def train_and_test(self, dataset, batch_size):
self.batch_size = batch_size
x_train, y_train = dataset.get_training_data()
num_examples = dataset.get_num_training_examples()
dropped = 0
for idx in range(num_examples):
if idx % 100 == 0:
print('Training on {} out of {} examples'.format(
idx, num_examples))
mfcc_vec_seq = x_train[idx]
phoneme_idx = y_train[idx]
# drop sequences that are too short
if len(mfcc_vec_seq) < self.n_states:
print('dropped')
dropped += 1
else:
# find an appropriate hmm and fit using EM
hmm = self.hmms[phoneme_idx]
# add phonem frequency
if not phoneme_idx in self.phonem_freq:
self.phonem_freq[phoneme_idx] = 0
self.phonem_freq[phoneme_idx] += 1
hmm.fit(mfcc_vec_seq)
# calculate phonem frequences
phonem_num = sum(self.phonem_freq.values())
for key in self.phonem_freq:
self.phonem_freq[key] /= float(phonem_num)
print(self.phonem_freq[key])
print(key)
print('Done training, dropped {} out of {}'.format(
dropped, num_examples))
self.test_on_random_training_batch(dataset, self.batch_size)
self.test(dataset)
def train(dataset):
# Get all vectors in the datasets
all_vectors = np.concatenate(
[np.concatenate(v, axis=0) for k, v in dataset.items()], axis=0)
print("vectors", all_vectors.shape)
# Run K-Means algorithm to get clusters
kmeans = clustering(all_vectors)
print("centers", kmeans.cluster_centers_.shape)
models = {}
for cname in CLASS_NAMES:
# print(cname[:4])
# class_vectors = dataset[cname]
# convert all vectors to the cluster index
# dataset['one'] = [O^1, ... O^R]
# O^r = (c1, c2, ... ct, ... cT)
# O^r size T x 1
dataset[cname] = list(
[kmeans.predict(v).reshape(-1, 1) for v in dataset[cname]])
#define model
hmm = hmm_model()
if 'test' not in cname:
X = np.concatenate(dataset[cname])
lengths = list([len(x) for x in dataset[cname]])
print("training class", cname)
print(X.shape, lengths, len(lengths))
hmm.fit(X, lengths=lengths)
models[cname] = hmm
print("Training done")
return models
def train_hmm_and_keep_track_of_log_likelihood(hmm, obs, n_iter=1, **kwargs):
hmm.n_iter = 1
hmm.fit(obs)
loglikelihoods = []
for n in range(n_iter):
hmm.n_iter = 1
hmm.init_params = ''
hmm.fit(obs)
loglikelihoods.append(sum(hmm.score(x) for x in obs))
return loglikelihoods
def train_hmm_and_keep_track_of_log_likelihood(hmm, obs, n_iter=1, **kwargs):
hmm.n_iter = 1
hmm.fit(obs)
loglikelihoods = []
for n in range(n_iter):
hmm.n_iter = 1
hmm.init_params = ''
hmm.fit(obs)
loglikelihoods.append(sum(hmm.score(x) for x in obs))
return loglikelihoods
def test_ala2():
# creates a 4-state HMM on the ALA2 data. Nothing fancy, just makes
# sure the code runs without erroring out
trajectories = AlanineDipeptide().get_cached().trajectories
topology = trajectories[0].topology
indices = topology.select('symbol C or symbol O or symbol N')
featurizer = SuperposeFeaturizer(indices, trajectories[0][0])
sequences = featurizer.transform(trajectories)
hmm = GaussianHMM(n_states=4, n_init=3)
hmm.fit(sequences)
assert len(hmm.timescales_ == 3)
assert np.any(hmm.timescales_ > 50)
def test_ala2():
# creates a 4-state HMM on the ALA2 data. Nothing fancy, just makes
# sure the code runs without erroring out
trajectories = AlanineDipeptide().get_cached().trajectories
topology = trajectories[0].topology
indices = topology.select('symbol C or symbol O or symbol N')
featurizer = SuperposeFeaturizer(indices, trajectories[0][0])
sequences = featurizer.transform(trajectories)
hmm = GaussianHMM(n_states=4, n_init=3, random_state=rs)
hmm.fit(sequences)
assert len(hmm.timescales_ == 3)
assert np.any(hmm.timescales_ > 50)
def test_pickle():
"""Test pickling an HMM"""
trajectories = AlanineDipeptide().get_cached().trajectories
