{"id":4679,"date":"2024-02-27T18:58:35","date_gmt":"2024-02-27T09:58:35","guid":{"rendered":"https:\/\/blog.since2020.jp\/?p=4679"},"modified":"2024-02-27T18:58:35","modified_gmt":"2024-02-27T09:58:35","slug":"deepar%e3%81%a7%e8%a4%87%e6%95%b0%e6%99%82%e7%b3%bb%e5%88%97%e4%ba%88%e6%b8%ac%e3%81%99%e3%82%8b%e6%96%b9%e6%b3%95%e3%82%92%e8%a7%a3%e8%aa%ac%e3%80%90%e5%a4%96%e9%83%a8%e5%a4%89%e6%95%b0%e3%81%82","status":"publish","type":"post","link":"https:\/\/since2020.jp\/media\/deepar%e3%81%a7%e8%a4%87%e6%95%b0%e6%99%82%e7%b3%bb%e5%88%97%e4%ba%88%e6%b8%ac%e3%81%99%e3%82%8b%e6%96%b9%e6%b3%95%e3%82%92%e8%a7%a3%e8%aa%ac%e3%80%90%e5%a4%96%e9%83%a8%e5%a4%89%e6%95%b0%e3%81%82\/","title":{"rendered":"DeepAR\u3067\u8907\u6570\u6642\u7cfb\u5217\u4e88\u6e2c\u3059\u308b\u65b9\u6cd5\u3092\u89e3\u8aac\u3010\u5916\u90e8\u5909\u6570\u3042\u308a\u3011"},"content":{"rendered":"\n

DeepAR\u3067\u8907\u6570\u6642\u7cfb\u5217\u4e88\u6e2c\u3059\u308b\u65b9\u6cd5\u3092\u89e3\u8aac\u3010\u5916\u90e8\u5909\u6570\u3042\u308a\u3011<\/p>\n\n\n

\u306f\u3058\u3081\u306b<\/h2>\n

DeepAR\u3067\u6642\u7cfb\u5217\u4e88\u6e2c\u3092\u3059\u308b\u8a18\u4e8b\u304c\u65e5\u672c\u3067\u306f\u3042\u307e\u308a\u306a\u3044\u3067\u3059\u306d\u3002\u3053\u306e\u8a18\u4e8b\u3067\u306f\u5c11\u3057\u3084\u3063\u3066\u307f\u305f\u3067\u306f\u306a\u304f\u3001\u5916\u90e8\u5909\u6570\u306e\u53d6\u308a\u7d44\u307f\u3084\u51fa\u529b\u7d50\u679c\u3092\u3069\u3046\u5909\u63db\u3059\u308b\u304b\u306a\u3069\uff11\u901a\u308a\u306e\u3084\u308a\u65b9\u3092\u8a18\u8f09\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\r\n

\u30c7\u30fc\u30bf\u306e\u5909\u63db\u306e\u4ed5\u65b9\u3082\u72ec\u7279\u306a\u306e\u3067\u516c\u5f0f\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3092\u8aad\u307f\u89e3\u304f\u3053\u3068\u304c\u5927\u5909\u3067\u3057\u305f\u2026\u3002<\/p>\n\n

DeepAR\u306fAmazon\u3067\u63d0\u4f9b\u3055\u308c\u305f\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0<\/h2>\n

DeepAR\u306f\u3001Amazon SageMaker\u306a\u3069\u306eAWS\u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u3067\u63d0\u4f9b\u3055\u308c\u308b\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3067\u3059\u304c\u3001\u540c\u69d8\u306bAWS\u4ee5\u5916\u306e\u74b0\u5883\u3067\u3082GluonTS\u30e9\u30a4\u30d6\u30e9\u30ea\u3092\u4f7f\u7528\u3057\u3066\u5229\u7528\u3067\u304d\u307e\u3059\u3002\u3053\u306e\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306f\u3001RNN\u306e\u3088\u3046\u306a\u6df1\u5c64\u5b66\u7fd2\u30e2\u30c7\u30eb\u3092\u57fa\u306b\u3057\u3066\u304a\u308a\u3001\u6642\u7cfb\u5217\u30c7\u30fc\u30bf\u306b\u7279\u5316\u3057\u305f\u6df1\u5c64\u5b66\u7fd2\u30e2\u30c7\u30eb\u3092\u69cb\u7bc9\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u3067\u3059\u3002<\/p>\r\n

