def main(params: dict,
output_dir: str):
import mlflow
print("start params={}".format(params))
logger = get_logger()
df = pd.read_pickle("../input/riiid-test-answer-prediction/train_merged.pickle")
# df = pd.read_pickle("../input/riiid-test-answer-prediction/split10/train_0.pickle").sort_values(["user_id", "timestamp"]).reset_index(drop=True)
if is_debug:
df = df.head(30000)
df["prior_question_had_explanation"] = df["prior_question_had_explanation"].fillna(-1)
column_config = {
("content_id", "content_type_id"): {"type": "category"},
"user_answer": {"type": "leakage_feature"},
"answered_correctly": {"type": "leakage_feature"},
"part": {"type": "category"},
"prior_question_elapsed_time_bin300": {"type": "category"},
"duration_previous_content_bin300": {"type": "category"},
"prior_question_had_explanation": {"type": "category"},
"rating_diff_content_user_id": {"type": "numeric"}
}
with open(f"{output_dir}/transformer_param.json", "w") as f:
json.dump(params, f)
if is_make_feature_factory:
# feature factory
feature_factory_dict = {"user_id": {}}
feature_factory_dict["user_id"]["DurationPreviousContent"] = DurationPreviousContent(is_partial_fit=True)
feature_factory_dict["user_id"]["ElapsedTimeBinningEncoder"] = ElapsedTimeBinningEncoder()
feature_factory_dict["user_id"]["UserContentRateEncoder"] = UserContentRateEncoder(rate_func="elo",
column="user_id")
feature_factory_manager = FeatureFactoryManager(feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=1,
model_id=model_id,
load_feature=not is_debug,
save_feature=not is_debug)
ff_for_transformer = FeatureFactoryForTransformer(column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
ff_for_transformer.make_dict(df=df)
if is_debug:
df = df.head(10000)
df = df.sort_values(["user_id", "timestamp"]).reset_index(drop=True)
feature_factory_manager.fit(df)
df = feature_factory_manager.all_predict(df)
for dicts in feature_factory_manager.feature_factory_dict.values():
for factory in dicts.values():
factory.logger = None
feature_factory_manager.logger = None
with open(f"{output_dir}/feature_factory_manager.pickle", "wb") as f:
pickle.dump(feature_factory_manager, f)
ff_for_transformer.fit(df)
ff_for_transformer.logger = None
with open(f"{output_dir}/feature_factory_manager_for_transformer.pickle", "wb") as f:
pickle.dump(ff_for_transformer, f)
def main(params: dict, output_dir: str):
import mlflow
print("start params={}".format(params))
model_id = "train_0"
logger = get_logger()
# df = pd.read_pickle("../input/riiid-test-answer-prediction/train_merged.pickle")
df = pd.read_pickle(
"../input/riiid-test-answer-prediction/split10/train_0.pickle"
).sort_values(["user_id", "timestamp"]).reset_index(drop=True)
if is_debug:
df = df.head(30000)
df["prior_question_had_explanation"] = df[
"prior_question_had_explanation"].fillna(-1)
df["answered_correctly"] = df["answered_correctly"].replace(-1, np.nan)
column_config = {
("content_id", "content_type_id"): {
"type": "category"
},
"user_answer": {
"type": "leakage_feature"
},
"answered_correctly": {
"type": "leakage_feature"
},
"part": {
"type": "category"
},
"prior_question_elapsed_time_bin300": {
"type": "category"
},
"duration_previous_content_bin300": {
"type": "category"
},
"prior_question_had_explanation": {
"type": "category"
},
"rating_diff_content_user_id": {
"type": "numeric"
},
"task_container_id_bin300": {
"type": "category"
},
"previous_answer_index_content_id": {
"type": "category"
},
"previous_answer_content_id": {
"type": "category"
},
"timediff-elapsedtime_bin500": {
"type": "category"
},
"timedelta_log10": {
"type": "category"
}
}
if not load_pickle or is_debug:
feature_factory_dict = {"user_id": {}}
feature_factory_dict["user_id"][
"DurationPreviousContent"] = DurationPreviousContent(
is_partial_fit=True)
feature_factory_dict["user_id"][
"ElapsedTimeBinningEncoder"] = ElapsedTimeBinningEncoder()
feature_factory_dict["user_id"][
"UserContentRateEncoder"] = UserContentRateEncoder(
rate_func="elo", column="user_id")
feature_factory_dict["user_id"]["PreviousAnswer2"] = PreviousAnswer2(
groupby="user_id",
column="content_id",
is_debug=is_debug,
model_id=model_id,
n=300)
feature_factory_dict["user_id"][
"StudyTermEncoder2"] = StudyTermEncoder2(is_partial_fit=True)
feature_factory_dict["user_id"][
f"MeanAggregatorStudyTimebyUserId"] = MeanAggregator(
column="user_id", agg_column="study_time", remove_now=False)
feature_factory_dict["user_id"][
"ElapsedTimeMeanByContentIdEncoder"] = ElapsedTimeMeanByContentIdEncoder(
)
feature_factory_dict["post"] = {
"DurationFeaturePostProcess": DurationFeaturePostProcess()
}
feature_factory_manager = FeatureFactoryManager(
feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=1,
model_id=model_id,
load_feature=not is_debug,
save_feature=not is_debug)
print("all_predict")
df = feature_factory_manager.all_predict(df)
def f(x):
x = x // 1000
if x < -100:
return -100
if x > 400:
return 400
return x
df["task_container_id_bin300"] = [
