def _test_elbs(model):
elb_pred = model.predict(elb_data)
score = score_util.ScoreReport(elb_y, elb_pred)
elb_cv_results.append(score.abs_99)
print(score)
print(f"elb yield stats: {_print_stats(elb_y)}")
print(f"elb prediction stats: {_print_stats(elb_pred)}")
def _test_elbs(model):
if elb_data.num_row() == 0:
print("no test elbs found, skipping scoring")
return
elb_pred = model.predict(elb_data)
score = score_util.ScoreReport(elb_y, elb_pred)
elb_cv_results.append(score.abs_99)
print(score)
print(f"elb yield stats: {_print_stats(elb_y)}")
print(f"elb prediction stats: {_print_stats(elb_pred)}")
def _eval(pred, d: xgb.DMatrix):
_scr = score_util.ScoreReport(d.get_label(), pred)
return '3std_100x', int(round(_scr.abs_99 * 100, 0))
return '3std_100x', int(round(_scr.abs_99 * 100, 0))
print(f"training xgb model")
model: xgb.Booster = xgb.train(
default_xgb_params,
train,
num_boost_round=xgb_n_rounds,
early_stopping_rounds=xgb_early_stopping,
evals=eval_list,
evals_result=eval_result,
feval=_eval,
verbose_eval=verbose_eval)
predictions = model.predict(test,
ntree_limit=model.best_ntree_limit)
score = score_util.ScoreReport(test.get_label(), predictions)
_results = grid_cv_results[i]
_results.add(score.abs_99,
iteration=model.best_iteration,
mean=score.abs_mean,
std_dev=score.abs_std,
y_min=np.min(train.get_label()),
y_max=np.max(train.get_label()),
pred_min=np.min(predictions),
pred_max=np.max(predictions))
print(f"""
GRID SEARCH RESULT: ({curr_model_training} of total: {total_model_trainings})
{test_run_idx + 1}, {idx + 1}, {i + 1} of {len(cv_params) - 1}, {p}
training yield stats: {_print_stats(train.get_label())}
pred yield stats: {_print_stats(predictions)}""")
epoch = 0
epoch_scores: List[score_util.ScoreReport] = []
test_diff = float('inf')
while epoch < epochs_max:
print(f"model current estimators: {model.n_estimators}")
n_estimators = model.n_estimators + n_threads
print(f"setting estimators: {n_estimators}")
model.n_estimators = n_estimators
print("fitting model")
model.fit(train, train_label)
print("scoring")
predictions = model.predict(test)
score = score_util.ScoreReport(test_label, predictions, store_predictions=False)
print(f"""
EPOCH RESULT: (training {curr_model_training} of total: {total_model_trainings})
epoch: {epoch}
n_estimators: {n_estimators}""")
print(f"abs_99: {score.abs_99:.2f}")
mem_util.print_mem_usage()
if epoch > 0:
test_diff = epoch_scores[epoch - 1].abs_std_3 - score.abs_std_3
print(f"score diff: {test_diff:.2f}")
epoch += 1
if test_diff < train_tol and epoch >= epochs_min:
print(f"ending early. test diff: {test_diff:.2f}, epochs: {epoch}")
# def _run_dummies():
column_categories = categorical_util.get_categories_lookup(dtrain)
categorical_util.set_categories(dtrain, column_categories)
categorical_util.set_categories(dtest, column_categories)
dtrain_y = dtrain.pop('Dry_Yield')
dtrain = categorical_util.encode_dummies(df)
dummy_cols = dtrain.columns
print(dummy_cols)
dtrain = dtrain.to_coo()
dtest_y = dtest.pop('Dry_Yield')
dtest = pandas.get_dummies(dtest,
sparse=True,
drop_first=True,
dummy_na=False,
prefix_sep='__DUMMY__').to_sparse().to_coo()
print(f"dmatrix sizes: {dtrain.shape[1]}, {dtest.shape[1]}")
dtrain: xgb.DMatrix = xgb.DMatrix(dtrain, dtrain_y)
dtest: xgb.DMatrix = xgb.DMatrix(dtest, label=dtest_y)
model: xgb.Booster = xgb.train({'max_depth': 2}, dtrain, num_boost_round=2)
pred = model.predict(dtest)
scr = score_util.ScoreReport(dtest_y, pred)
print(scr)
# _run_dummies()
column_categories = categorical_util.encode_categories(data)
dummy_enc = categorical_util.DummyEncoder(data.columns, column_categories)
print("Fitting dummy enc")
data: np.ndarray = dummy_enc.fit_transform(data.as_matrix())
kf_outer = GroupKFold()
split = next(kf_outer.split(data, groups=df_year_id))
train_idx, test_idx = split
train, train_y = data[train_idx], data_label[train_idx]
test, test_y = data[test_idx], data_label[test_idx]
model = XGBRegressor(max_depth=5, n_estimators=100, silent=False, n_jobs=2)
model.fit(train, train_y)
scr = score_util.ScoreReport(test_y, model.predict(test))
print(scr)
model = RandomForestRegressor(verbose=99, n_estimators=50, n_jobs=2)
model.fit(train, train_y)
scr = score_util.ScoreReport(test_y, model.predict(test))
print(scr)
model = ExtraTreesRegressor(verbose=99, n_estimators=50, n_jobs=2)
model.fit(train, train_y)
scr = score_util.ScoreReport(test_y, model.predict(test))
print(scr)
model = SVR(degree=5, verbose=99)
model.fit(train, train_y)
scr = score_util.ScoreReport(test_y, model.predict(test))
def _score_cv(est: GridSearchCV, X, y):
if est.best_estimator_ not in cv_results:
cv_results.append(est.best_estimator_)
predictions = est.predict(X)
return score_util.ScoreReport(y, predictions).abs_99
def _score_grid_search(est: Pipeline, X, y):
predictions = est.predict(X)
_print_mem_usage()
return score_util.ScoreReport(y, predictions).abs_99