topology = trajectories[0].topology
indices = topology.select('symbol C or symbol O or symbol N')
featurizer = SuperposeFeaturizer(indices, trajectories[0][0])
sequences = featurizer.transform(trajectories)
hmm = GaussianHMM(n_states=4, n_init=3, random_state=rs)
hmm.fit(sequences)
logprob, hidden = hmm.predict(sequences)
with tempfile.TemporaryFile() as savefile:
pickle.dump(hmm, savefile)
savefile.seek(0, 0)
hmm2 = pickle.load(savefile)
logprob2, hidden2 = hmm2.predict(sequences)
assert (logprob == logprob2)
def test_pickle():
"""Test pickling an HMM"""
trajectories = AlanineDipeptide().get_cached().trajectories
topology = trajectories[0].topology
indices = topology.select('symbol C or symbol O or symbol N')
featurizer = SuperposeFeaturizer(indices, trajectories[0][0])
sequences = featurizer.transform(trajectories)
hmm = GaussianHMM(n_states=4, n_init=3, random_state=rs)
hmm.fit(sequences)
logprob, hidden = hmm.predict(sequences)
with tempfile.TemporaryFile() as savefile:
pickle.dump(hmm, savefile)
savefile.seek(0, 0)
hmm2 = pickle.load(savefile)
logprob2, hidden2 = hmm2.predict(sequences)
assert(logprob == logprob2)
def train(dataset):
# Get all vectors in the datasets
all_vectors = np.concatenate(
[np.concatenate(v, axis=0) for k, v in dataset.items()], axis=0)
print("vectors", all_vectors.shape)
# Run K-Means algorithm to get clusters
kmeans = clustering(all_vectors)
print("centers", kmeans.cluster_centers_.shape)
models = {}
for cname in CLASS_NAMES:
# print(cname[:4])
# class_vectors = dataset[cname]
# convert all vectors to the cluster index
# dataset['one'] = [O^1, ... O^R]
# O^r = (c1, c2, ... ct, ... cT)
# O^r size T x 1
dataset[cname] = list(
[kmeans.predict(v).reshape(-1, 1) for v in dataset[cname]])
if cname == "benh_nhan":
hmm = hmm_model(N_COMPONENT_BN, START_PROB_BN, TRANSMAT_PRIOR_BN)
elif cname == "cua":
hmm = hmm_model(N_COMPONENT_CUA, START_PROB_CUA,
TRANSMAT_PRIOR_CUA)
elif cname == "khong":
hmm = hmm_model(N_COMPONENT_KHONG, START_PROB_KHONG,
TRANSMAT_PRIOR_KHONG)
elif cname == "nguoi":
hmm = hmm_model(N_COMPONENT_NGUOI, START_PROB_NGUOI,
TRANSMAT_PRIOR_NGUOI)
#define model
# hmm = hmm_model()
if 'test' not in cname:
X = np.concatenate(dataset[cname])
lengths = list([len(x) for x in dataset[cname]])
print("training class", cname)
print(X.shape, lengths, len(lengths))
hmm.fit(X, lengths=lengths)
models[cname] = hmm
print("Training done")
return models
def train(data_train, state_num):
models = {}
dataset = data_train.copy()
for cname in dataset.keys():
n = state_num[cname]
startprob = np.zeros(n)
startprob[0] = 1
transmat = np.diag(np.full(n, 1))
hmm = hmmlearn.hmm.MultinomialHMM(
n_components=n,
random_state=0,
n_iter=1000,
verbose=False,
startprob_prior=startprob,
transmat_prior=transmat,
)
X = np.concatenate(dataset[cname])
lengths = list([len(x) for x in dataset[cname]])
hmm.fit(X, lengths=lengths)
models[cname] = hmm
return models
params='mctw',
init_params='mst')
hmm.startprob_ = np.array([1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0])
hmm.transmat_ = np.array([
[0.7, 0.2, 0.1, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.7, 0.2, 0.1, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.7, 0.2, 0.1, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.7, 0.2, 0.1, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.7, 0.2, 0.1],
[0.0, 0.0, 0.0, 0.0, 0.0, 0.7, 0.3],
[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0],
])
X = np.concatenate(dataset_train[cname])
lengths = list([len(x) for x in dataset_train[cname]])
hmm.fit(X)
models[cname] = hmm
print("Training done")
with open("gmm_hmm.pkl", "wb") as file:
pickle.dump(models, file)
print("Saved!")