x if x < 300 else 300 for x in df["task_container_id"]
]
df["timediff-elapsedtime_bin500"] = [
f(x) for x in df["timediff-elapsedtime"].values
]
df["timedelta_log10"] = np.log10(
df["duration_previous_content"].values)
df["timedelta_log10"] = df["timedelta_log10"].replace(
-np.inf, -1).replace(np.inf, -1).fillna(-1).astype("int8")
df = df[[
"user_id", "content_id", "content_type_id", "part", "user_answer",
"answered_correctly", "prior_question_elapsed_time_bin300",
"duration_previous_content_bin300",
"prior_question_had_explanation", "rating_diff_content_user_id",
"task_container_id_bin300", "previous_answer_index_content_id",
"previous_answer_content_id", "row_id",
"timediff-elapsedtime_bin500", "timedelta_log10"
]]
print(df.head(10))
print("data preprocess")
ff_for_transformer = FeatureFactoryForTransformer(
column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
ff_for_transformer.make_dict(df=df)
n_skill = len(ff_for_transformer.embbed_dict[("content_id",
"content_type_id")])
if not load_pickle or is_debug:
df_val_row = pd.read_feather(
"../../riiid_takoi/notebook/fe/validation_row_id.feather").head(
len(df))
if is_debug:
df_val_row = df_val_row.head(3000)
df_val_row["is_val"] = 1
df = pd.merge(df, df_val_row, how="left", on="row_id")
df["is_val"] = df["is_val"].fillna(0)
print(df["is_val"].value_counts())
w_df = df[df["is_val"] == 0]
w_df["group"] = (
w_df.groupby("user_id")["user_id"].transform("count") -
w_df.groupby("user_id").cumcount()) // params["max_seq"]
w_df["user_id"] = w_df["user_id"].astype(
str) + "_" + w_df["group"].astype(str)
group = ff_for_transformer.all_predict(w_df)
dataset_train = SAKTDataset(group,
n_skill=n_skill,
max_seq=params["max_seq"])
del w_df
gc.collect()
ff_for_transformer = FeatureFactoryForTransformer(
column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
if not load_pickle or is_debug:
group = ff_for_transformer.all_predict(df[df["content_type_id"] == 0])
dataset_val = SAKTDataset(group,
is_test=True,
n_skill=n_skill,
max_seq=params["max_seq"])
os.makedirs("../input/feature_engineering/model256", exist_ok=True)
if not is_debug and not load_pickle:
with open(f"../input/feature_engineering/model256/train.pickle",
"wb") as f:
pickle.dump(dataset_train, f)
with open(f"../input/feature_engineering/model256/val.pickle",
"wb") as f:
pickle.dump(dataset_val, f)
if not is_debug and load_pickle:
with open(f"../input/feature_engineering/model256/train.pickle",
"rb") as f:
dataset_train = pickle.load(f)
with open(f"../input/feature_engineering/model256/val.pickle",
"rb") as f:
dataset_val = pickle.load(f)
print("loaded!")
dataloader_train = DataLoader(dataset_train,
batch_size=params["batch_size"],
shuffle=True)
dataloader_val = DataLoader(dataset_val,
batch_size=params["batch_size"],
shuffle=False)
model = SAKTModel(n_skill,
embed_dim=params["embed_dim"],
max_seq=params["max_seq"],
dropout=dropout,
cont_emb=params["cont_emb"])
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = AdamW(
optimizer_grouped_parameters,
lr=params["lr"],
weight_decay=0.01,
)
num_train_optimization_steps = int(len(dataloader_train) * 20)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=params["num_warmup_steps"],
num_training_steps=num_train_optimization_steps)
criterion = nn.BCEWithLogitsLoss()
model.to(device)
criterion.to(device)
for epoch in range(epochs):
loss, acc, auc, auc_val = train_epoch(model, dataloader_train,
dataloader_val, optimizer,
criterion, scheduler, epoch,
device)
print("epoch - {} train_loss - {:.3f} auc - {:.4f} auc-val: {:.4f}".
format(epoch, loss, auc, auc_val))
preds = []
labels = []
with torch.no_grad():
for item in tqdm(dataloader_val):
label = item["label"].to(device).float()
output = model(item, device)
preds.extend(torch.nn.Sigmoid()(
output[:, -1]).view(-1).data.cpu().numpy().tolist())
labels.extend(label[:, -1].view(-1).data.cpu().numpy().tolist())
auc_transformer = roc_auc_score(labels, preds)
print("single transformer: {:.4f}".format(auc_transformer))
df_oof = pd.DataFrame()
# df_oof["row_id"] = df.loc[val_idx].index
print(len(dataloader_val))
print(len(preds))
df_oof["predict"] = preds
df_oof["target"] = labels
df_oof.to_csv(f"{output_dir}/transformers1.csv", index=False)
"""
df_oof2 = pd.read_csv("../output/ex_237/20201213110353/oof_train_0_lgbm.csv")
df_oof2.columns = ["row_id", "predict_lgbm", "target"]
df_oof2 = pd.merge(df_oof, df_oof2, how="inner")
auc_lgbm = roc_auc_score(df_oof2["target"].values, df_oof2["predict_lgbm"].values)
print("lgbm: {:.4f}".format(auc_lgbm))
print("ensemble")
max_auc = 0
max_nn_ratio = 0
for r in np.arange(0, 1.05, 0.05):
auc = roc_auc_score(df_oof2["target"].values, df_oof2["predict_lgbm"].values*(1-r) + df_oof2["predict"].values*r)
print("[nn_ratio: {:.2f}] AUC: {:.4f}".format(r, auc))
if max_auc < auc:
max_auc = auc
max_nn_ratio = r
print(len(df_oof2))
"""
if not is_debug:
mlflow.start_run(experiment_id=10, run_name=os.path.basename(__file__))
for key, value in params.items():
mlflow.log_param(key, value)
