\nI think that case 1 is more reasonable considering leakage. (ex. imputation nan with next day in lag)
\n[Case1]
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6610354%2Ffd930e9643fd4d9508fbecaf94ba5a88%2Fcase1.png?generation=1732517447727717&alt=media\")
\n[Case2]
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6610354%2Fb358c8ec93f321fb756eeadf33e11308%2Fcase2.png?generation=1732517454750318&alt=media\")
I am confusing yet. Which case is right?
\nI think that case 1 is more reasonable considering leakage. (ex. imputation nan with next day in lag)
\n[Case1]
\n
\n[Case2]
\n
I am confusing yet. Which case is right?
\nI think that case 1 is more reasonable considering leakage. (ex. imputation nan with next day in lag)
\n[Case1]
\n
\n[Case2]
\n
It is the first one.
\nFor example:
\nSay your predict function is being called for date_id = D. Then, inside your predict function, at the first time_id = T of date_id = D, you are given:
test dataframe - data (feature_01, 02, weight… etc) corresponding to time_id = T of date_id D for all symbol_ids available at that date_id_time_id.lags dataframe - 1 day lagged responder values, as well as time_id, symbol_id for all time_ids, and all available symbol_ids for date_id D-1. The unique values of time_ids and symbol_ids are not guaranteed to be exactly the same as what is given in the test dataframe. Note that in this lags df, the date_id = D i.e. same as test df, even though the columns represent the lagged responders of previous day.On all other time_ids for date_id = D, lags will be None. i.e. lags df will be non-null only on the first time_id of a given date_id.
", "rawMarkdown": "It is the first one.\n\nFor example:\n\nSay your predict function is being called for ```date_id = D```. Then, inside your predict function, at the first ```time_id = T``` of ```date_id = D```, you are given:\n1. ```test``` dataframe - data (feature_01, 02, weight... etc) corresponding to ```time_id = T``` of ```date_id D``` for all ```symbol_ids``` available at that ```date_id_time_id```.\n2. ```lags``` dataframe - 1 day lagged responder values, as well as ```time_id, symbol_id``` for all ```time_ids```, and all available ```symbol_ids``` for ```date_id D-1```. The unique values of ```time_ids``` and ```symbol_ids``` are not guaranteed to be exactly the same as what is given in the ```test``` dataframe. Note that in this ```lags``` df, the date_id = D i.e. same as test df, even though the columns represent the lagged responders of previous day.\n\n\nOn all other time_ids for date_id = D, lags will be None. i.e. lags df will be non-null only on the first time_id of a given date_id.", "votes": null }, { "id": "3055177", "postDate": "11/25/2024 15:59:59", "content": "however, the lags will be saved in a global variable lags_ which we can access when the time_id is not 0, correct?
\ni`m getting an error as i try to merge the lags_ dataframe with the test when time_id is not zero.
\nAny tips?
`lags_ : pl.DataFrame | None = None
\ndef predict(test: pl.DataFrame, lags: pl.DataFrame | None) -> pl.DataFrame | pd.DataFrame:
\n lags is not None:\n lags_ = lags \n\ntest1= test(lags_, on=, how=)\n)\nfeat = test1()\npred = model(feat)\npredictions = test()(pl(, pred()))\nassert (predictions, (pl, pd.DataFrame))\nassert (predictions.) == [, ]\nassert (predictions) == (test)\nreturn predictions\n`
", "rawMarkdown": "however, the lags will be saved in a global variable lags_ which we can access when the time_id is not 0, correct?\ni`m getting an error as i try to merge the lags_ dataframe with the test when time_id is not zero.\nAny tips? \n\n`lags_ : pl.DataFrame | None = None\n\ndef predict(test: pl.DataFrame, lags: pl.DataFrame | None) -> pl.DataFrame | pd.DataFrame:\n\n if lags is not None:\n lags_ = lags \n\n test1= test.join(lags_, on=['date_id', 'time_id', 'symbol_id'], how='left')\n print(test1.head(10))\n feat = test1[feature_names].to_numpy()\n pred = model.predict(feat)\n predictions = test.select('row_id').with_columns(pl.Series('responder_6', pred.ravel()))\n assert isinstance(predictions, (pl.DataFrame, pd.DataFrame))\n assert list(predictions.columns) == ['row_id', 'responder_6']\n assert len(predictions) == len(test)\n return predictions\n`", "votes": null }, { "id": "3055205", "postDate": "11/25/2024 16:27:45", "content": "what was your error?
