def train_and_save(parameters, dataset, filename):
"""Creates a HMMChain and, loads its parameters, trains it on dataset and
saves the results in filename.
"""
chain = hmm_chain.HMMChain()
chain.set_params(**parameters)
chain.my_class = None
chain.other_classes = None
dataset, _, _ = create_dataset_crossvalidation(dataset)
chain.fit(dataset)
score = chain.score(dataset)
display_name = os.path.split(filename)[1].split(".")[0]
longline = "========%s========" % display_name
print
print longline
print "After training the score is ", score
print "with parameters: %s" % parameters
print "=" * len(longline)
print
with open(filename, "w") as f:
print "Saving file: ", filename
cPickle.dump(chain, f, protocol=cPickle.HIGHEST_PROTOCOL)
def train_and_save(parameters, dataset, filename):
chain = hmm_chain.HMMChain()
chain.set_params(**parameters)
chain.my_class = None
chain.other_classes = None
chain.fit(dataset)
score = chain.score(dataset)
print "After training the score is ", score
with open(filename, "w") as f:
print "Saving file: ", filename
cPickle.dump(chain, f, protocol=cPickle.HIGHEST_PROTOCOL)
def train_dataset(dataset):
"""Uses cross validation to train a HMM chain. Returns the parameters
that yield the best result.
CHANGE THE PARAMETERS HERE TO USE A SUITABLE RANGE, THESE ARE ONLY HERE FOR
TESTING!
"""
dataset, train_indexes, test_indexes = create_dataset_crossvalidation(dataset)
cv = [(train_indexes, test_indexes)]
parameters = [
dict(n_pca_components = [0.97],
n_hidden_components=[12, 15, 18],
resampling_size=[20, 25, 30],
n_discretization_symbols=[5, 10, 12],
hmm_max_iter = [100],
#kmeans_max_iter = [1000]
),
#dict(n_pca_components = [0.97],
#n_hidden_components=[40, 50],
#resampling_size=[20],
#n_discretization_symbols=[30, ],
#hmm_max_iter = [2000],
##kmeans_max_iter = [1000]
#),
]
'''parameters = [
dict(n_pca_components = [0.97],
n_hidden_components=[15, 18, 25],
resampling_size=[30, 40, 50],
n_discretization_symbols=[10, 12, 15],
hmm_max_iter = [300],
),
]'''
#print "Using parameters:\n", parameters
chain = hmm_chain.HMMChain()
grid = sklearn.grid_search.GridSearchCV(chain, parameters,
cv = cv,
verbose = 10,
n_jobs = 6,
refit = False
)
grid.fit(dataset)
return grid.best_params_
def train_dataset(dataset, all_adjectives, adjective):
parameters = [
dict(
n_pca_components=[0.97],
n_hidden_components=[35, 40, 45],
resampling_size=[20],
n_discretization_symbols=[
25,
],
hmm_max_iter=[2000],
#kmeans_max_iter = [1000]
),
#dict(n_pca_components = [0.97],
#n_hidden_components=[40, 50],
#resampling_size=[20],
#n_discretization_symbols=[30, ],
#hmm_max_iter = [2000],
##kmeans_max_iter = [1000]
#),
]
print "Using parameters:\n", parameters
chain = hmm_chain.HMMChain()
cross_validator = sklearn.cross_validation.ShuffleSplit(len(dataset),
n_iterations=2,
train_size=3. / 4.)
for p in parameters:
p.update(my_class=[adjective], other_classes=[all_adjectives])
grid = sklearn.grid_search.GridSearchCV(chain,
parameters,
cv=cross_validator,
verbose=10,
n_jobs=6,
refit=False)
grid.fit(dataset)
return grid.best_params_
from pylab import *
import utilities
import hmm_chain
import cPickle
bumpy = cPickle.load(
open("/home/pezzotto/log/bigbags/bag_files/databases/bumpy.pkl"))
pdc = bumpy['SLIDE_5CM']['pdc']
splits = [len(d) for d in pdc]
hmm = hmm_chain.HMMChain(data_splits=splits,
n_pca_components=1,
resampling_size=50,
n_discretization_symbols=5)
hmm.update_splits(pdc)
pca = hmm.pca
pca.fit(vstack(pdc))
Xt = hmm.splitter.transform(pca.transform(hmm.combiner.transform(pdc)))
Xt = hmm.resample.fit_transform(Xt)
Xt = hmm.combiner.transform(Xt)
hmm.discretizer.fit(Xt)
Xt = hmm.discretizer.transform(Xt)
Xt = hmm.splitter2.transform(Xt)
hmm.hmm.fit(Xt)
print "Score: ", hmm.score(pdc)
print "Using the whole training"
pdc = bumpy['SLIDE_5CM']['pdc']
hmm.fit(pdc)
print "Score: ", hmm.score(pdc)