mlflow.log_metric("auc_val", auc_transformer)
mlflow.end_run()
torch.save(model.state_dict(), f"{output_dir}/transformers.pth")
del model
torch.cuda.empty_cache()
with open(f"{output_dir}/transformer_param.json", "w") as f:
json.dump(params, f)
if is_make_feature_factory:
# feature factory
feature_factory_dict = {"user_id": {}}
feature_factory_dict["user_id"][
"DurationPreviousContent"] = DurationPreviousContent(
is_partial_fit=True)
feature_factory_dict["user_id"][
"ElapsedTimeBinningEncoder"] = ElapsedTimeBinningEncoder()
feature_factory_manager = FeatureFactoryManager(
feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=1,
model_id="all",
load_feature=not is_debug,
save_feature=not is_debug)
ff_for_transformer = FeatureFactoryForTransformer(
column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
df = pd.read_pickle(
"../input/riiid-test-answer-prediction/train_merged.pickle")
if is_debug:
df = df.head(10000)
df = df.sort_values(["user_id", "timestamp"]).reset_index(drop=True)
feature_factory_manager.fit(df)
df = feature_factory_manager.all_predict(df)
for dicts in feature_factory_manager.feature_factory_dict.values():
for factory in dicts.values():
factory.logger = None
feature_factory_manager.logger = None
with open(f"{output_dir}/feature_factory_manager.pickle", "wb") as f:
pickle.dump(feature_factory_manager, f)
ff_for_transformer.fit(df)
ff_for_transformer.logger = None
with open(
f"{output_dir}/feature_factory_manager_for_transformer.pickle",
"wb") as f:
pickle.dump(ff_for_transformer, f)
def main(params: dict,
output_dir: str):
import mlflow
print("start params={}".format(params))
logger = get_logger()
df = pd.read_pickle("../input/riiid-test-answer-prediction/train_merged.pickle")
# df = pd.read_pickle("../input/riiid-test-answer-prediction/split10/train_0.pickle").sort_values(["user_id", "timestamp"]).reset_index(drop=True)
if is_debug:
df = df.head(30000)
df["prior_question_had_explanation"] = df["prior_question_had_explanation"].fillna(-1)
column_config = {
("content_id", "content_type_id"): {"type": "category"},
"user_answer": {"type": "category"},
"part": {"type": "category"},
"prior_question_elapsed_time_bin300": {"type": "category"},
"duration_previous_content_bin300": {"type": "category"},
"prior_question_had_explanation": {"type": "category"},
}
if not load_pickle or is_debug:
feature_factory_dict = {"user_id": {}}
feature_factory_dict["user_id"]["DurationPreviousContent"] = DurationPreviousContent()
feature_factory_dict["user_id"]["ElapsedTimeBinningEncoder"] = ElapsedTimeBinningEncoder()
feature_factory_manager = FeatureFactoryManager(feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=1,
model_id=model_id,
load_feature=not is_debug,
save_feature=not is_debug)
print("all_predict")
df = feature_factory_manager.all_predict(df)
df = df[["user_id", "content_id", "content_type_id", "part", "user_answer", "answered_correctly",
"prior_question_elapsed_time_bin300", "duration_previous_content_bin300",
"prior_question_had_explanation"]].replace(-99, -1)
print(df.head(10))
print("data preprocess")
train_idx = []
val_idx = []
np.random.seed(0)
for _, w_df in df[df["content_type_id"] == 0].groupby("user_id"):
if np.random.random() < 0.01:
# all val
val_idx.extend(w_df.index.tolist())
else:
train_num = int(len(w_df) * 0.95)
train_idx.extend(w_df[:train_num].index.tolist())
val_idx.extend(w_df[train_num:].index.tolist())
ff_for_transformer = FeatureFactoryForTransformer(column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
ff_for_transformer.make_dict(df=df)
n_skill = len(ff_for_transformer.embbed_dict[("content_id", "content_type_id")])
if not load_pickle or is_debug:
df["is_val"] = 0
df["is_val"].loc[val_idx] = 1
w_df = df[df["is_val"] == 0]
w_df["group"] = (w_df.groupby("user_id")["user_id"].transform("count") - w_df.groupby("user_id").cumcount()) // params["max_seq"]
w_df["user_id"] = w_df["user_id"].astype(str) + "_" + w_df["group"].astype(str)
group = ff_for_transformer.all_predict(w_df)
dataset_train = SAKTDataset(group,
n_skill=n_skill,
max_seq=params["max_seq"])
del w_df
gc.collect()
ff_for_transformer = FeatureFactoryForTransformer(column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
if not load_pickle or is_debug:
group = ff_for_transformer.all_predict(df[df["content_type_id"] == 0])
dataset_val = SAKTDataset(group,
is_test=True,
n_skill=n_skill,
max_seq=params["max_seq"])
os.makedirs("../input/feature_engineering/model106_all", exist_ok=True)
if not is_debug and not load_pickle:
with open(f"../input/feature_engineering/model106_all/train.pickle", "wb") as f:
pickle.dump(dataset_train, f)
with open(f"../input/feature_engineering/model106_all/val.pickle", "wb") as f:
pickle.dump(dataset_val, f)
if not is_debug and load_pickle:
with open(f"../input/feature_engineering/model106_all/train.pickle", "rb") as f:
dataset_train = pickle.load(f)
with open(f"../input/feature_engineering/model106_all/val.pickle", "rb") as f:
dataset_val = pickle.load(f)
print("loaded!")