\nyour code to join the lags with test seems ok. However, your issue might be related to handling of NaNs when calling model.predict.
", "rawMarkdown": "what was your error?\n\nyour code to join the lags with test seems ok. However, your issue might be related to handling of NaNs when calling model.predict.", "votes": null }, { "id": "3055245", "postDate": "11/25/2024 16:44:36", "content": "lgbm, no issues with predict. Funny thing is that I look at the logs and there is no errors there. Only at the submissions page:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F23236619%2Fee9b3f4289f6de1cf298fa9a2219a04b%2Fsub.png?generation=1732552997114172&alt=media\")
And the log (I was trying to print some shapes to see what I was missing) No errors:
\n32.3s 1 CPU times: user 13.8 s, sys: 16 s, total: 29.8 s
\n32.3s 2 Wall time: 18 s
\n43.9s 3 /opt/conda/lib/python3.10/site-packages/lightgbm/engine.py:172: UserWarning: Found n_estimators in params. Will use it instead of argument
\n43.9s 4 log_warning(f\"Found {alias} in params. Will use it instead of argument\")\n843.2s 5 CPU times: user 48min 19s, sys: 13.7 s, total: 48min 32s\n843.2s 6 Wall time: 13min 19s\n843.9s 7 (39, 12)\n843.9s 8 shape: (1, 1)\n843.9s 9 ┌─────────┐\n843.9s 10 │ date_id │\n843.9s 11 │ --- │\n843.9s 12 │ i16 │\n843.9s 13 ╞═════════╡\n843.9s 14 │ 0 │\n843.9s 15 └─────────┘\n843.9s 16 shape: (1, 1)\n843.9s 17 ┌─────────┐\n843.9s 18 │ time_id │\n843.9s 19 │ --- │\n843.9s 20 │ i16 │\n843.9s 21 ╞═════════╡\n843.9s 22 │ 0 │\n843.9s 23 └─────────┘\n843.9s 24 shape: (39, 1)\n843.9s 25 ┌───────────┐\n843.9s 26 │ symbol_id │\n843.9s 27 │ --- │\n843.9s 28 │ i8 │\n843.9s 29 ╞═══════════╡\n843.9s 30 │ 23 │\n843.9s 31 │ 4 │\n843.9s 32 │ 26 │\n843.9s 33 │ 32 │\n843.9s 34 │ 21 │\n843.9s 35 │ … │\n843.9s 36 │ 31 │\n843.9s 37 │ 33 │\n843.9s 38 │ 28 │\n843.9s 39 │ 17 │\n843.9s 40 │ 11 │\n843.9s 41 └───────────┘\n843.9s 42 test1\n843.9s 43 shape: (10, 94)\n843.9s 44 ┌────────┬─────────┬─────────┬───────────┬───┬─────────────┬─────────────┬────────────┬────────────┐\n843.9s 45 │ row_id ┆ date_id ┆ time_id ┆ symbol_id ┆ … ┆ responder_5 ┆ responder_6 ┆ responder ┆ responder_ │
\n843.9s 46 │ --- ┆ --- ┆ --- ┆ --- ┆ ┆ _lag_1 ┆ _lag_1 ┆ 7_lag_1 ┆ 8_lag_1 │
\n843.9s 47 │ i64 ┆ i16 ┆ i16 ┆ i8 ┆ ┆ --- ┆ --- ┆ --- ┆ --- │
\n843.9s 48 │ ┆ ┆ ┆ ┆ ┆ f32 ┆ f32 ┆ f32 ┆ f32 │
\n843.9s 49 ╞════════╪═════════╪═════════╪═══════════╪═══╪═════════════╪═════════════╪════════════╪════════════╡
\n843.9s 50 │ 0 ┆ 0 ┆ 0 ┆ 0 ┆ … ┆ -0.036595 ┆ -1.305746 ┆ -0.795677 ┆ -0.143724 │
