Update catboost

- added exp description optn
- renamed dataset description => feature description
- update total_features_config according to fe version (=v1.2)

catboost/dataset.py

@@ -2,12 +2,13 @@
 
 import pandas as pd
 
+from features import Features
 
 class GBMDataset:
-    def __init__(self, data_path, descript=None, feat_path="/opt/ml/input/code/catboost/feature_config.json"):
-        features = self._get_features_from(feat_path)
-        self.features = features["FEAT"]
-        self.cat_features = features["CAT_FEAT"]
+    def __init__(self, data_path, descript=None):
+        features = Features()
+        self.features = features.FEAT
+        self.cat_features = features.CAT_FEAT
         self.num_features = list(set(self.features) - set(self.cat_features))
         self.df = pd.read_csv(data_path)
 
@@ -26,8 +27,9 @@ def split_data(self, pseudo_labeling=False):
 
         train_valid_df = self.get_train_data(pseudo_labeling)
 
-        train = train_valid_df[train_valid_df.user == train_valid_df.user.shift(-1)]
-        valid = train_valid_df[train_valid_df.user != train_valid_df.user.shift(-1)]
+        valid = train_valid_df[(train_valid_df.user != train_valid_df.user.shift(-1)) &
+                               (train_valid_df["label"] == "train")]
+        train = train_valid_df[~train_valid_df.index.isin(valid.index)]
 
         X_train, y_train = train[self.features], train["answer"]
         X_valid, y_valid = valid[self.features], valid["answer"]

catboost/features.py

@@ -0,0 +1,34 @@
+class Features:
+    FEAT = ["user", "assessmentItemID", "testId", "tag", "category", "test", "item", 
+        "month", "day", "weekday", "hour", 
+        # "elapsed", "test_elapsed",
+        # "prev_elapsed", 
+        "prev_test_elapsed",
+
+        "user_correct_answer", "user_total_answer", "user_acc",
+
+        "user_category_correct_answer", "user_category_total_answer", "user_category_acc", 
+        # "user_tag_correct_answer", "user_tag_total_answer", "user_tag_acc",
+        "user_testId_correct_answer", "user_testId_total_answer", "user_testId_acc", 
+
+        "user_category_cum_telapsed", "user_category_mean_telapsed", 
+        # "user_tag_cum_telapsed", "user_tag_mean_telapsed", 
+        "user_testId_cum_telapsed", "user_testId_mean_telapsed",
+
+        "testId_answer_mean", "testId_test_elapsed_mean", 
+        # "testId_answer_sum",
+
+        # "tag_answer_mean", "tag_test_elapsed_mean", 
+        # "tag_answer_sum",
+
+        "assessmentItemID_answer_mean", "assessmentItemID_test_elapsed_mean", 
+        # "assessmentItemID_answer_sum",
+
+        "category_answer_mean", "category_test_elapsed_mean", 
+        # "category_answer_sum",
+
+        # "last_prob"
+    ]
+    CAT_FEAT = ["user", "assessmentItemID", "testId", "tag", "category", "test", "item", 
+                "month", "day", "weekday", "hour"
+                ]

catboost/main.py

@@ -1,6 +1,7 @@
 import argparse
 import os
 import random
+from sys import float_info
 
 import pandas as pd
 import numpy as np
@@ -17,13 +18,14 @@ def seed_everythings(seed):
 def main(args):
     # dataset setting
     print(">>> load dataset...")
-    dataset = GBMDataset(args.data_dir, descript=args.dataset_descript, feat_path=args.feat_path)
+    dataset = GBMDataset(args.data_dir, descript=args.feature_descript)
     X_train, X_valid, y_train, y_valid = dataset.split_data()
-    print("<<< done!\n")
 
-    if dataset.descript:
-        args.output_dir = os.path.join(args.output_dir, dataset.descript)
+    print(f"# of features: {len(dataset.features)}")
+    args.output_dir = os.path.join(args.output_dir, dataset.descript + "_" + args.exp_descript)
     os.makedirs(args.output_dir, exist_ok=True)
+    print(f"The result will be saved in {args.output_dir}")
+    print("<<< done!\n")
 
     # model setting & run
     print(">>>load model with configurations...")
@@ -36,7 +38,7 @@ def main(args):
         verbose=args.verbose,
     )
     # save model feature information
-    model.save_features(dataset.features, dataset.cat_features)
+    model.save_features(dataset.features, dataset.cat_features, args.feature_descript)
     if args.save_model:
         model.save_model(args.output_dir)
 
@@ -50,17 +52,19 @@ def main(args):
     parser = argparse.ArgumentParser()
 
     parser.add_argument("--data_dir", type=str, default="/opt/ml/input/data/feature_engineering/processed_data.csv")
-    parser.add_argument("--feat_path", type=str, default="./feature_config.json")
     parser.add_argument("--inference_dir", type=str, default="/opt/ml/input/data/sample_submission.csv")
-    parser.add_argument("--dataset_descript", type=str, default=None)
+    parser.add_argument("--feature_descript", type=str, default=None)
 
     parser.add_argument("--output_dir", type=str, default="/opt/ml/output/catboost")
+    parser.add_argument("--lr",type=float, default=None)
     parser.add_argument("--iteration", type=int, default=10000)
-    parser.add_argument("--early_stopping", type=int, default=300)
+    parser.add_argument("--early_stopping", type=int, default=1000)
     parser.add_argument("--seed", type=int, default=2022)
     parser.add_argument("--verbose", type=int, default=100)
     parser.add_argument("--save_model", type=bool, default=False)
 
+    parser.add_argument("--exp_descript", type=str, default=None)
+
     args = parser.parse_args()
 
     seed_everythings(args.seed)

catboost/model.py

@@ -14,6 +14,7 @@ def __init__(self, args, output_dir):
                                         eval_metric="AUC",
                                         early_stopping_rounds=args.early_stopping,
                                         train_dir=self.output_dir,
+                                        learning_rate=args.lr,
 
                                         task_type="GPU",
                                         devices="0")
@@ -24,11 +25,12 @@ def fit(self, X, y, cat_features, eval_set, verbose=100):
                        eval_set=eval_set,
                        verbose=verbose)
 
-    def save_features(self, features, cat_features):
+    def save_features(self, features, cat_features, feature_descript):
         # feature names
         with open(os.path.join(self.output_dir, "features.json"), "w") as f:
             feature_dict = {
                 "num_feats": len(features),
+                "description": feature_descript,
                 "FEAT": features,
                 "CAT_FEAT": cat_features
             }
@@ -47,4 +49,4 @@ def save_features(self, features, cat_features):
         plt.savefig(os.path.join(self.output_dir, "feature_importances.png"))
 
     def inference(self, X_test):
-        return self.model.predict_proba(X_test)[:, 1]
+        return self.model.predict_proba(X_test)[:, 1]