dataloader_train = DataLoader(dataset_train, batch_size=params["batch_size"], shuffle=True, num_workers=1)
dataloader_val = DataLoader(dataset_val, batch_size=params["batch_size"], shuffle=False, num_workers=1)
model = SAKTModel(n_skill, embed_dim=params["embed_dim"], max_seq=params["max_seq"], dropout=dropout)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
optimizer = AdamW(optimizer_grouped_parameters,
lr=params["lr"],
weight_decay=0.01,
)
num_train_optimization_steps = int(len(dataloader_train) * epochs)
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps=params["num_warmup_steps"],
num_training_steps=num_train_optimization_steps)
criterion = nn.BCEWithLogitsLoss()
model.to(device)
criterion.to(device)
for epoch in range(epochs):
loss, acc, auc, auc_val = train_epoch(model, dataloader_train, dataloader_val, optimizer, criterion, scheduler, device)
print("epoch - {} train_loss - {:.3f} auc - {:.4f} auc-val: {:.4f}".format(epoch, loss, auc, auc_val))
preds = []
labels = []
with torch.no_grad():
for item in tqdm(dataloader_val):
x = item["x"].to(device).long()
target_id = item["target_id"].to(device).long()
part = item["part"].to(device).long()
label = item["label"].to(device).float()
elapsed_time = item["elapsed_time"].to(device).long()
duration_previous_content = item["duration_previous_content"].to(device).long()
prior_question_had_explanation = item["prior_q"].to(device).long()
user_answer = item["user_answer"].to(device).long()
output = model(x, target_id, part, elapsed_time,
duration_previous_content, prior_question_had_explanation, user_answer)
preds.extend(torch.nn.Sigmoid()(output[:, -1]).view(-1).data.cpu().numpy().tolist())
labels.extend(label[:, -1].view(-1).data.cpu().numpy().tolist())
auc_transformer = roc_auc_score(labels, preds)
print("single transformer: {:.4f}".format(auc_transformer))
df_oof = pd.DataFrame()
# df_oof["row_id"] = df.loc[val_idx].index
print(len(dataloader_val))
print(len(preds))
df_oof["predict"] = preds
df_oof["target"] = labels
df_oof.to_csv(f"{output_dir}/transformers1.csv", index=False)
"""
df_oof2 = pd.read_csv("../output/ex_237/20201213110353/oof_train_0_lgbm.csv")
df_oof2.columns = ["row_id", "predict_lgbm", "target"]
df_oof2 = pd.merge(df_oof, df_oof2, how="inner")
auc_lgbm = roc_auc_score(df_oof2["target"].values, df_oof2["predict_lgbm"].values)
print("lgbm: {:.4f}".format(auc_lgbm))
print("ensemble")
max_auc = 0
max_nn_ratio = 0
for r in np.arange(0, 1.05, 0.05):
auc = roc_auc_score(df_oof2["target"].values, df_oof2["predict_lgbm"].values*(1-r) + df_oof2["predict"].values*r)
print("[nn_ratio: {:.2f}] AUC: {:.4f}".format(r, auc))
if max_auc < auc:
max_auc = auc
max_nn_ratio = r
print(len(df_oof2))
"""
if not is_debug:
mlflow.start_run(experiment_id=10,
run_name=os.path.basename(__file__))
for key, value in params.items():
mlflow.log_param(key, value)
mlflow.log_metric("auc_val", auc_transformer)
mlflow.end_run()
torch.save(model.state_dict(), f"{output_dir}/transformers.pth")
del model
torch.cuda.empty_cache()
with open(f"{output_dir}/transformer_param.json", "w") as f:
json.dump(params, f)
if is_make_feature_factory:
# feature factory
feature_factory_dict = {"user_id": {}}
feature_factory_dict["user_id"]["DurationPreviousContent"] = DurationPreviousContent(is_partial_fit=True)
feature_factory_dict["user_id"]["ElapsedTimeBinningEncoder"] = ElapsedTimeBinningEncoder()
feature_factory_manager = FeatureFactoryManager(feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=1,
model_id="all",
load_feature=not is_debug,
save_feature=not is_debug)
ff_for_transformer = FeatureFactoryForTransformer(column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
df = pd.read_pickle("../input/riiid-test-answer-prediction/train_merged.pickle")
if is_debug:
df = df.head(10000)
df = df.sort_values(["user_id", "timestamp"]).reset_index(drop=True)
feature_factory_manager.fit(df)
df = feature_factory_manager.all_predict(df)
for dicts in feature_factory_manager.feature_factory_dict.values():
for factory in dicts.values():
factory.logger = None
feature_factory_manager.logger = None
with open(f"{output_dir}/feature_factory_manager.pickle", "wb") as f:
pickle.dump(feature_factory_manager, f)
ff_for_transformer.fit(df)
ff_for_transformer.logger = None
with open(f"{output_dir}/feature_factory_manager_for_transformer.pickle", "wb") as f:
pickle.dump(ff_for_transformer, f)
CategoryLevelEncoder(groupby_column="user_id",
agg_column="user_count_bin",
categories=[0, 1, 2, 3, 4, 5])
feature_factory_dict["prior_question_elapsed_time"] = {
"PriorQuestionElapsedTimeBinningEncoder": PriorQuestionElapsedTimeBinningEncoder()
}
feature_factory_dict[("part", "prior_question_elapsed_time_bin")] = {
"CountEncoder": CountEncoder(column=["part", "prior_question_elapsed_time_bin"]),
"TargetEncoder": TargetEncoder(column=["part", "prior_question_elapsed_time_bin"])
}