\"\"<\/p>\r\n

DeepAR\u306e\u7279\u9577\u306e\u4e00\u3064\u306f\u3001\u5916\u90e8\u5909\u6570\u306e\u7d44\u307f\u8fbc\u307f\u304c\u5bb9\u6613\u3067\u3042\u308b\u3053\u3068\u3067\u3059\u3002\u5916\u90e8\u5909\u6570\u306f\u6642\u7cfb\u5217\u30c7\u30fc\u30bf\u306b\u95a2\u9023\u3059\u308b\u88dc\u52a9\u7684\u306a\u60c5\u5831\u3067\u3042\u308a\u3001\u3053\u308c\u3092\u30e2\u30c7\u30eb\u306b\u7d44\u307f\u8fbc\u3080\u3053\u3068\u3067\u7cbe\u5ea6\u5411\u4e0a\u304c\u671f\u5f85\u3067\u304d\u307e\u3059\u3002\u307e\u305f\u3001DeepAR\u306f\u30e6\u30fc\u30b6\u30fc\u30d5\u30ec\u30f3\u30c9\u30ea\u30fc\u3067\u624b\u8efd\u306b\u30e2\u30c7\u30eb\u3092\u69cb\u7bc9\u3067\u304d\u308b\u305f\u3081\u3001\u521d\u5fc3\u8005\u3067\u3082\u5229\u7528\u3057\u3084\u3059\u3044\u3068\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u521d\u3081\u3066\u5229\u7528\u3059\u308b\u969b\u306f\u7406\u89e3\u304c\u96e3\u3057\u3044\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u304c\u3001\u305d\u306e\u67d4\u8edf\u6027\u3068\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u304b\u3089\u9ad8\u3044\u8a55\u4fa1\u3092\u53d7\u3051\u3066\u3044\u307e\u3059\u3002<\/p>\r\n

\u8a73\u3057\u3044\u89e3\u8aac\u306f\u3053\u3061\u3089\u306b\u3042\u308a\u307e\u3059\u3002<\/p>\r\n

\u301015%\u306e\u7cbe\u5ea6\u5411\u4e0a\u3011Amazon\u304c\u958b\u767a\u3057\u305f\u6642\u7cfb\u5217\u30e2\u30c7\u30eb\u3001DeepAR\u304c\u51c4\u3044 | Data Driven Knowledgebase (since2020.jp)<\/a><\/p>\r\n

<\/p>\n\n

\u30e9\u30a4\u30d6\u30e9\u30ea\u306e\u30a4\u30f3\u30dd\u30fc\u30c8<\/h2>\n

\u5168\u30b3\u30fc\u30c9\u306f\u4e00\u756a\u4e0b\u306b\u3042\u308b\u306e\u3067\u5148\u306b\u898b\u305f\u3044\u4eba\u306f\u30b9\u30ad\u30c3\u30d7\u3057\u3066\u4e0b\u3055\u3044\u3002<\/p>\r\n

GluonTS\u306b\u306fmxnet\u30e2\u30c7\u30eb\u3068torch\u30e2\u30c7\u30eb\u306e\uff12\u7a2e\u985e\u3042\u308b\uff06DeepAR\u3082\u3069\u3061\u3089\u306b\u3082\u3042\u308b\u3088\u3046\u3067\u3059\u304c\u3001\u79c1\u306ftorch\u306e\u65b9\u3067\u4e00\u901a\u308a\u51fa\u6765\u307e\u3057\u305f\u3002\u4ed6\u306e\u4eba\u3067\u306f\u4e21\u65b9\u4f7f\u3063\u3066\u3044\u308b\u4eba\u306a\u3069\u3044\u307e\u3057\u305f\u3002<\/p>\r\n