\n843.9s 51 │ 1 ┆ 0 ┆ 0 ┆ 1 ┆ … ┆ -0.615652 ┆ -1.162801 ┆ -1.205924 ┆ -1.245934 │
\n843.9s 52 │ 2 ┆ 0 ┆ 0 ┆ 2 ┆ … ┆ -0.378265 ┆ -1.57429 ┆ -1.863071 ┆ -0.027343 │
\n843.9s 53 │ 3 ┆ 0 ┆ 0 ┆ 3 ┆ … ┆ -0.054984 ┆ 0.329152 ┆ -0.965471 ┆ 0.576635 │
\n843.9s 54 │ 4 ┆ 0 ┆ 0 ┆ 4 ┆ … ┆ -0.597093 ┆ 0.219856 ┆ -0.276356 ┆ -0.90479 │
\n843.9s 55 │ 5 ┆ 0 ┆ 0 ┆ 5 ┆ … ┆ -0.24035 ┆ -0.913801 ┆ -0.548867 ┆ -1.283726 │
\n843.9s 56 │ 6 ┆ 0 ┆ 0 ┆ 6 ┆ … ┆ 0.517728 ┆ 0.896325 ┆ 1.068884 ┆ 1.57929 │
\n843.9s 57 │ 7 ┆ 0 ┆ 0 ┆ 7 ┆ … ┆ -0.182887 ┆ -0.492168 ┆ -0.142915 ┆ -0.202081 │
\n843.9s 58 │ 8 ┆ 0 ┆ 0 ┆ 8 ┆ … ┆ -0.225042 ┆ 0.95646 ┆ 2.185598 ┆ -0.435856 │
\n843.9s 59 │ 9 ┆ 0 ┆ 0 ┆ 9 ┆ … ┆ -0.774299 ┆ -0.716492 ┆ -1.471419 ┆ -1.107083 │
\n843.9s 60 └────────┴─────────┴─────────┴───────────┴───┴─────────────┴─────────────┴────────────┴────────────┘
\n849.4s 61 [NbConvertApp] Converting notebook notebook.ipynb to notebook
\n850.0s 62 [NbConvertApp] Writing 40602 bytes to notebook.ipynb
\n852.4s 63 [NbConvertApp] Converting notebook notebook.ipynb to html
\n853.6s 64 [NbConvertApp] Writing 332474 bytes to results.html
It is the code below I did test.
\nI found that I explicitly select the samples on time_id=0 in lag data.
\nI don't know why I do this… (since I don't know how the lag is provided actually)
() -> pl.DataFrame | pd.DataFrame:\n lags_\n lags :\n \n lags_ = lags.group_by([, ], maintain_order=).first().drop()\n \n \n lags_ :\n lags_ = test.select(\n pl.col(), pl.col(),\n *[pl.lit().alias() col CFG.responder_cols]\n )\n\n \n df_test = preprocessor.run(test.drop([, ]), lags_, mode=)\n df_test = df_test.drop(CFG.metadata_cols).to_pandas()\n\n fold_pred = []\n fold (CFG.n_folds):\n ()\n \n test_x = df_test.copy()\n \n scaler = fold_output[fold][]\n test_x[var_info[]] = scaler.transform(test_x[var_info[]])\n encoder = fold_output[fold][]\n ohe_cols = fold_output[fold][]\n test_x = pd.concat([\n test_x.drop(var_info[], axis=),\n pd.DataFrame(encoder.transform(test_x[var_info[]]), index=test_x.index.copy(), columns=ohe_cols)\n ], axis=)\n (, test_x.shape)\n \n model = fold_output[fold][]\n \n test_pred = model.predict(test_x)\n \n test_pred = np.clip(test_pred, a_min=-, a_max=)\n fold_pred.append(test_pred)\n\n fold_pred = np.stack(fold_pred, axis=).mean(axis=)\n predictions = test.select(, pl.Series(CFG.target_col, fold_pred, dtype=pl.datatypes.Float32))\n predictions\nThanks for your explanation :)
", "rawMarkdown": "Thanks for your explanation :)", "votes": null }, { "id": "3055969", "postDate": "11/26/2024 11:14:45", "content": "Thanks! Really helpful!