feature_factory_manager = FeatureFactoryManager(feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=10)
df = feature_factory_manager.all_predict(df)
os.makedirs(output_dir, exist_ok=True)
params = {
'objective': 'binary',
'num_leaves': 32,
'min_data_in_leaf': 15, # 42,
'max_depth': -1,
'learning_rate': 0.3,
'boosting': 'gbdt',
'bagging_fraction': 0.7, # 0.5,
'feature_fraction': 0.3,
'bagging_seed': 0,
'reg_alpha': 5, # 1.728910519108444,
'reg_lambda': 5,
'random_state': 0,
'metric': 'auc',
def main(params: dict,
output_dir: str):
import mlflow
print("start params={}".format(params))
model_id = "all"
logger = get_logger()
column_config = {
("content_id", "content_type_id"): {"type": "category", "dtype": np.int16},
"user_answer": {"type": "leakage_feature", "dtype": np.int8},
"answered_correctly": {"type": "leakage_feature", "dtype": np.int8},
"part": {"type": "category", "dtype": np.int8},
"prior_question_elapsed_time_bin300": {"type": "category", "dtype": np.int16},
"duration_previous_content_bin300": {"type": "category", "dtype": np.int16},
"prior_question_had_explanation": {"type": "category", "dtype": np.int8},
"rating_diff_content_user_id": {"type": "numeric", "dtype": np.float16},
"task_container_id_bin300": {"type": "category", "dtype": np.int16},
"previous_answer_index_content_id": {"type": "category", "dtype": np.int16},
"previous_answer_content_id": {"type": "category", "dtype": np.int8},
"timediff-elapsedtime_bin500": {"type": "category", "dtype": np.int16}
}
if is_make_feature_factory:
# feature factory
feature_factory_dict = {"user_id": {}}
feature_factory_dict["user_id"]["DurationPreviousContent"] = DurationPreviousContent(is_partial_fit=True)
feature_factory_dict["user_id"]["ElapsedTimeBinningEncoder"] = ElapsedTimeBinningEncoder()
feature_factory_dict["user_id"]["UserContentRateEncoder"] = UserContentRateEncoder(rate_func="elo",
column="user_id")
feature_factory_dict["user_id"]["PreviousAnswer2"] = PreviousAnswer2(groupby="user_id",
column="content_id",
is_debug=is_debug,
model_id=model_id,
n=300)
feature_factory_dict["user_id"]["StudyTermEncoder2"] = StudyTermEncoder2(is_partial_fit=True)
feature_factory_dict["user_id"][f"MeanAggregatorStudyTimebyUserId"] = MeanAggregator(column="user_id",
agg_column="study_time",
remove_now=False)
feature_factory_dict["user_id"]["ElapsedTimeMeanByContentIdEncoder"] = ElapsedTimeMeanByContentIdEncoder()
feature_factory_dict["post"] = {
"DurationFeaturePostProcess": DurationFeaturePostProcess()
}
feature_factory_manager = FeatureFactoryManager(feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=1,
model_id="all",
load_feature=not is_debug,
save_feature=not is_debug)
ff_for_transformer = FeatureFactoryForTransformer(column_config=column_config,
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
df = pd.read_pickle("../input/riiid-test-answer-prediction/train_merged.pickle")
df["prior_question_had_explanation"] = df["prior_question_had_explanation"].fillna(-1)
df["answered_correctly"] = df["answered_correctly"].replace(-1, np.nan)
if is_debug:
df = df.head(10000)
df = df.sort_values(["user_id", "timestamp"]).reset_index(drop=True)
ff_for_transformer.make_dict(df=df)
feature_factory_manager.fit(df)
df = feature_factory_manager.all_predict(df)
def f(x):
x = x // 1000
if x < -90:
return -90
if x > 90:
return 90
return x
df["task_container_id_bin300"] = [x if x < 300 else 300 for x in df["task_container_id"].values]
df["timediff-elapsedtime_bin500"] = [f(x) for x in df["timediff-elapsedtime"].values]
df = df[["user_id", "content_id", "content_type_id", "part", "user_answer", "answered_correctly",
"prior_question_elapsed_time_bin300", "duration_previous_content_bin300",
"prior_question_had_explanation", "rating_diff_content_user_id", "task_container_id_bin300",
"previous_answer_index_content_id", "previous_answer_content_id", "row_id",
"timediff-elapsedtime_bin500"]]
for dicts in feature_factory_manager.feature_factory_dict.values():
for factory in dicts.values():
factory.logger = None
feature_factory_manager.logger = None
with open(f"{output_dir}/feature_factory_manager.pickle", "wb") as f:
pickle.dump(feature_factory_manager, f)
ff_for_transformer.fit(df)
ff_for_transformer.logger = None
with open(f"{output_dir}/feature_factory_manager_for_transformer.pickle", "wb") as f:
pickle.dump(ff_for_transformer, f)
def train_lgbm_cv_newuser_with_iteration(
df: pd.DataFrame,
feature_factory_manager: FeatureFactoryManager,
params: dict,
output_dir: str,
model_id: int,
exp_name: str,
drop_user_id: bool,
categorical_feature: list = [],
experiment_id: int = 0,
is_debug: bool = False):
if not is_debug:
mlflow.start_run(experiment_id=experiment_id, run_name=exp_name)
mlflow.log_param("model_id", model_id)
mlflow.log_param("count_row", len(df))
mlflow.log_param("count_column", len(df.columns))
for key, value in params.items():
mlflow.log_param(key, value)
