def test_slfit(self):
sequences = io.load_sequences(self._train_filename)
hmm = HMM.from_file(self._model_filename)
learner = SLHMM(self._num_hidden, self._num_observ)
learner.fit(sequences, verbose=True)
for sequence in sequences:
pprint("True probability: %f" % hmm.predict(sequence))
pprint("Infered probability: %f" % learner.predict(sequence))
def test_slfit(self):
sequences = io.load_sequences(self._train_filename)
hmm = HMM.from_file(self._model_filename)
learner = SLHMM(self._num_hidden, self._num_observ)
learner.fit(sequences, verbose=True)
for sequence in sequences:
pprint("True probability: %f" % hmm.predict(sequence))
pprint("Infered probability: %f" % learner.predict(sequence))
class Experimenter(object):
'''
This class is built to facilitate the experiments of different learning
algorithms.
'''
def __init__(self, training_filename, test_filename, model_filename, num_hidden,
num_observ, num_em_restarts=20):
self._training_data = [np.loadtxt(training_filename, dtype=np.int, delimiter=",")]
# self._test_data = np.loadtxt(test_filename, dtype=np.int, delimiter=",")
self._test_data = []
with file(test_filename, "rb") as fin:
reader = csv.reader(fin)
for line in reader:
self._test_data.append(np.asarray(map(int, line)))
self._model = HMM.from_file(model_filename)
self._num_hidden = num_hidden
self._num_observ = num_observ
self._num_em_restarts = num_em_restarts
@property
def training_size(self):
return self._training_data[0].shape[0]
@property
def test_size(self):
return len(self._test_data)
@property
def num_em_restarts(self):
return self._num_em_restarts
def _train(self, num_train_inst):
'''
Train a Hidden Markov Model with differnt learning algorithms
'''
num_train_inst = min(num_train_inst, self._training_data[0].shape[0])
training_data = self._training_data[0][:num_train_inst]
pprint("=" * 50)
pprint("Training set length: %d" % num_train_inst)
# Spectral learning algorithm
start_time = time.time()
self._sl_learner = SLHMM(self._num_hidden, self._num_observ)
self._sl_learner.fit([training_data])
end_time = time.time()
pprint("Time used for Spectral Learner: %f" % (end_time - start_time))
sl_time = end_time - start_time
# Expectation Maximization algorithm
#self._em_learners = []
em_times = np.zeros(self._num_em_restarts, dtype=np.float)
#for i in xrange(self._num_em_restarts):
#self._em_learners.append(EMHMM(self._num_hidden, self._num_observ))
#start_time = time.time()
#self._em_learners[i].fit([training_data], max_iters=20, verbose=True)
#end_time = time.time()
#pprint("Time used for Expectation Maximization: %f" % (end_time - start_time))
#em_times[i] = end_time - start_time
return (sl_time, np.mean(em_times))
def run_experiment(self, num_train_inst):
'''
@log_filename: string, filepath of the output log
'''
sl_time, em_time = self._train(num_train_inst)
true_probs = np.zeros(len(self._test_data), dtype=np.float)
sl_probs = np.zeros(len(self._test_data), dtype=np.float)
em_probs = np.zeros((self._num_em_restarts, len(self._test_data)), dtype=np.float)
for i, seq in enumerate(self._test_data):
true_probs[i] = self._model.predict(seq)
sl_probs[i] = self._sl_learner.predict(seq)
#for j in xrange(self._num_em_restarts):
#em_probs[j, i] = self._em_learners[j].predict(seq)
# L1-distance between true probability and inference probability by spectral learning
sl_variation_dist = np.abs(true_probs - sl_probs)
# L1-distance between true probability and inference probability by expectation maximization
em_variation_dist = np.abs(true_probs - em_probs)
# Sum of L1-distance
sl_variation_measure = np.sum(sl_variation_dist)
em_variation_measure = np.sum(em_variation_dist, axis=1)
return (sl_time, em_time, sl_variation_measure, em_variation_measure)
class Experimenter(object):
'''
This class is built to facilitate the experiments of different learning
algorithms.
'''
def __init__(self,
training_filename,
test_filename,
model_filename,
num_hidden,
num_observ,
num_em_restarts=20):
self._training_data = [
np.loadtxt(training_filename, dtype=np.int, delimiter=",")
]
# self._test_data = np.loadtxt(test_filename, dtype=np.int, delimiter=",")
self._test_data = []
with file(test_filename, "rb") as fin:
reader = csv.reader(fin)
for line in reader:
self._test_data.append(np.asarray(map(int, line)))
self._model = HMM.from_file(model_filename)
self._num_hidden = num_hidden
self._num_observ = num_observ
self._num_em_restarts = num_em_restarts
@property
def training_size(self):
return self._training_data[0].shape[0]
@property
def test_size(self):
return len(self._test_data)
@property
def num_em_restarts(self):
return self._num_em_restarts
def _train(self, num_train_inst):
'''
Train a Hidden Markov Model with differnt learning algorithms
'''
num_train_inst = min(num_train_inst, self._training_data[0].shape[0])
training_data = self._training_data[0][:num_train_inst]
pprint("=" * 50)
pprint("Training set length: %d" % num_train_inst)
# Spectral learning algorithm
start_time = time.time()
self._sl_learner = SLHMM(self._num_hidden, self._num_observ)
self._sl_learner.fit([training_data])
end_time = time.time()
pprint("Time used for Spectral Learner: %f" % (end_time - start_time))
sl_time = end_time - start_time
# Expectation Maximization algorithm
#self._em_learners = []
em_times = np.zeros(self._num_em_restarts, dtype=np.float)
#for i in xrange(self._num_em_restarts):
#self._em_learners.append(EMHMM(self._num_hidden, self._num_observ))
#start_time = time.time()
#self._em_learners[i].fit([training_data], max_iters=20, verbose=True)
#end_time = time.time()
#pprint("Time used for Expectation Maximization: %f" % (end_time - start_time))
#em_times[i] = end_time - start_time
return (sl_time, np.mean(em_times))
def run_experiment(self, num_train_inst):
'''
@log_filename: string, filepath of the output log
'''
sl_time, em_time = self._train(num_train_inst)
true_probs = np.zeros(len(self._test_data), dtype=np.float)
sl_probs = np.zeros(len(self._test_data), dtype=np.float)
em_probs = np.zeros((self._num_em_restarts, len(self._test_data)),
dtype=np.float)
for i, seq in enumerate(self._test_data):
true_probs[i] = self._model.predict(seq)
sl_probs[i] = self._sl_learner.predict(seq)
#for j in xrange(self._num_em_restarts):
#em_probs[j, i] = self._em_learners[j].predict(seq)
# L1-distance between true probability and inference probability by spectral learning
sl_variation_dist = np.abs(true_probs - sl_probs)
# L1-distance between true probability and inference probability by expectation maximization
em_variation_dist = np.abs(true_probs - em_probs)
# Sum of L1-distance
sl_variation_measure = np.sum(sl_variation_dist)
em_variation_measure = np.sum(em_variation_dist, axis=1)
return (sl_time, em_time, sl_variation_measure, em_variation_measure)