", "rawMarkdown": "Thanks! Really helpful!", "votes": null } ], "comments": [ { "id": 3055139, "author_name": "alamjs", "author_url": "", "post_date": "11/25/2024 15:15:15", "content": "It is the first one.
\nFor example:
\nSay your predict function is being called for date_id = D. Then, inside your predict function, at the first time_id = T of date_id = D, you are given:
test dataframe - data (feature_01, 02, weight… etc) corresponding to time_id = T of date_id D for all symbol_ids available at that date_id_time_id.lags dataframe - 1 day lagged responder values, as well as time_id, symbol_id for all time_ids, and all available symbol_ids for date_id D-1. The unique values of time_ids and symbol_ids are not guaranteed to be exactly the same as what is given in the test dataframe. Note that in this lags df, the date_id = D i.e. same as test df, even though the columns represent the lagged responders of previous day.On all other time_ids for date_id = D, lags will be None. i.e. lags df will be non-null only on the first time_id of a given date_id.
", "votes": null, "replies": [ { "id": 3055177, "author_name": "tomazrochadamota", "author_url": "", "post_date": "11/25/2024 15:59:59", "content": "however, the lags will be saved in a global variable lags_ which we can access when the time_id is not 0, correct?
\ni`m getting an error as i try to merge the lags_ dataframe with the test when time_id is not zero.
\nAny tips?
`lags_ : pl.DataFrame | None = None
\ndef predict(test: pl.DataFrame, lags: pl.DataFrame | None) -> pl.DataFrame | pd.DataFrame:
\n lags is not None:\n lags_ = lags \n\ntest1= test(lags_, on=, how=)\n)\nfeat = test1()\npred = model(feat)\npredictions = test()(pl(, pred()))\nassert (predictions, (pl, pd.DataFrame))\nassert (predictions.) == [, ]\nassert (predictions) == (test)\nreturn predictions\n`
", "votes": null, "replies": [ { "id": 3055205, "author_name": "alamjs", "author_url": "", "post_date": "11/25/2024 16:27:45", "content": "what was your error?
\nyour code to join the lags with test seems ok. However, your issue might be related to handling of NaNs when calling model.predict.
", "votes": null, "replies": [ { "id": 3055245, "author_name": "tomazrochadamota", "author_url": "", "post_date": "11/25/2024 16:44:36", "content": "lgbm, no issues with predict. Funny thing is that I look at the logs and there is no errors there. Only at the submissions page:
\n
And the log (I was trying to print some shapes to see what I was missing) No errors:
\n32.3s 1 CPU times: user 13.8 s, sys: 16 s, total: 29.8 s
\n32.3s 2 Wall time: 18 s
\n43.9s 3 /opt/conda/lib/python3.10/site-packages/lightgbm/engine.py:172: UserWarning: Found n_estimators in params. Will use it instead of argument