if drop_user_id:
features = [
x for x in df.columns if x not in [
"answered_correctly", "user_id", "user_answer", "tags",
"type_of", "bundle_id", "previous_5_ans"
]
]
else:
features = [
x for x in df.columns if x not in [
"answered_correctly", "user_answer", "tags", "type_of",
"bundle_id", "previous_5_ans"
]
]
df_imp = pd.DataFrame()
df_imp["feature"] = features
df1 = feature_factory_manager.all_predict(df.copy())
train_idx = []
val_idx = []
np.random.seed(0)
for _, w_df in df.groupby("user_id"):
if np.random.random() < 0.1:
# all val
val_idx.extend(w_df.index.tolist())
else:
train_num = int(len(w_df) * 0.9)
train_idx.extend(w_df[:train_num].index.tolist())
val_idx.extend(w_df[train_num:].index.tolist())
val_idx = val_idx[:1000000]
df1 = df1.drop(
["user_answer", "tags", "type_of", "bundle_id", "previous_5_ans"],
axis=1)
df1.columns = [
x.replace("[", "_").replace("]", "_").replace("'", "_").replace(
" ", "_").replace(",", "_") for x in df1.columns
]
df_train = df1.loc[train_idx]
df_train = df_train[df_train["answered_correctly"].notnull()]
df_val = df1.loc[val_idx]
df_val = df_val[df_val["answered_correctly"].notnull()]
# valid2
feature_factory_manager.fit(df.loc[train_idx])
df2 = []
for i in tqdm.tqdm(range(len(val_idx) // 100)):
w_df = df.loc[val_idx[i * 100:(i + 1) * 100]]
df2.append(feature_factory_manager.partial_predict(w_df))
feature_factory_manager.fit(w_df)
df2 = pd.concat(df2)
df2.columns = [
x.replace("[", "_").replace("]", "_").replace("'", "_").replace(
" ", "_").replace(",", "_") for x in df2.columns
]
df2_val = df2[df2["answered_correctly"].notnull()]
print(df_val)
print(df2_val)
assert len(df_val) == len(df2_val)
print(f"make_train_data len={len(train_idx)}")
train_data = lgb.Dataset(df_train[features],
label=df_train["answered_correctly"])
print(f"make_test_data len={len(val_idx)}")
valid_data1 = lgb.Dataset(df_val[features],
label=df_val["answered_correctly"])
valid_data2 = lgb.Dataset(df2_val[features],
label=df2_val["answered_correctly"])
model = lgb.train(params,
train_data,
categorical_feature=categorical_feature,
valid_sets=[train_data, valid_data1, valid_data2],
verbose_eval=100)
print(
roc_auc_score(df_val["answered_correctly"],
model.predict(df_val[features])))
print(
roc_auc_score(df2_val["answered_correctly"],
model.predict(df2_val[features])))
if not is_debug:
mlflow.log_metric("auc_train", model.best_score["training"]["auc"])
mlflow.log_metric("auc_val", model.best_score["valid_1"]["auc"])
mlflow.end_run()
df_imp["importance"] = model.feature_importance(
"gain") / model.feature_importance("gain").sum()
df_imp.sort_values(
"importance",
ascending=False).to_csv(f"{output_dir}/imp_{model_id}.csv")
with open(f"{output_dir}/model_{model_id}_lgbm.pickle", "wb") as f:
pickle.dump(model, f)
y_oof = model.predict(df.loc[val_idx][features])
df_oof = pd.DataFrame()
df_oof["row_id"] = df.loc[val_idx].index
df_oof["predict"] = y_oof
df_oof["target"] = df.loc[val_idx]["answered_correctly"].values
df_oof.to_csv(f"{output_dir}/oof_{model_id}_lgbm.csv", index=False)
if random.random() < new_user_ratio:
val_idx.extend(w_df.index.tolist())
else:
train_num = (np.random.random(len(w_df)) < 0.9 /
(1 - new_user_ratio)).sum()
train_idx.extend(w_df[:train_num].index.tolist())
val_idx.extend(w_df[train_num:].index.tolist())
print(len(train_idx))
print(len(df))
df_train = df.iloc[train_idx]
df_val = df.iloc[val_idx]
df = merge(df=df, df_question=df_question, df_lecture=df_lecture)
df_train = merge(df=df_train, df_question=df_question, df_lecture=df_lecture)
df_all_fit = feature_factory_manager.all_predict(df)
df_train_all_fit = df_all_fit[df_all_fit["row_id"].isin(
df_train["row_id"].values)]
df_val_all_fit = df_all_fit[~df_all_fit["row_id"].isin(df_train["row_id"].
values)]
feature_factory_manager.fit(df_train, all_predict_mode=True)
gen = MyEnvironment(df_test=df_val, interval=1).iter_test()
env_manager = EnvironmentManager(
feature_factory_manager=feature_factory_manager,
gen=gen,
fit_interval=150,
df_question=df_question,
df_lecture=df_lecture)
def test_interval1(self):
logger = get_logger()
df = pd.DataFrame({"row_id": [0, 1, 2, 3, 4, 5, 6, 7],
"user_id": ["a", "a", "a", "b", "a", "b", "a", "b"],
"timestamp": [0, 1, 2, 3, 4, 5, 6, 7],
"content_id": [0, 0, 1, 1, 0, 0, 1, 1],
"content_type_id": [0, 1, 0, 0, 0, 0, 0, 1],
"user_answer": [0, 1, 2, 3, 4, 5, 6, 7],
"answered_correctly": [0, -1, 1, -1, 0, 0, 1, -1],
"prior_question_had_explanation": [0, 0, 0, 0, 0, 0, 0, 0]}).sort_values(["user_id", "timestamp"])
df_question = pd.DataFrame({"question_id": [0, 1],
"bundle_id": [0, 1],
"correct_answer": [0, 1],
"part": [0, 1],
"tags": ["0", "1"]})
df_lecture = pd.DataFrame({"lecture_id": [0, 1],
"tag": [0, 1],
"part": [0, 1],
"type_of": ["0", "1"]})
feature_factory_dict = {
"user_id": {
"CountEncoder": CountEncoder(column="user_id"),
"TargetEncoder": TargetEncoder(column="user_id")}
