def learning_curve():
n = 50000
nsteps = 10
full = cu.get_sample_data_frame(n)
data = full.ix[0 : int(n * 0.6) - 1].reset_index()
cval = full.ix[int(n * 0.6) : int(n * 0.8) - 1].reset_index()
test = full.ix[int(n * 0.8) : n - 1].reset_index()
step = len(data) / nsteps
ndata = len(data)
mvec = range(step, ndata + step, step)
test_features = features.extract_features(test)
data_error = []
cval_error = []
for i in range(len(mvec)):
m = mvec[i]
print "running for size", m
train = data.ix[0 : m - 1].reset_index()
fea = features.extract_features(train)
rf = RandomForestClassifier(n_estimators=50, verbose=0, compute_importances=False, n_jobs=5)
rf.fit(fea, train["OpenStatus"])
new_priors = cu.load_priors("train.csv")
old_priors = cu.compute_priors(train.OpenStatus)
# predict train
probs = rf.predict_proba(fea)
# probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(train)
score = multiclass_log_loss(y_true, probs)
data_error.append(score)
# predict cval
probs = rf.predict_proba(test_features)
# probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
cval_error.append(score)
return mvec, data_error, cval_error
def learning_curve():
n = 50000
nsteps = 10
full = cu.get_sample_data_frame(n)
data = full.ix[0:int(n*.6)-1].reset_index()
cval = full.ix[int(n*.6):int(n*.8)-1].reset_index()
test = full.ix[int(n*.8):n-1].reset_index()
step = len(data) / nsteps
ndata = len(data)
mvec = range(step, ndata + step, step)
test_features = features.extract_features(test)
data_error = []
cval_error = []
for i in range(len(mvec)):
m = mvec[i]
print 'running for size', m
train = data.ix[0:m-1].reset_index()
fea = features.extract_features(train)
rf = RandomForestClassifier(n_estimators=50, verbose=0, compute_importances=False, n_jobs=5)
rf.fit(fea, train["OpenStatus"])
new_priors = cu.load_priors('train.csv')
old_priors = cu.compute_priors(train.OpenStatus)
# predict train
probs = rf.predict_proba(fea)
#probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(train)
score = multiclass_log_loss(y_true, probs)
data_error.append(score)
# predict cval
probs = rf.predict_proba(test_features)
#probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
cval_error.append(score)
return mvec, data_error, cval_error
def measure_bayes(datasize=1000, testsize=500):
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
nbfd = features.naive_features(data)
nbft = features.naive_features(test)
nb = nltk.NaiveBayesClassifier.train(nbfd)
probs = []
for i in range(len(nbft)):
p = nb.prob_classify(nbft[i][0])
probs.append([p.prob(s) for s in p.samples()])
probs = np.array(probs)
new_priors = cu.load_priors("train.csv")
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, nb, y_true, probs
def measure_bayes(datasize=1000, testsize=500):
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
nbfd = features.naive_features(data)
nbft = features.naive_features(test)
nb = nltk.NaiveBayesClassifier.train(nbfd)
probs = []
for i in range(len(nbft)):
p = nb.prob_classify(nbft[i][0])
probs.append([p.prob(s) for s in p.samples()])
probs = np.array(probs)
new_priors = cu.load_priors('train.csv')
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, nb, y_true, probs
def measure_model(datasize=1000, testsize=500):
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
# data = full.ix[len(full)/4:].reset_index() # last n/4 * 3 records
# test = full.ix[:(len(full)/4)-1].reset_index() # first n/4 records
# data = cu.get_dataframe('train-sample.csv')
# test = cu.get_dataframe('public_leaderboard.csv')
fea = features.extract_features(data)
test_features = features.extract_features(test)
rf = RandomForestClassifier(n_estimators=50, verbose=2, compute_importances=True, n_jobs=5)
rf.fit(fea, data["OpenStatus"])
probs = rf.predict_proba(test_features)
new_priors = cu.load_priors("train.csv")
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, rf, fea
def measure_model(datasize=1000, testsize=500):
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
#data = full.ix[len(full)/4:].reset_index() # last n/4 * 3 records
#test = full.ix[:(len(full)/4)-1].reset_index() # first n/4 records
#data = cu.get_dataframe('train-sample.csv')
#test = cu.get_dataframe('public_leaderboard.csv')
fea = features.extract_features(data)
test_features = features.extract_features(test)