\n43.9s 4 log_warning(f\"Found {alias} in params. Will use it instead of argument\")\n843.2s 5 CPU times: user 48min 19s, sys: 13.7 s, total: 48min 32s\n843.2s 6 Wall time: 13min 19s\n843.9s 7 (39, 12)\n843.9s 8 shape: (1, 1)\n843.9s 9 ┌─────────┐\n843.9s 10 │ date_id │\n843.9s 11 │ --- │\n843.9s 12 │ i16 │\n843.9s 13 ╞═════════╡\n843.9s 14 │ 0 │\n843.9s 15 └─────────┘\n843.9s 16 shape: (1, 1)\n843.9s 17 ┌─────────┐\n843.9s 18 │ time_id │\n843.9s 19 │ --- │\n843.9s 20 │ i16 │\n843.9s 21 ╞═════════╡\n843.9s 22 │ 0 │\n843.9s 23 └─────────┘\n843.9s 24 shape: (39, 1)\n843.9s 25 ┌───────────┐\n843.9s 26 │ symbol_id │\n843.9s 27 │ --- │\n843.9s 28 │ i8 │\n843.9s 29 ╞═══════════╡\n843.9s 30 │ 23 │\n843.9s 31 │ 4 │\n843.9s 32 │ 26 │\n843.9s 33 │ 32 │\n843.9s 34 │ 21 │\n843.9s 35 │ … │\n843.9s 36 │ 31 │\n843.9s 37 │ 33 │\n843.9s 38 │ 28 │\n843.9s 39 │ 17 │\n843.9s 40 │ 11 │\n843.9s 41 └───────────┘\n843.9s 42 test1\n843.9s 43 shape: (10, 94)\n843.9s 44 ┌────────┬─────────┬─────────┬───────────┬───┬─────────────┬─────────────┬────────────┬────────────┐\n843.9s 45 │ row_id ┆ date_id ┆ time_id ┆ symbol_id ┆ … ┆ responder_5 ┆ responder_6 ┆ responder ┆ responder_ │
\n843.9s 46 │ --- ┆ --- ┆ --- ┆ --- ┆ ┆ _lag_1 ┆ _lag_1 ┆ 7_lag_1 ┆ 8_lag_1 │
\n843.9s 47 │ i64 ┆ i16 ┆ i16 ┆ i8 ┆ ┆ --- ┆ --- ┆ --- ┆ --- │
\n843.9s 48 │ ┆ ┆ ┆ ┆ ┆ f32 ┆ f32 ┆ f32 ┆ f32 │
\n843.9s 49 ╞════════╪═════════╪═════════╪═══════════╪═══╪═════════════╪═════════════╪════════════╪════════════╡
\n843.9s 50 │ 0 ┆ 0 ┆ 0 ┆ 0 ┆ … ┆ -0.036595 ┆ -1.305746 ┆ -0.795677 ┆ -0.143724 │
\n843.9s 51 │ 1 ┆ 0 ┆ 0 ┆ 1 ┆ … ┆ -0.615652 ┆ -1.162801 ┆ -1.205924 ┆ -1.245934 │
\n843.9s 52 │ 2 ┆ 0 ┆ 0 ┆ 2 ┆ … ┆ -0.378265 ┆ -1.57429 ┆ -1.863071 ┆ -0.027343 │
\n843.9s 53 │ 3 ┆ 0 ┆ 0 ┆ 3 ┆ … ┆ -0.054984 ┆ 0.329152 ┆ -0.965471 ┆ 0.576635 │
\n843.9s 54 │ 4 ┆ 0 ┆ 0 ┆ 4 ┆ … ┆ -0.597093 ┆ 0.219856 ┆ -0.276356 ┆ -0.90479 │
\n843.9s 55 │ 5 ┆ 0 ┆ 0 ┆ 5 ┆ … ┆ -0.24035 ┆ -0.913801 ┆ -0.548867 ┆ -1.283726 │
\n843.9s 56 │ 6 ┆ 0 ┆ 0 ┆ 6 ┆ … ┆ 0.517728 ┆ 0.896325 ┆ 1.068884 ┆ 1.57929 │
\n843.9s 57 │ 7 ┆ 0 ┆ 0 ┆ 7 ┆ … ┆ -0.182887 ┆ -0.492168 ┆ -0.142915 ┆ -0.202081 │
\n843.9s 58 │ 8 ┆ 0 ┆ 0 ┆ 8 ┆ … ┆ -0.225042 ┆ 0.95646 ┆ 2.185598 ┆ -0.435856 │
\n843.9s 59 │ 9 ┆ 0 ┆ 0 ┆ 9 ┆ … ┆ -0.774299 ┆ -0.716492 ┆ -1.471419 ┆ -1.107083 │
\n843.9s 60 └────────┴─────────┴─────────┴───────────┴───┴─────────────┴─────────────┴────────────┴────────────┘
\n849.4s 61 [NbConvertApp] Converting notebook notebook.ipynb to notebook
\n850.0s 62 [NbConvertApp] Writing 40602 bytes to notebook.ipynb
\n852.4s 63 [NbConvertApp] Converting notebook notebook.ipynb to html
\n853.6s 64 [NbConvertApp] Writing 332474 bytes to results.html
It is the code below I did test.
\nI found that I explicitly select the samples on time_id=0 in lag data.