}
print(df.iloc[4:])
gen = MyEnvironment(df_test=df.iloc[4:],
interval=1).iter_test()
feature_factory_manager = FeatureFactoryManager(feature_factory_dict=feature_factory_dict,
logger=logger)
env_manager = EnvironmentManager(feature_factory_manager=feature_factory_manager,
gen=gen,
fit_interval=1,
df_question=df_question,
df_lecture=df_lecture)
w_df1 = pd.merge(df[df["content_type_id"] == 0], df_question, how="left", left_on="content_id",
right_on="question_id")
w_df2 = pd.merge(df[df["content_type_id"] == 1], df_lecture, how="left", left_on="content_id",
right_on="lecture_id")
df2 = pd.concat([w_df1, w_df2]).sort_values(["user_id", "timestamp"])
df_expect = feature_factory_manager.all_predict(df2).iloc[4:]
df_expect["tag"] = df_expect["tag"].fillna(-1)
df_expect["correct_answer"] = df_expect["correct_answer"].fillna(-1)
df_expect["bundle_id"] = df_expect["bundle_id"].fillna(-1)
df_expect["prior_question_had_explanation"] = df_expect["prior_question_had_explanation"].astype("float16").fillna(-1).astype("int8")
df_expect.columns = [x.replace(" ", "_") for x in df_expect.columns]
df_actual = pd.DataFrame()
feature_factory_manager.fit(df2.iloc[:4])
while True:
x = env_manager.step()
if x is None:
break
df_test = x[0]
df_sub = x[1]
df_actual = pd.concat([df_actual, df_test], axis=0)
pd.testing.assert_frame_equal(df_expect.reset_index(drop=True),
df_actual.reset_index(drop=True),
check_dtype=False)
"type_of"
]:
feature_factory_dict["user_id"][f"Counter{col}"] = Counter(
groupby_column="user_id",
agg_column=col,
categories=df[col].drop_duplicates())
feature_factory_manager = FeatureFactoryManager(
feature_factory_dict=feature_factory_dict, logger=logger, split_num=10)
train_idx = []
val_idx = []
for _, w_df in df.groupby("user_id"):
train_num = (np.random.random(len(w_df)) < 0.8).sum()
train_idx.extend(w_df[:train_num].index.tolist())
val_idx.extend(w_df[train_num:].index.tolist())
df = feature_factory_manager.all_predict(
pd.concat([df.iloc[train_idx], df.iloc[val_idx]]))
df = df.drop(["user_answer", "tags", "type_of"], axis=1)
df_train = df.iloc[:len(train_idx)]
df_val = df.iloc[len(train_idx):]
print(df_train)
df2 = pd.read_pickle(fname).sort_values(["user_id", "timestamp"
]).reset_index(drop=True)
# df2 = pd.concat([pd.read_pickle(fname).head(500), pd.read_pickle(fname).tail(500)]).sort_values(["user_id", "timestamp"]).reset_index(drop=True)
df2["answered_correctly"] = df2["answered_correctly"].replace(-1, np.nan)
df2["prior_question_had_explanation"] = df2[
"prior_question_had_explanation"].fillna(-1).astype("int8")
df2_train = feature_factory_manager.all_predict(df2.iloc[train_idx])
print(df2_train)
feature_factory_manager.fit(df2.iloc[train_idx], is_first_fit=True)
df2_val = []
def main(params: dict):
import mlflow
logger = get_logger()
print("start params={}".format(params))
df = pd.read_pickle(
"../input/riiid-test-answer-prediction/train_merged.pickle")
# df = pd.read_pickle("../input/riiid-test-answer-prediction/split10/train_0.pickle").sort_values(["user_id", "timestamp"]).reset_index(drop=True)
if is_debug:
df = df.head(30000)
logger = get_logger()
feature_factory_dict = {"user_id": {}}
feature_factory_dict["user_id"][
"DurationPreviousContent"] = DurationPreviousContent(
is_partial_fit=True)
feature_factory_manager = FeatureFactoryManager(
feature_factory_dict=feature_factory_dict,
logger=logger,
split_num=1,
model_id="all",
load_feature=not is_debug,
save_feature=not is_debug)
print("all_predict")
df = feature_factory_manager.all_predict(df)
df = df[[
"user_id", "content_id", "content_type_id", "part", "user_answer",
"answered_correctly", "prior_question_elapsed_time",
"duration_previous_content"
]]
print("data preprocess")
df["prior_question_elapsed_time_bin300"] = [
x // 1000 if x // 1000 < 300 else 300
for x in df["prior_question_elapsed_time"].fillna(0).values
]
df["duration_previous_content_bin300"] = [
x // 1000 if x // 1000 < 300 else 300
for x in df["duration_previous_content"].fillna(0).values
]
train_idx = []
val_idx = []
np.random.seed(0)
for k, idx in tqdm(
df[df["content_type_id"] == 0].groupby("user_id").indices.items()):
if np.random.random() < 0.1:
# all val
val_idx.extend(idx)
else:
train_num = int(len(idx) * 0.9)
train_idx.extend(idx[:train_num])
val_idx.extend(idx[train_num:])
ff_for_transformer = FeatureFactoryForTransformer(
column_config={
("content_id", "content_type_id"): {
"type": "category"
},
"user_answer": {
"type": "category"
},
"part": {
"type": "category"
},
"prior_question_elapsed_time_bin300": {
"type": "category"
},
"duration_previous_content_bin300": {
"type": "category"
}
},
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
ff_for_transformer.make_dict(df)
n_skill = len(ff_for_transformer.embbed_dict[("content_id",
"content_type_id")])
if not load_pickle:
df["is_val"] = 0
df["is_val"].loc[val_idx] = 1
w_df = df[df["is_val"] == 0]