rf = RandomForestClassifier(n_estimators=50, verbose=2,
compute_importances=True, n_jobs=5)
rf.fit(fea, data["OpenStatus"])
probs = rf.predict_proba(test_features)
new_priors = cu.load_priors('train.csv')
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, rf, fea
def measure_svm(datasize=1000, testsize=500):
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
vocab = [w.strip() for w in file("vocab4.txt")][0:1000]
vidx = get_vocab_index_lookup(vocab)
print "extracting data features"
xdata = extract_svm_features(vidx, data)
print "extracting test features"
xtest = extract_svm_features(vidx, test)
labels = sorted(cu.labels)
ydata = data.OpenStatus.apply(labels.index).tolist()
model = svm.sparse.SVC(probability=True)
print "fitting model"
model.fit(xdata, ydata)
print "rest"
probs = model.predict_proba(xtest)
new_priors = cu.load_priors("train.csv")
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, model
def measure_svm(datasize=1000, testsize=500):
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
vocab = [w.strip() for w in file('vocab4.txt')][0:1000]
vidx = get_vocab_index_lookup(vocab)
print 'extracting data features'
xdata = extract_svm_features(vidx, data)
print 'extracting test features'
xtest = extract_svm_features(vidx, test)
labels = sorted(cu.labels)
ydata = data.OpenStatus.apply(labels.index).tolist()
model = svm.sparse.SVC(probability=True)
print 'fitting model'
model.fit(xdata, ydata)
print 'rest'
probs = model.predict_proba(xtest)
new_priors = cu.load_priors('train.csv')
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, model
def measure_lda(datasize=1000, testsize=500):
# The number of documents to analyze each iteration
batchsize = 100
# The total number of questions on Stack Overflow
D = 3.3e6
# The number of topics
K = 100
# How many documents to look at
documentstoanalyze = datasize / batchsize
# Our vocabulary
vocab = [w.strip() for w in file("./vocab2.txt")]
W = len(vocab)
# the data
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
# Initialize the algorithm with alpha=1/K, eta=1/K, tau_0=1024, kappa=0.7
lda = OnlineLDA(vocab, K, D, 1.0 / K, 1.0 / K, 1024.0, 0.7)
make_topic_columns(lda, data, K, D, batchsize)
make_topic_columns(lda, test, K, D, batchsize)
# data = full.ix[len(full)/4:].reset_index() # last n/4 * 3 records
# test = full.ix[:(len(full)/4)-1].reset_index() # first n/4 records
# data = cu.get_dataframe('train-sample.csv')
# test = cu.get_dataframe('public_leaderboard.csv')
fea = features.extract_features(data)
test_features = features.extract_features(test)
rf = RandomForestClassifier(n_estimators=50, verbose=2, compute_importances=True, n_jobs=5)
rf.fit(fea, data["OpenStatus"])
probs = rf.predict_proba(test_features)
new_priors = cu.load_priors("train.csv")
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, rf, fea, data, datagamma
def measure_lda(datasize=1000, testsize=500):
# The number of documents to analyze each iteration
batchsize = 100
# The total number of questions on Stack Overflow
D = 3.3e6
# The number of topics
K = 100
# How many documents to look at
documentstoanalyze = datasize / batchsize
# Our vocabulary
vocab = [w.strip() for w in file('./vocab2.txt')]
W = len(vocab)
# the data
data = cu.get_sample_data_frame(datasize)
test = cu.get_test_data_frame(testsize)
# Initialize the algorithm with alpha=1/K, eta=1/K, tau_0=1024, kappa=0.7
lda = OnlineLDA(vocab, K, D, 1./K, 1./K, 1024., 0.7)
make_topic_columns(lda, data, K, D, batchsize)
make_topic_columns(lda, test, K, D, batchsize)
#data = full.ix[len(full)/4:].reset_index() # last n/4 * 3 records
#test = full.ix[:(len(full)/4)-1].reset_index() # first n/4 records
#data = cu.get_dataframe('train-sample.csv')
#test = cu.get_dataframe('public_leaderboard.csv')
fea = features.extract_features(data)
test_features = features.extract_features(test)
rf = RandomForestClassifier(n_estimators=50, verbose=2, compute_importances=True, n_jobs=5)
rf.fit(fea, data["OpenStatus"])
probs = rf.predict_proba(test_features)
new_priors = cu.load_priors('train.csv')
old_priors = cu.compute_priors(data.OpenStatus)
probs = cu.cap_and_update_priors(old_priors, probs, new_priors, 0.001)
y_true = compute_y_true(test)
score = multiclass_log_loss(y_true, probs)
return score, rf, fea, data, datagamma