\nI don't know why I do this… (since I don't know how the lag is provided actually)
() -> pl.DataFrame | pd.DataFrame:\n lags_\n lags :\n \n lags_ = lags.group_by([, ], maintain_order=).first().drop()\n \n \n lags_ :\n lags_ = test.select(\n pl.col(), pl.col(),\n *[pl.lit().alias() col CFG.responder_cols]\n )\n\n \n df_test = preprocessor.run(test.drop([, ]), lags_, mode=)\n df_test = df_test.drop(CFG.metadata_cols).to_pandas()\n\n fold_pred = []\n fold (CFG.n_folds):\n ()\n \n test_x = df_test.copy()\n \n scaler = fold_output[fold][]\n test_x[var_info[]] = scaler.transform(test_x[var_info[]])\n encoder = fold_output[fold][]\n ohe_cols = fold_output[fold][]\n test_x = pd.concat([\n test_x.drop(var_info[], axis=),\n pd.DataFrame(encoder.transform(test_x[var_info[]]), index=test_x.index.copy(), columns=ohe_cols)\n ], axis=)\n (, test_x.shape)\n \n model = fold_output[fold][]\n \n test_pred = model.predict(test_x)\n \n test_pred = np.clip(test_pred, a_min=-, a_max=)\n fold_pred.append(test_pred)\n\n fold_pred = np.stack(fold_pred, axis=).mean(axis=)\n predictions = test.select(, pl.Series(CFG.target_col, fold_pred, dtype=pl.datatypes.Float32))\n predictions\nThanks! Really helpful!
", "votes": null, "replies": [] } ] } ] }, { "id": 3055890, "author_name": "cafelatte1", "author_url": "", "post_date": "11/26/2024 08:50:36", "content": "Thanks for your explanation :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3054786": "I am confusing yet. Which case is right?\nI think that case 1 is more reasonable considering leakage. (ex. imputation nan with next day in lag)\n[Case1]\n\n[Case2]\n", "3055139": "It is the first one.\n\nFor example:\n\nSay your predict function is being called for ```date_id = D```. Then, inside your predict function, at the first ```time_id = T``` of ```date_id = D```, you are given:\n1. ```test``` dataframe - data (feature_01, 02, weight... etc) corresponding to ```time_id = T``` of ```date_id D``` for all ```symbol_ids``` available at that ```date_id_time_id```.\n2. ```lags``` dataframe - 1 day lagged responder values, as well as ```time_id, symbol_id``` for all ```time_ids```, and all available ```symbol_ids``` for ```date_id D-1```. The unique values of ```time_ids``` and ```symbol_ids``` are not guaranteed to be exactly the same as what is given in the ```test``` dataframe. Note that in this ```lags``` df, the date_id = D i.e. same as test df, even though the columns represent the lagged responders of previous day.\n\n\nOn all other time_ids for date_id = D, lags will be None. i.e. lags df will be non-null only on the first time_id of a given date_id.", "3055177": "however, the lags will be saved in a global variable lags_ which we can access when the time_id is not 0, correct?\ni`m getting an error as i try to merge the lags_ dataframe with the test when time_id is not zero.