w_df["group"] = (
w_df.groupby("user_id")["user_id"].transform("count") -
w_df.groupby("user_id").cumcount()) // params["max_seq"]
w_df["user_id"] = w_df["user_id"].astype(
str) + "_" + w_df["group"].astype(str)
group = ff_for_transformer.all_predict(w_df)
dataset_train = SAKTDataset(group,
n_skill=n_skill,
max_seq=params["max_seq"])
del w_df
gc.collect()
ff_for_transformer = FeatureFactoryForTransformer(
column_config={
("content_id", "content_type_id"): {
"type": "category"
},
"user_answer": {
"type": "category"
},
"part": {
"type": "category"
},
"prior_question_elapsed_time_bin300": {
"type": "category"
},
"duration_previous_content_bin300": {
"type": "category"
}
},
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
if not load_pickle:
group = ff_for_transformer.all_predict(df[df["content_type_id"] == 0])
dataset_val = SAKTDataset(group,
is_test=True,
n_skill=n_skill,
max_seq=params["max_seq"])
os.makedirs("../input/feature_engineering/model048", exist_ok=True)
if not load_pickle:
with open(f"../input/feature_engineering/model048/train.pickle",
"wb") as f:
pickle.dump(dataset_train, f)
with open(f"../input/feature_engineering/model048/val.pickle",
"wb") as f:
pickle.dump(dataset_val, f)
if load_pickle:
with open(f"../input/feature_engineering/model048/train.pickle",
"rb") as f:
dataset_train = pickle.load(f)
with open(f"../input/feature_engineering/model048/val.pickle",
"rb") as f:
dataset_val = pickle.load(f)
dataloader_train = DataLoader(dataset_train,
batch_size=64,
shuffle=True,
num_workers=1)
dataloader_val = DataLoader(dataset_val,
batch_size=64,
shuffle=False,
num_workers=1)
model = SAKTModel(n_skill,
embed_dim=params["embed_dim"],
max_seq=params["max_seq"],
dropout=dropout)
optimizer = torch.optim.Adam(model.parameters(), lr=params["lr"])
criterion = nn.BCEWithLogitsLoss()
model.to(device)
criterion.to(device)
for epoch in range(epochs):
loss, acc, auc, auc_val = train_epoch(model, dataloader_train,
dataloader_val, optimizer,
criterion, device)
print("epoch - {} train_loss - {:.3f} auc - {:.4f} auc-val: {:.4f}".
format(epoch, loss, auc, auc_val))
preds = []
labels = []
for item in tqdm(dataloader_val):
x = item["x"].to(device).long()
target_id = item["target_id"].to(device).long()
part = item["part"].to(device).long()
label = item["label"].to(device).float()
elapsed_time = item["elapsed_time"].to(device).long()
duration_previous_content = item["duration_previous_content"].to(
device).long()
output, atten_weight = model(x, target_id, part, elapsed_time,
duration_previous_content)
preds.extend(torch.nn.Sigmoid()(
output[:, -1]).view(-1).data.cpu().numpy().tolist())
labels.extend(label[:, -1].view(-1).data.cpu().numpy().tolist())
df_oof = pd.DataFrame()
df_oof["row_id"] = df.loc[val_idx].index
df_oof["predict"] = preds
df_oof["target"] = df.loc[val_idx]["answered_correctly"].values
df_oof.to_csv(f"{output_dir}/transformers1.csv", index=False)
df_oof2 = pd.read_csv(
"../output/ex_237/20201213110353/oof_train_0_lgbm.csv")
df_oof2.columns = ["row_id", "predict_lgbm", "target"]
df_oof2 = pd.merge(df_oof, df_oof2, how="inner")
auc_transformer = roc_auc_score(df_oof2["target"].values,
df_oof2["predict"].values)
auc_lgbm = roc_auc_score(df_oof2["target"].values,
df_oof2["predict_lgbm"].values)
print("single transformer: {:.4f}".format(auc_transformer))
print("lgbm: {:.4f}".format(auc_lgbm))
print("ensemble")
max_auc = 0
max_nn_ratio = 0
for r in np.arange(0, 1.05, 0.05):
auc = roc_auc_score(
df_oof2["target"].values, df_oof2["predict_lgbm"].values *
(1 - r) + df_oof2["predict"].values * r)
print("[nn_ratio: {:.2f}] AUC: {:.4f}".format(r, auc))
if max_auc < auc:
max_auc = auc
max_nn_ratio = r
print(len(df_oof2))
if not is_debug:
mlflow.start_run(experiment_id=10, run_name=os.path.basename(__file__))
mlflow.log_param("count_row", len(df))
for key, value in params.items():
mlflow.log_param(key, value)
mlflow.log_metric("auc_val", auc_transformer)
mlflow.log_metric("auc_lgbm", auc_lgbm)
mlflow.log_metric("auc_ensemble", max_auc)
mlflow.log_metric("ensemble_nn_ratio", max_nn_ratio)
mlflow.end_run()
torch.save(model.state_dict(), f"{output_dir}/transformers.pth")
with open(f"{output_dir}/transformer_param.json", "w") as f:
json.dump(params, f)
if is_make_feature_factory:
ff_for_transformer = FeatureFactoryForTransformer(
column_config={
("content_id", "content_type_id"): {
"type": "category"
},
"part": {
"type": "category"
}
},
dict_path="../feature_engineering/",
sequence_length=params["max_seq"],
logger=logger)
df = pd.read_pickle(
"../input/riiid-test-answer-prediction/train_merged.pickle")
if is_debug:
df = df.head(10000)
df = df.sort_values(["user_id", "timestamp"]).reset_index(drop=True)
ff_for_transformer.fit(df)
ff_for_transformer.logger = None
with open(
f"{output_dir}/feature_factory_manager_for_transformer.pickle",
"wb") as f:
pickle.dump(ff_for_transformer, f)