\r\n
pip install \"gluonts[torch]\"<\/code><\/pre>\r\n<\/div>\r\n
<\/p>\n\n
DeepAR\u3067\u5b66\u7fd2\u3055\u305b\u308b\u7528\u306b\u30c7\u30fc\u30bf\u306e\u524d\u51e6\u7406\u3092\u884c\u3046<\/h2>\n
\r\n
def deepar_process_data_1(df,threshold_date='2022-11-24',id_column,target<\/span>): #\u65e5\u4ed8\u578b\u306b\u5909\u63db df['date'] = pd.to_datetime(df['date']) #\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf\u306e\u671f\u9593\u306e\u76ee\u7684\u5909\u6570\u3092\u6d88\u53bb\uff08\u6d88\u3057\u3066\u3082\u6d88\u3055\u306a\u304f\u3066\u3082\u4e88\u6e2c\u7d50\u679c\u306f\u5909\u308f\u3089\u306a\u3044) df.loc[df['date'] > threshold_date,target] = None list_pref = list(df[id_column].unique()) return df,list_pref<\/code><\/pre>\r\n<\/div>\r\n
\u3053\u3053\u3067\u306f\u6700\u4f4e\u9650\u306e\u3053\u3068\u3057\u304b\u66f8\u304d\u307e\u305b\u3093\u304c\u3001df\u3092\u8aad\u307f\u3053\u307f\u3001\u5b66\u7fd2\u3068\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf\u306e\u5206\u5272\u70b9\u3068\u306a\u308b\u65e5\u4ed8\u3092\u5165\u529b\u3057\u3001\u8907\u6570\u6642\u7cfb\u5217\u30c7\u30fc\u30bf\u306e\u8b58\u5225ID\u306e\u30ea\u30b9\u30c8\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002
\r\n<\/p>\r\n
\r\n
def df_to_deepar(df,id_column,sta_cal_list,dyn_real_list,taget):\r\n#dyn_real_list\u5217\u3067null\u304c\u3042\u308b\u884c\u306f\u524a\u9664\r\ndf_cleaned = df.dropna(subset=dyn_real_list, how='any')\r\ndf_cleaned = df_cleaned.reset_index(drop=True)\r\n\r\n# \u30ab\u30c6\u30b4\u30ea\u5909\u6570\u3092\u30ab\u30c6\u30b4\u30ea\u578b\u306b\u5909\u63db\u3059\u308b\r\nunique_counts = []\r\n\r\n# \u30ab\u30c6\u30b4\u30ea\u5909\u6570\u3092\u30ab\u30c6\u30b4\u30ea\u578b\u306b\u5909\u63db\u3059\u308b\r\nfor column in sta_cal_list:\r\ndf_cleaned = df_cleaned.copy()\r\ndf_cleaned[column] = df_cleaned[column].astype('category')\r\nunique_count = df[column].nunique()\r\nunique_counts.append((unique_count))\r\nprint(f'{column}: {unique_count}')\r\n\r\ncalframes = [df_cleaned[sta_cal_list]]\r\n\r\n#DeepAR\u7528\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u5909\u63db\r\nresult = PandasDataset.from_long_dataframe(df_cleaned, target=taget, item_id=id_column,\r\ntimestamp='date', freq='D',static_features=pd.concat(calframes, axis=1),\r\nfeat_dynamic_real = dyn_real_list)\r\n\r\n# unique_counts\u306b\u306f\u5404\u30ab\u30e9\u30e0\u3054\u3068\u306e (\u30ab\u30e9\u30e0\u540d, \u30e6\u30cb\u30fc\u30af\u306a\u30af\u30e9\u30b9\u6570) \u304c\u683c\u7d0d\u3055\u308c\u308b\r\nprint(\"Unique counts for each column:\", unique_counts)\r\n\r\nreturn result ,unique_counts<\/code><\/pre>\r\n<\/div>\r\n
null\u304c\u3042\u308b\u884c\u3092\u524a\u9664\u3057\u3001\u30ab\u30c6\u30b4\u30ea\u5909\u6570\u3092\u30ab\u30c6\u30b4\u30ea\u578b\u306b\u5909\u63db\u3057\u307e\u3059\u3002<\/p>\r\n
\u305d\u3057\u3066\u3001\u5404\u30ab\u30c6\u30b4\u30ea\u5217\u3054\u3068\u306b\u30b0\u30eb\u30fc\u30d7\u304c\u3044\u304f\u3064\u3042\u308b\u306e\u304b\u30ab\u30a6\u30f3\u30c8\u3057\u3001\u30ea\u30b9\u30c8\u3067\u8fd4\u3057\u307e\u3059\u3002\uff08\u5b66\u7fd2\u6642\u306b\u4f7f\u7528\u3059\u308b\u306e\u3067return\u3067\u8fd4\u3059\uff09<\/p>\r\n