\nAny tips? \n\n`lags_ : pl.DataFrame | None = None\n\ndef predict(test: pl.DataFrame, lags: pl.DataFrame | None) -> pl.DataFrame | pd.DataFrame:\n\n if lags is not None:\n lags_ = lags \n\n test1= test.join(lags_, on=['date_id', 'time_id', 'symbol_id'], how='left')\n print(test1.head(10))\n feat = test1[feature_names].to_numpy()\n pred = model.predict(feat)\n predictions = test.select('row_id').with_columns(pl.Series('responder_6', pred.ravel()))\n assert isinstance(predictions, (pl.DataFrame, pd.DataFrame))\n assert list(predictions.columns) == ['row_id', 'responder_6']\n assert len(predictions) == len(test)\n return predictions\n`", "3055205": "what was your error?\n\nyour code to join the lags with test seems ok. However, your issue might be related to handling of NaNs when calling model.predict.", "3055245": "lgbm, no issues with predict. Funny thing is that I look at the logs and there is no errors there. Only at the submissions page:\n\n\nAnd the log (I was trying to print some shapes to see what I was missing) No errors:\n32.3s\t1\tCPU times: user 13.8 s, sys: 16 s, total: 29.8 s\n32.3s\t2\tWall time: 18 s\n43.9s\t3\t/opt/conda/lib/python3.10/site-packages/lightgbm/engine.py:172: UserWarning: Found `n_estimators` in params. Will use it instead of argument\n43.9s\t4\t _log_warning(f\"Found `{alias}` in params. Will use it instead of argument\")\n843.2s\t5\tCPU times: user 48min 19s, sys: 13.7 s, total: 48min 32s\n843.2s\t6\tWall time: 13min 19s\n843.9s\t7\t(39, 12)\n843.9s\t8\tshape: (1, 1)\n843.9s\t9\t┌─────────┐\n843.9s\t10\t│ date_id │\n843.9s\t11\t│ --- │\n843.9s\t12\t│ i16 │\n843.9s\t13\t╞═════════╡\n843.9s\t14\t│ 0 │\n843.9s\t15\t└─────────┘\n843.9s\t16\tshape: (1, 1)\n843.9s\t17\t┌─────────┐\n843.9s\t18\t│ time_id │\n843.9s\t19\t│ --- │\n843.9s\t20\t│ i16 │\n843.9s\t21\t╞═════════╡\n843.9s\t22\t│ 0 │\n843.9s\t23\t└─────────┘\n843.9s\t24\tshape: (39, 1)\n843.9s\t25\t┌───────────┐\n843.9s\t26\t│ symbol_id │\n843.9s\t27\t│ --- │\n843.9s\t28\t│ i8 │\n843.9s\t29\t╞═══════════╡\n843.9s\t30\t│ 23 │\n843.9s\t31\t│ 4 │\n843.9s\t32\t│ 26 │\n843.9s\t33\t│ 32 │\n843.9s\t34\t│ 21 │\n843.9s\t35\t│ … │\n843.9s\t36\t│ 31 │\n843.9s\t37\t│ 33 │\n843.9s\t38\t│ 28 │\n843.9s\t39\t│ 17 │\n843.9s\t40\t│ 11 │\n843.9s\t41\t└───────────┘\n843.9s\t42\ttest1\n843.9s\t43\tshape: (10, 94)\n843.9s\t44\t┌────────┬─────────┬─────────┬───────────┬───┬─────────────┬─────────────┬────────────┬────────────┐\n843.9s\t45\t│ row_id ┆ date_id ┆ time_id ┆ symbol_id ┆ … ┆ responder_5 ┆ responder_6 ┆ responder_ ┆ responder_ │\n843.9s\t46\t│ --- ┆ --- ┆ --- ┆ --- ┆ ┆ _lag_1 ┆ _lag_1 ┆ 7_lag_1 ┆ 8_lag_1 │\n843.9s\t47\t│ i64 ┆ i16 ┆ i16 ┆ i8 ┆ ┆ --- ┆ --- ┆ --- ┆ --- │\n843.9s\t48\t│ ┆ ┆ ┆ ┆ ┆ f32 ┆ f32 ┆ f32 ┆ f32 │\n843.9s\t49\t╞════════╪═════════╪═════════╪═══════════╪═══╪═════════════╪═════════════╪════════════╪════════════╡\n843.9s\t50\t│ 0 ┆ 0 ┆ 0 ┆ 0 ┆ … ┆ -0.036595 ┆ -1.305746 ┆ -0.795677 ┆ -0.143724 │\n843.9s\t51\t│ 1 ┆ 0 ┆ 0 ┆ 1 ┆ … ┆ -0.615652 ┆ -1.162801 ┆ -1.205924 ┆ -1.245934 │\n843.9s\t52\t│ 2 ┆ 0 ┆ 0 ┆ 2 ┆ … ┆ -0.378265 ┆ -1.57429 ┆ -1.863071 ┆ -0.027343 │\n843.9s\t53\t│ 3 ┆ 0 ┆ 0 ┆ 3 ┆ … ┆ -0.054984 ┆ 