print("Testing")
#class_names = ['khong', 'toi', 'trong', 'amtinh',"test_toi","test_trong","test_khong","test_amtinh"]
for true_cname in class_names:
#if true_cname[:4] == 'test':
true_predict = 0
# for O in dataset[true_cname]:
for O in dataset_test[true_cname]:
0.1,
0.1,
0.5,
],
[
0.1,
0.1,
0.1,
0.1,
0.1,
0.5,
],
]),
)
if cname[:4] != 'test':
X = np.concatenate(dataset[cname])
lengths = list([len(x) for x in dataset[cname]])
print("training class", cname)
print(X.shape, lengths, len(lengths))
hmm.fit(X, lengths=lengths)
models[cname] = hmm
print("Training done")
print("Testing")
for true_cname in class_names:
for O in dataset[true_cname]:
score = {
cname: model.score(O, [len(O)])
for cname, model in models.items() if cname[:4] != 'test'
}
print(true_cname, score)
def train(evaluate=False):
''' if evaluate is True, test data will be choosen randomly from train data set
however, you need to backup your self-record test data before use this evaluate
'''
global trained
trained = True
class1 = ["toi", "test_toi"]
class2 = ["mot", "test_mot"]
class3 = ["trong", "test_trong"]
class4 = ["thoigian", "test_thoigian"]
class5 = ["chungta", "test_chungta"]
class_names = []
class_names.extend(class1)
class_names.extend(class2)
class_names.extend(class3)
class_names.extend(class4)
class_names.extend(class5)
dataset = {}
if evaluate is True:
for cname in class_names:
if cname[:4] == "test":
cname_ = cname.split("_")[1]
data_dir_src = os.path.join("data", cname_)
data_dir_dst = os.path.join("data", cname)
samples = random.sample([x for x in os.listdir(data_dir_src) if os.path.isfile(os.path.join(data_dir_src, x))], 10)
[copyfile(os.path.join(data_dir_src, x), os.path.join(data_dir_dst, x)) for x in samples]
for cname in class_names:
print(f"Load {cname} dataset")
dataset[cname] = get_class_data(os.path.join("data", cname))
all_vectors = np.concatenate([np.concatenate(v, axis=0) for k, v in dataset.items()], axis=0)
kmeans = clustering(all_vectors)
kmeans_model_filename = 'kmeans.joblib'
with open (kmeans_model_filename, 'wb') as f_kmeans:
joblib.dump(kmeans, f_kmeans)
print("centers", kmeans.cluster_centers_.shape)
config = {
'toi': {'n_components':5},
'mot': {'n_components': 5},
'trong': {'n_components': 5},
'thoigian': {'n_components': 8}, # 10
'chungta': {'n_components': 8},
'test_toi': {'n_components': 5},
'test_mot': {'n_components': 5},
'test_trong': {'n_components': 5},
'test_thoigian': {'n_components': 8}, # 10
'test_chungta': {'n_components': 8},
'demo': {'n_components': 3},
}
for cname in class_names:
dataset[cname] = list([kmeans.predict(v).reshape(-1,1) for v in dataset[cname]])
n_components = config[cname]['n_components']
start_prob = np.zeros(n_components)
start_prob[0] = 1.0
transmat = np.ndarray(shape=(n_components, n_components), dtype=float)
for i in range((n_components - 1)):
transmat[i][i] = 0.7
transmat[i][i + 1] = 0.3
transmat[n_components - 1][n_components - 1] = 1.0
hmm = hmmlearn.hmm.MultinomialHMM(
n_components=n_components, random_state=0, n_iter=1000, verbose=True,
transmat_prior=transmat,
startprob_prior=start_prob,
init_params='ste',
params='ste'
)
# hmm.startprob_ = start_prob
# hmm.transmat_ = transmat
if cname[:4] != 'test':
X = np.concatenate(dataset[cname])
lengths = list([len(x) for x in dataset[cname]])
print("training class", cname)
hmm.fit(X, lengths=lengths)
models[cname] = hmm
print("Training done")
model_filename = 'finalized_model.joblib'
with open (model_filename, 'wb') as f_hmm:
joblib.dump(models, f_hmm)
if evaluate is True:
print("Testing")
for true_cname in class_names:
if true_cname != 'demo':
count = 0
for O in dataset[true_cname]:
score = {cname : model.score(O, [len(O)]) for cname, model in models.items() if cname[:4] != 'test' and cname != 'demo' }
