def ret(series, n_folds, clfparams, featureparams, aggregateparams,
refineparams, include, exclude, save_test_predictions,
save_oob_predictions, skip_cross_validation, _run):
data = TelstraData(include=include, exclude=exclude, **featureparams)
X, y, Xtest, ytest = data.get_train_test_features(**aggregateparams)
time = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
pred_cols = ['predict_{}'.format(i) for i in range(3)]
rrfparams = dict(refineparams)
rrfparams['n_prunings'] = 1
kf = StratifiedKFold(y.values, n_folds=n_folds, shuffle=True)
pred = pd.DataFrame(0., index=y.index, columns=pred_cols)
testpreds = [] # list for storing test predictions
i = 1
_run.info['loss'] = []
_run.info['trainloss'] = []
_run.info['best_pruning'] = []
for itrain, itest in kf:
Xtr, ytr, Xte, yte = data.get_train_test_features(
itrain, itest, **aggregateparams)
clf = ET(**clfparams)
clf.fit(Xtr, ytr)
rrf = RRF(clf, **rrfparams)
best_loss = 1000.
train_loss = 1000.
no_improvement_in = 0
for k in range(refineparams['n_prunings']):
try:
rrf.fit(Xtr, ytr) # fit and do 1 pruning
except IndexError as e:
print('IndexError') # sometimes I get an index error when an
# unfortunate tree gets cut down to the root
# we'll just stop and use best-so-far prediction in this case
break
loss2tr = multiclass_log_loss(ytr.values, rrf.predict_proba(Xtr))
loss2te = multiclass_log_loss(yte.values, rrf.predict_proba(Xte))
print("Fold {} Pruning {} mlogloss train: {:.4f}, test: {:.4f}".
format(i, k, loss2tr, loss2te))
if loss2te < best_loss: # performance is better
no_improvement_in = 0
best_loss = loss2te + 0. # save new best loss
# predict oof samples with new model
pred.iloc[itest, :] = rrf.predict_proba(Xte)
# predict test with new model
testpred = rrf.predict_proba(Xtest)
# record current train loss
train_loss = loss2tr + 0.
else:
no_improvement_in += 1
if no_improvement_in >= 5:
break
# Append current testpred to testpreds list
testpreds.append(
pd.DataFrame(testpred, index=ytest.index, columns=pred_cols))
# Save loss and train loss from current fold
_run.info['loss'].append(best_loss)
_run.info['trainloss'].append(train_loss)
_run.info['best_pruning'].append(k + 1)
print("Fold {} mlogloss train: {:.4f}, test: {:.4f}".format(
i, train_loss, best_loss))
i += 1
loss = multiclass_log_loss(y.values, pred.values)
_run.info['features'] = list(Xtr.columns)
filename = '{}_{}_cv_{:.4f}.csv'.format(series, time, loss)
pred.to_csv(filename, index_label='id')
# Optionally generate test predictions
testpred = sum(testpreds) / len(testpreds)
filename = '{}_test_{}_cv_{:.4f}.csv'.format(series, time, loss)
testpred.to_csv(filename, index_label='id')
return loss
def xgbrun(series, n_folds, clfparams, featureparams, num_trees,
early_stopping_rounds, verbose_eval, _seed, _run, aggregateparams,
include, exclude, save_test_predictions, save_oob_predictions,
skip_cross_validation):
clfparams['seed'] = _seed
data = TelstraData(include=include, exclude=exclude, **featureparams)
time = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
pred_cols = ['predict_{}'.format(i) for i in range(3)]
num_rounds = num_trees + 0
params = {"verbose_eval": verbose_eval}
if skip_cross_validation:
loss = 999.
else:
y = data.get_y()
kf = StratifiedKFold(y.values, n_folds=n_folds, shuffle=True)
pred = pd.DataFrame(0., index=y.index, columns=pred_cols)
i = 1
_run.info['loss'] = []
_run.info['trainloss'] = []
for itrain, itest in kf:
Xtr, ytr, Xte, yte = data.get_train_test_features(
itrain, itest, **aggregateparams)
dtrain = xgb.DMatrix(Xtr, ytr)
dvalid = xgb.DMatrix(Xte, yte)
params = {"verbose_eval": verbose_eval}
if (i == 1) and (early_stopping_rounds > 0):
watchlist = [(dtrain, 'train'), (dvalid, 'eval')]
params["evals"] = watchlist
params["early_stopping_rounds"] = early_stopping_rounds
gbm = xgb.train(clfparams, dtrain, num_rounds, **params)
if i == 1:
num_rounds = gbm.best_ntree_limit
_run.info['num_rounds'] = num_rounds
pred.iloc[itest, :] = gbm.predict(dvalid,
ntree_limit=num_rounds).reshape(
yte.shape[0], len(pred_cols))
predtrain = gbm.predict(dtrain, ntree_limit=num_rounds).reshape(
ytr.shape[0], len(pred_cols))
loss = multiclass_log_loss(yte, pred.iloc[itest].values)
trainloss = multiclass_log_loss(ytr, predtrain)
#print("Fold {:02d}: trainloss = {:.4f}, testloss = {:.4f}".format(i,trainloss, loss))
_run.info['loss'].append(loss)
_run.info['trainloss'].append(trainloss)
i += 1
loss = multiclass_log_loss(y, pred.values)
_run.info['features'] = list(Xtr.columns)
# Optionally save oob predictions
if save_oob_predictions:
filename = '{}_{}.csv'.format(series, time)
pred.to_csv(filename, index_label='id')
# Optionally generate test predictions
if save_test_predictions:
filename = '{}_test_{}.csv'.format(series, time)
Xtr, ytr, Xte, yte = data.get_train_test_features(**aggregateparams)
dtrain = xgb.DMatrix(Xtr, ytr)
dtest = xgb.DMatrix(Xte)
gbm = xgb.train(clfparams, dtrain, num_rounds, **params)
predtest = pd.DataFrame(gbm.predict(dtest).reshape(
yte.shape[0], len(pred_cols)),
index=yte.index,
columns=pred_cols)
predtest.to_csv(filename, index_label='id')
return loss