def greedyWeight(p1, weight1, p2, weight2_range, true_label):
best_weight2 = 1.0
best_logloss = 10
for weight2 in weight2_range:
p = (weight1 * p1 + weight2 * p2) / (weight1 + weight2)
logloss = computeLogloss(p, true_label)
#print( "w1: %s, w2: %s, logloss: %s" % (weight1, weight2, logloss) )
if logloss < best_logloss:
best_logloss, best_weight2 = logloss, weight2
return best_weight2
def greedyWeight(p1, weight1, p2, weight2_range, true_label):
best_weight2 = 1.0
best_logloss = 10
for weight2 in weight2_range:
p = (weight1 * p1 + weight2 * p2) / (weight1 + weight2)
logloss = computeLogloss(p, true_label)
#print( "w1: %s, w2: %s, logloss: %s" % (weight1, weight2, logloss) )
if logloss < best_logloss:
best_logloss, best_weight2 = logloss, weight2
return best_weight2
def greedyEnsemble(list_file, fout, mode, task, fins):
list_in = np.loadtxt(list_file, dtype=str)
true_label = np.asarray(list_in[:, 1], dtype="int")
numValid = true_label.shape[0]
numTest = 130400
p0_valid = pd.read_csv(fins[-1] + "_valid.csv", index_col=0)
numLabel = p0_valid.shape[1]
p_ens_valid = np.zeros((numValid, numLabel), dtype="float")
if task == "test":
p0_test = pd.read_csv(fins[-1] + "_test.csv", index_col=0)
p_ens_test = np.zeros((numTest, numLabel), dtype="float")
best_dict = defaultdict(lambda: 0)
fins_tmp = copy(fins)
if mode == "average":
# naive average
for f in fins:
this_p_valid = pd.read_csv(f + "_valid.csv",
index_col=0).values[-numValid:]
p_ens_valid += this_p_valid
if task == "test":
this_p_test = pd.read_csv(f + "_test.csv",
index_col=0).values[-numTest:]
p_ens_test += this_p_test
p_ens_valid /= len(fins)
if task == "test":
p_ens_test /= len(fins)
best_logloss = computeLogloss(p_ens_valid, true_label)
print("\nAverage result:")
print(("logloss: %s" % (best_logloss)))
elif mode == "greedy":
# greedy ensemble
best_logloss = 10
best_accuracy = 0
best_fin = None
first = True
while True:
for f in fins_tmp:
this_p_valid = pd.read_csv(f + "_valid.csv",
index_col=0).values[-numValid:]
if first:
w_ens, this_w = 0.0, 1.0
else:
w_ens = 1.0
this_w = greedyWeight(p_ens_valid, w_ens, this_p_valid,
np.arange(0.1, 2, 0.01), true_label)
# all the current prediction to the ensemble
tmp = (w_ens * p_ens_valid + this_w * this_p_valid) / (w_ens +
this_w)
logloss = computeLogloss(tmp, true_label)
accuracy = computeAccuracy(tmp, true_label)
if logloss < best_logloss:
best_logloss, best_accuracy, best_fin, best_fin_w = logloss, accuracy, f, this_w
if best_fin == None:
break
print(best_fin)
print(best_fin_w)
print(best_logloss)
print(best_accuracy)
best_dict[best_fin] += best_fin_w
# valid
this_p_valid = pd.read_csv(best_fin + "_valid.csv",
index_col=0).values[-numValid:]
p_ens_valid = (w_ens * p_ens_valid +
best_fin_w * this_p_valid) / (w_ens + best_fin_w)
# test
if task == "test":
this_p_test = pd.read_csv(best_fin + "_test.csv",
index_col=0).values[-numTest:]
p_ens_test = (w_ens * p_ens_test +
best_fin_w * this_p_test) / (w_ens + best_fin_w)
#fins_tmp.remove(best_fin)
best_fin = None
first = False
# report the best weights and the corresponding logloss found
print("\nGreedy ensemble result:")
print(("logloss: %s" % (best_logloss)))
print(("accuracy: %s" % (best_accuracy)))
for i, f in enumerate(fins, start=1):
if f not in best_dict:
print((" w%s=%s" % (i, 0)))
else:
print((" w%s=%s" % (i, best_dict[f])))
# the final ensemble
p_out_valid = pd.DataFrame(p_ens_valid,
columns=p0_valid.columns,
index=p0_valid.index)