\u305d\u3057\u3066\u3001pandas\u578b\u304b\u3089PandasDataset\u578b\u306b\u5909\u63db\u3057\u307e\u3059\u3002\uff08\u540d\u524d\u306f\u4f3c\u3066\u3044\u308b\u3051\u3069\u5225\u7269\uff09<\/p>\r\n
timestamp\u3084freq\u306a\u3069\u5fc5\u8981\u306b\u5fdc\u3058\u3066\u5909\u3048\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\r\n
<\/p>\n\n
DeepAR\u3067\u5b66\u7fd2\u3057\u3001\u4e88\u6e2c\u5024\u3092\u5f97\u308b<\/h2>\n
\r\n
def do_deepar(df,unique_counts,list_pref,id_column):\r\n# \u958b\u59cb\u6642\u9593\u3092\u8a18\u9332\r\nstart_time = time.time()\r\n\r\n# DeepAREstimator\u306e\u4f5c\u6210\r\nestimator = DeepAREstimator(\r\nfreq='D',\r\nprediction_length=366,\r\nnum_layers=5,\r\ntrainer_kwargs={'max_epochs': 1},\r\nnonnegative_pred_samples=True,\r\ncontext_length=366,\r\ncardinality=unique_counts,\r\nnum_feat_dynamic_real=3\r\n)\r\n\r\npredictor = estimator.train(df)\r\nprint('\u5b66\u7fd2\u5b8c\u4e86')\r\nforecast_it, ts_it = make_evaluation_predictions(dataset=df, predictor=predictor, num_samples=100)\r\nforecast = list(forecast_it) # \u5c06\u6765\u4e88\u6e2c\u5024\u306e\u60c5\u5831\u304c\u5165\u3063\u305f\u30ea\u30b9\u30c8\r\ntss = list(ts_it) #\u30c7\u30fc\u30bf\u306b\u95a2\u3059\u308b\u30ea\u30b9\u30c8\r\n# tss\u3092\u30c7\u30fc\u30bf\u30d5\u30ec\u30fc\u30e0\u5316\r\nts_entry = pd.concat([tss[i] for i in range(len(list_pref))], axis=1)\r\nprint(ts_entry)\r\nts_entry.columns = list_pref\r\n\r\n# forecast\u3092\u8f9e\u66f8\u306b\u683c\u7d0d\r\nforecast_entry = dict(zip(list_pref, forecast))\r\n\r\n# \u7d42\u4e86\u6642\u9593\u3092\u8a18\u9332\r\nend_time = time.time()\r\n\r\n# \u7d4c\u904e\u6642\u9593\u3092\u8868\u793a\r\nelapsed_time = end_time - start_time\r\nprint(f\"\u51e6\u7406\u306b\u304b\u304b\u3063\u305f\u6642\u9593: {elapsed_time}\u79d2\")\r\n\r\nreturn forecast,tss<\/code><\/pre>\r\n<\/div>\r\n
cardinality\u306e\u5f15\u6570\u3067\u5148\u307b\u3069\u306e\u30ea\u30b9\u30c8\u3092\u683c\u7d0d\u3057\u307e\u3059\u3002<\/p>\r\n
\u305d\u3057\u3066\u3001prediction_length\u306b\u306f\u4e88\u6e2c\u671f\u9593\u3001num_feat_dynamic_real\u306b\u306f\u52d5\u7684\u9023\u7d9a\u306e\u6570\u5024\u306e\u7279\u5fb4\u91cf\u306e\u6570\u3092\u5165\u308c\u3066\u4e0b\u3055\u3044\u3002\u57fa\u672ccontext_length\u3068\u4e88\u6e2c\u671f\u9593\u306f\u540c\u3058<\/strong>\u306b\u306a\u308b\u3088\u3046\u3067\u3059\u3002<\/p>\r\n
\u305d\u3057\u3066make_evaluation_predictions\u3067\u4e88\u6e2c\u5024\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002tss\u306f\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf\u3067\u3042\u308a\u3001\u4e88\u6e2c\u3057\u305f\u5024\u306fforecast<\/strong>\u306b\u683c\u7d0d\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\r\n