0.329152 ┆ -0.965471 ┆ 0.576635 │\n843.9s\t54\t│ 4 ┆ 0 ┆ 0 ┆ 4 ┆ … ┆ -0.597093 ┆ 0.219856 ┆ -0.276356 ┆ -0.90479 │\n843.9s\t55\t│ 5 ┆ 0 ┆ 0 ┆ 5 ┆ … ┆ -0.24035 ┆ -0.913801 ┆ -0.548867 ┆ -1.283726 │\n843.9s\t56\t│ 6 ┆ 0 ┆ 0 ┆ 6 ┆ … ┆ 0.517728 ┆ 0.896325 ┆ 1.068884 ┆ 1.57929 │\n843.9s\t57\t│ 7 ┆ 0 ┆ 0 ┆ 7 ┆ … ┆ -0.182887 ┆ -0.492168 ┆ -0.142915 ┆ -0.202081 │\n843.9s\t58\t│ 8 ┆ 0 ┆ 0 ┆ 8 ┆ … ┆ -0.225042 ┆ 0.95646 ┆ 2.185598 ┆ -0.435856 │\n843.9s\t59\t│ 9 ┆ 0 ┆ 0 ┆ 9 ┆ … ┆ -0.774299 ┆ -0.716492 ┆ -1.471419 ┆ -1.107083 │\n843.9s\t60\t└────────┴─────────┴─────────┴───────────┴───┴─────────────┴─────────────┴────────────┴────────────┘\n849.4s\t61\t[NbConvertApp] Converting notebook __notebook__.ipynb to notebook\n850.0s\t62\t[NbConvertApp] Writing 40602 bytes to __notebook__.ipynb\n852.4s\t63\t[NbConvertApp] Converting notebook __notebook__.ipynb to html\n853.6s\t64\t[NbConvertApp] Writing 332474 bytes to __results__.html", "3055889": "It is the code below I did test.\nI found that **I explicitly select the samples on time_id=0 in lag data**.\nI don't know why I do this... (since I don't know how the lag is provided actually)\n\n```python\ndef predict(test: pl.DataFrame, lags: pl.DataFrame | None) -> pl.DataFrame | pd.DataFrame:\n global lags_\n if lags is not None:\n # IMPORTANT -> select the samples time_id=0 with group_by before merge with test data\n lags_ = lags.group_by([\"date_id\", \"symbol_id\"], maintain_order=True).first().drop(\"time_id\")\n # Below lags_ cause error when merging with test\n # lags_ = lags.drop(\"time_id\")\n if lags_ is None:\n lags_ = test.select(\n pl.col(\"date_id\"), pl.col(\"symbol_id\"),\n *[pl.lit(None).alias(f\"{col}_lag_1\") for col in CFG.responder_cols]\n )\n \n \"\"\"Make a prediction.\"\"\"\n df_test = preprocessor.run(test.drop([\"row_id\", \"is_scored\"]), lags_, mode=\"test\")\n df_test = df_test.drop(CFG.metadata_cols).to_pandas()\n \n fold_pred = []\n for fold in range(CFG.n_folds):\n print(f\"\\n=== FOLD {fold} ===\")\n # split train & test\n test_x = df_test.copy()\n # scaling & encoding\n scaler = fold_output[fold][\"scaler\"]\n test_x[var_info[\"num\"]] = scaler.transform(test_x[var_info[\"num\"]])\n encoder = fold_output[fold][\"encoder\"]\n ohe_cols = fold_output[fold][\"ohe_cols\"]\n test_x = pd.concat([\n test_x.drop(var_info[\"cat\"], axis=1),\n pd.DataFrame(encoder.transform(test_x[var_info[\"cat\"]]), index=test_x.index.copy(), columns=ohe_cols)\n ], axis=1)\n print(\"shape info ->\", test_x.shape)\n # create model\n model = fold_output[fold][\"model\"]\n # inference\n test_pred = model.predict(test_x)\n # target clipping\n test_pred = np.clip(test_pred, a_min=-5.0, a_max=5.0)\n fold_pred.append(test_pred)\n\n fold_pred = np.stack(fold_pred, axis=0).mean(axis=0)\n predictions = test.select(\"row_id\", pl.Series(CFG.target_col, fold_pred, dtype=pl.datatypes.Float32))\n return predictions\n```", "3055890": "Thanks for your explanation :)", "3055969": "Thanks! Really helpful!" }, "source": "meta" }