p_out_valid.index.name = p0_valid.index.name
p_out_valid.to_csv(fout + "_[merge_nll" + str(np.round(best_logloss, 8)) +
"]_valid.csv")
# test
if task == "test":
p_out_test = pd.DataFrame(p_ens_test,
columns=p0_test.columns,
index=p0_test.index)
p_out_test.index.name = p0_test.index.name
p_out_test.to_csv(fout + "_[merge_nll" +
str(np.round(best_logloss, 8)) + "]_test.csv")
def greedyEnsemble(list_file, fout, mode, task, fins):
list_in = np.loadtxt(list_file, dtype=str)
true_label = np.asarray(list_in[:,1], dtype="int")
numValid = true_label.shape[0]
numTest = 130400
p0_valid = pd.read_csv(fins[-1] + "_valid.csv", index_col=0)
numLabel = p0_valid.shape[1]
p_ens_valid = np.zeros((numValid, numLabel), dtype="float")
if task == "test":
p0_test = pd.read_csv(fins[-1] + "_test.csv", index_col=0)
p_ens_test = np.zeros((numTest, numLabel), dtype="float")
best_dict = defaultdict(lambda : 0)
fins_tmp = copy(fins)
if mode == "average":
# naive average
for f in fins:
this_p_valid = pd.read_csv(f + "_valid.csv", index_col=0).values[-numValid:]
p_ens_valid += this_p_valid
if task == "test":
this_p_test = pd.read_csv(f + "_test.csv", index_col=0).values[-numTest:]
p_ens_test += this_p_test
p_ens_valid /= len(fins)
if task == "test":
p_ens_test /= len(fins)
best_logloss = computeLogloss(p_ens_valid, true_label)
print( "\nAverage result:" )
print( "logloss: %s" % (best_logloss) )
elif mode == "greedy":
# greedy ensemble
best_logloss = 10
best_accuracy = 0
best_fin = None
first = True
while True:
for f in fins_tmp:
this_p_valid = pd.read_csv(f + "_valid.csv", index_col=0).values[-numValid:]
if first:
w_ens, this_w = 0.0, 1.0
else:
w_ens = 1.0
this_w = greedyWeight(p_ens_valid, w_ens, this_p_valid, np.arange(0.1, 2, 0.01), true_label)
# all the current prediction to the ensemble
tmp = (w_ens * p_ens_valid + this_w * this_p_valid) / (w_ens + this_w)
logloss = computeLogloss(tmp, true_label)
accuracy = computeAccuracy(tmp, true_label)
if logloss < best_logloss:
best_logloss, best_accuracy, best_fin, best_fin_w = logloss, accuracy, f, this_w
if best_fin == None:
break
print best_fin
print best_fin_w
print best_logloss
print best_accuracy
best_dict[best_fin] += best_fin_w
# valid
this_p_valid = pd.read_csv(best_fin + "_valid.csv", index_col=0).values[-numValid:]
p_ens_valid = (w_ens * p_ens_valid + best_fin_w * this_p_valid) / (w_ens + best_fin_w)
# test
if task == "test":
this_p_test = pd.read_csv(best_fin + "_test.csv", index_col=0).values[-numTest:]
p_ens_test = (w_ens * p_ens_test + best_fin_w * this_p_test) / (w_ens + best_fin_w)
#fins_tmp.remove(best_fin)
best_fin = None
first = False
# report the best weights and the corresponding logloss found
print( "\nGreedy ensemble result:" )
print( "logloss: %s" % (best_logloss) )
print( "accuracy: %s" % (best_accuracy) )
for i,f in enumerate(fins, start=1):
if f not in best_dict:
print( " w%s=%s" % (i, 0) )
else:
print( " w%s=%s" % (i, best_dict[f]) )
# the final ensemble
p_out_valid = pd.DataFrame(p_ens_valid, columns=p0_valid.columns, index=p0_valid.index)
p_out_valid.index.name = p0_valid.index.name
p_out_valid.to_csv( fout + "_[merge_nll" + str(np.round(best_logloss, 8)) + "]_valid.csv" )
# test
if task == "test":
p_out_test = pd.DataFrame(p_ens_test, columns=p0_test.columns, index=p0_test.index)
p_out_test.index.name = p0_test.index.name
p_out_test.to_csv( fout + "_[merge_nll" + str(np.round(best_logloss, 8)) + "]_test.csv" )
