def main():
# Todo:
# 1. Automatically pick best model for every question.
# 4. Try new signals.
# 2. Try kappa as loss function.
# 3. Try better loss function.
N_EST = 150
boost_params = dict(min_samples_leaf=5, min_samples_split=5, max_depth=1, n_estimators=N_EST, learn_rate=.3)
list_of_signals = [G.num_words, G.is_crap, G.word_length, G.answer_length, G.num_sentences, G.choice]
ml_params.PlotNumberOfEstimators(9, list_of_signals, mods_for_training=range(4), boost_params=boost_params, steps=N_EST)
return
vs = []
for q in range(10):
if q == 9:
list_of_signals.append(G.choice)
dd = []
for msl in [3, 4, 5]:
for n_estimators in range(30, 50, 5):
for max_depth in range(1, 6):
varz = dict(min_samples_leaf=msl, min_samples_split=msl, n_estimators=n_estimators, max_depth=max_depth)
v = PredictForSignals(q, list_of_signals, boost_params=varz)
dd.append((-v, varz))
# vs.append()
#print kappa.mean_quadratic_weighted_kappa(vs)
return
# TODO: show difference between models!
GenerateBasicFeatures()
GenerateTrainingFeatures()
return
# TODO: try to integrate some other signal into "all zeroes" model
# TODO: Find most popular words for each question, try to use
# TODO: use different target function
'''
for id, elem in enumerate(G.num_words):
if elem < 3:
print id, G.question[id], G.score[id], G.other_score[id], G.answer[id]
'''
vars = {'min_word_cutoff': 3.,
'ans_len': .5,
'num_sent': 10., }
# TODO: compute scores for pub leaderboard.
#scores = ScoresForModel(VerySimple2, vars)
print 'Current'
for k in range(5):
CheckModel(models.Version0, util.FMod(k))
for k in range(5):
CheckModel(models.Version1, util.FMod(k))
return
data0 = []
data1 = []
data2 = []
data3 = []
ks = range(5, 100, 5)
for k in ks:
vars['min_word_cutoff'] = k
scores0 = RawScoresForModel(VerySimple0, vars, extra_filter=util.FMod(1))
data0.append(metrics.EvalPerQuestion(scores0, extra_filter=util.FMod(1)))
scores1 = RawScoresForModel(VerySimple1, vars, extra_filter=util.FMod(1))
data1.append(metrics.EvalPerQuestion(scores1, extra_filter=util.FMod(1)))
scores2 = RawScoresForModel(VerySimple2, vars, extra_filter=util.FMod(1))
data2.append(metrics.EvalPerQuestion(scores2, extra_filter=util.FMod(1)))
scores3 = RawScoresForModel(VerySimple3, vars, extra_filter=util.FMod(1))
data3.append(metrics.EvalPerQuestion(scores3, extra_filter=util.FMod(1)))
# print k, '%.2f' % (sum(v2) / len(v2)), ['%.3f' % v for v in v2]
cutoff = []
value = []
score = []
for q in range(10):
s0 = [(-data0[i][q], (k, 0)) for i, k in enumerate(ks)]
s1 = [(-data1[i][q], (k, 30)) for i, k in enumerate(ks)]
s2 = [(-data2[i][q], (k, 50)) for i, k in enumerate(ks)]
s3 = [(-data3[i][q], (k, 100)) for i, k in enumerate(ks)]
x = sorted(s0 + s1 + s2 + s3)
score.append(-x[0][0])
cutoff.append(x[0][1][0])
value.append(x[0][1][1])
print cutoff
print value
print kappa.mean_quadratic_weighted_kappa(score), util.PrintList(score)
def Eval(raw_scores, extra_filter=FTrue, only_questions=None): vals = EvalPerQuestion(raw_scores, extra_filter=extra_filter, only_questions=only_questions) known_vals = filter(lambda v: v != signal.UNKNOWN, vals) return kappa.mean_quadratic_weighted_kappa(known_vals)
