transition_matrix = np.load(path_transition_matrix_file)
viterbi_path = viterbi(y_predict,
transition_matrix=transition_matrix,
prior=prior,
penalty=0,
scaled=True)
pitch_MIDI = np.zeros([y_predict.shape[0], 1])
pitch_freq = np.zeros([y_predict.shape[0], 1])
for i in range(y_predict.shape[0]):
# for test : origianl
# index_predict[i] = np.argmax(y_predict[i,:])
# pitch_MIDI[i] = pitch_range[index_predict[i]]
#viterbi_path
pitch_MIDI[i] = pitch_range[viterbi_path[i]]
pitch_freq[i] = 2**((pitch_MIDI[i] - 69) / 12.) * 440
est_pitch = np.multiply(pitch_freq, voice_frame_vad)
#==================================
#adjust frame
#==================================
idx_shift = 2
shift_array = np.zeros(idx_shift)
est_pitch = np.append(shift_array, est_pitch[:-idx_shift])
#==================================
# save result
#==================================
PATH_est_pitch = PATH_SAVE_FILE
if not os.path.exists(os.path.dirname(PATH_est_pitch)):
os.makedirs(os.path.dirname(PATH_est_pitch))
f = open(PATH_est_pitch, 'w')
for j in range(len(est_pitch)):
est = "%f\t%f\n" % (0.01 * j, est_pitch[j])
f.write(est)
f.close()
print PATH_est_pitch
if random.random() < P[1]:
state = 1
for i in range(60):
switched_state = (state + 1) % 2
if i > 0 and random.random() < Tm[state][switched_state]:
state = switched_state
lines[0] += str(random.choices(list(range(1, 7)), Em[state])[0])
lines[1] += 'F' if state == 0 else 'L'
return lines
probabilities_file, rolls_file = parse_args()
P, Tm, Em = parse_probabilities(probabilities_file)
if rolls_file is None:
lines = generate_rolls()
else:
with open(rolls_file) as input:
lines = input.read().splitlines()
print('Rolls: ' + lines[0])
print('Die: ' + lines[1])
observations = [int(i) for i in lines[0]]
viter = viterbi(S, P, observations, Tm, Em)
forw_back = forward_backward(observations, S, P, Tm, Em)
percent = lambda obs: str(
round(
sum([obs[i] == lines[1][i]
for i in range(len(observations))]) / len(observations) * 100, 2))
print('Viterbi: ' + ''.join(viter) + ' (' + percent(viter) + '%)')
print('Posterior: ' + ''.join(forw_back) + ' (' + percent(forw_back) + '%)')