predict = max(score, key = score.get)
if predict == true_cname or true_cname[:4] == 'test' and predict == true_cname.split('_')[1]:
count+=1
# print(true_cname, score, predict)
print(f"true: {count}/{len(dataset[true_cname])}")
for cname in class_names:
if cname[:4] == "test":
data_dir_dst = os.path.join("data", cname)
samples = os.listdir(data_dir_dst)
[os.remove(os.path.join(data_dir_dst, x)) for x in samples]
return models
def main():
n_states = {
'khong': 11,
'vietnam': 24,
'nguoi': 11,
'benhvien': 24,
'trong': 11
}
#Load dataset
class_names = [
f for f in os.listdir('data') if os.path.isdir(os.path.join('data', f))
]
dataset = {}
for cname in class_names:
print(f"-->Load {cname} dataset")
dataset[cname] = get_class_data(cname)
# Get all vectors in the datasets
all_vectors = np.concatenate(
[np.concatenate(v, axis=0) for k, v in dataset.items()], axis=0)
print("vectors", all_vectors.shape)
# Run K-Means algorithm to get clusters
kmeans = clustering(all_vectors, n_clusters=35)
#Train model
models = {}
class_vectors = dataset.copy()
for cname in class_names:
# convert all vectors to the cluster index
# dataset['one'] = [O^1, ... O^R]
# O^r = (c1, c2, ... ct, ... cT)
# O^r size T x 1
class_vectors[cname] = list(
[kmeans.predict(v).reshape(-1, 1) for v in class_vectors[cname]])
if cname[:4] != 'test':
hmm = hmmlearn.hmm.MultinomialHMM(
n_components=n_states[cname],
random_state=2020,
n_iter=1000,
verbose=False,
init_params='e',
params='te',
)
hmm.startprob_ = np.array(make_pi(n_states[cname]))
hmm.transmat_ = np.array(make_A(n_states[cname]))
X = np.concatenate(class_vectors[cname])
lengths = list([len(x) for x in class_vectors[cname]])
print("training class", cname)
print(X.shape, lengths, len(lengths))
hmm.fit(X, lengths=lengths)
models[cname] = hmm
print("<--Training done-->\n")
print("-----Testing-----")
print("Test in Datatrain")
for test_cname in class_names:
cnt = 0
if test_cname[:4] != "test":
for O in class_vectors[test_cname]:
score = {
cname: model.score(O, [len(O)])
for cname, model in models.items() if cname[:4] != 'test'
}
max_value = max(v for k, v in score.items())
#print("Max: ", max_value)
for k, v in score.items():
if v == max_value:
if k == test_cname:
cnt += 1
#print(test_cname, score)
print(f"{test_cname} -- Score: ",
cnt / len(class_vectors[test_cname]))
print()
print("Test in Datatest")
for test_cname in class_names:
cnt = 0
if test_cname[:4] == "test":
for O in class_vectors[test_cname]:
score = {
cname: model.score(O, [len(O)])
for cname, model in models.items() if cname[:4] != 'test'
}
max_value = max(v for k, v in score.items())
#print("Max: ", max_value)
for k, v in score.items():
if v == max_value:
predict = k
if predict.strip() == test_cname[5:].strip():
cnt += 1
#print(test_cname, score)
print(f"{test_cname} -- Score: ",
cnt / len(class_vectors[test_cname]))
#Extract models parameters
with open("Models_parameters.txt", "w") as f:
for cname, model in models.items():
f.write(f"Model_name : {cname}\n")
f.write("Startprob matrix:\n")
f.write(" ".join(map(str, model.startprob_)))
f.write("\nTransition Matrix\n")
f.write(" ".join(map(str, model.transmat_)))
f.write("\nEmissionProb Matrix\n")
f.write(" ".join(map(str, model.emissionprob_)))
f.write("\n\n")
print("Extracted models to Models_parameters.txt successfully")
#Save models
if "models" not in os.listdir():
os.mkdir("models")
#Kmeans
with open(os.path.join("models", "kmeans.pkl"), "wb") as f:
pickle.dump(kmeans, f)
print("Saved Kmeans model to 'models/kmeans.pkl' successfully")
#HMM
with open(os.path.join("models", "models.pkl"), "wb") as f:
pickle.dump(models, f)
print(f"Saved HMMs model to 'models/models.pkl' successfully")