\u4f8b\u3068\u3057\u3066\u4e88\u6e2c\u671f\u9593\u3092366\u65e5\u306b\u3057\u3066\u3044\u307e\u3059\u3002\u306a\u306e\u3067\u3082\u30571366\u65e5\u5206\u306e\u30c7\u30fc\u30bfdf\u3092\u5165\u308c\u3066\u4e88\u6e2c\u3059\u308b\u3068\u3001\u6700\u3082\u6700\u8fd1\u306e366\u65e5\u306e\u4e88\u6e2c\u304c\u8fd4\u3063\u3066\u304d\u307e\u3059\u3002\u9006\u306b\u8a00\u3046\u3068\u3001\u76f4\u8fd1\u306e366\u65e5\u306e\u76ee\u7684\u5909\u6570\u306e\u5024\u306f\u4f55\u306b\u3082\u4f7f\u7528\u3055\u308c\u307e\u305b\u3093\u3002<\/p>\r\n
<\/p>\r\n
\r\n
def return_all_preds(pred,id_column):\r\nall_preds = list()\r\n\r\nfor item in pred:\r\nID = item.item_id\r\np = item.samples.mean(axis=0)\r\np10 = np.percentile(item.samples, 10, axis=0)\r\np90 = np.percentile(item.samples, 90, axis=0)\r\ndates = pd.date_range(start=item.start_date.to_timestamp(), periods=len(p), freq='D')\r\nID_pred = pd.DataFrame({'date': dates, id_column: ID, 'prediction': p, 'p10': p10, 'p90': p90})\r\nall_preds += [ID_pred]\r\n\r\nall_preds = pd.concat(all_preds, ignore_index=True)\r\nall_preds[id_column] = all_preds[id_column].astype(int)\r\n\r\nreturn all_preds<\/code><\/pre>\r\n<\/div>\r\n
\u6700\u5f8c\u306b\u4e88\u6e2c\u5024\u306e\u6700\u4f4e\u9650\u306e\u52a0\u5de5\u3092\u884c\u3063\u3066\u7d42\u4e86\u3067\u3059\u3002<\/p>\n\n
\u5168\u30b3\u30fc\u30c9<\/h2>\n
\u5168\u30b3\u30fc\u30c9\u306f\u3053\u306e\u3088\u3046\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\r\n
\r\n
!pip install \"gluonts[torch]\"\r\n\r\nimport pandas as pd\r\nimport numpy as np\r\nimport matplotlib.pyplot as plt\r\nimport seaborn as sns\r\nimport time\r\nimport copy\r\nfrom gluonts.torch.model import deepar\r\nfrom gluonts.dataset.pandas import PandasDataset\r\nfrom gluonts.torch.model.deepar import DeepAREstimator\r\nfrom gluonts.evaluation.backtest import make_evaluation_predictions\r\n\r\ndef deepar_process_data_1(df,threshold_date='2022-11-24',id_column,taget):\r\n#\u65e5\u4ed8\u578b\u306b\u5909\u63db\r\ndf['date'] = pd.to_datetime(df['date'])\r\n#\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf\u306e\u671f\u9593\u306e\u76ee\u7684\u5909\u6570\u3092\u6d88\u53bb\uff08\u6d88\u3057\u3066\u3082\u6d88\u3055\u306a\u304f\u3066\u3082\u4e88\u6e2c\u7d50\u679c\u306f\u5909\u308f\u3089\u306a\u3044)\r\ndf.loc[df['date'] > threshold_date,taget] = None\r\nlist_pref = list(df[id_column].unique())\r\nreturn df,list_pref\r\n\r\ndef df_to_deepar(df,id_column,sta_cal_list,dyn_real_list,taget):\r\n#dyn_real_list\u5217\u3067null\u304c\u3042\u308b\u884c\u306f\u524a\u9664\r\ndf_cleaned = df.dropna(subset=dyn_real_list, how='any')\r\ndf_cleaned = df_cleaned.reset_index(drop=True)\r\n\r\n# \u30ab\u30c6\u30b4\u30ea\u5909\u6570\u3092\u30ab\u30c6\u30b4\u30ea\u578b\u306b\u5909\u63db\u3059\u308b\r\nunique_counts = []\r\n\r\n# \u30ab\u30c6\u30b4\u30ea\u5909\u6570\u3092\u30ab\u30c6\u30b4\u30ea\u578b\u306b\u5909\u63db\u3059\u308b\r\nfor column in sta_cal_list:\r\ndf_cleaned = df_cleaned.copy()\r\ndf_cleaned[column] = df_cleaned[column].astype('category')\r\nunique_count = df[column].nunique()\r\nunique_counts.append((unique_count))\r\nprint(f'{column}: {unique_count}')\r\n\r\ncalframes = [df_cleaned[sta_cal_list]]\r\n\r\n#DeepAR\u7528\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u5909\u63db\r\nresult = PandasDataset.from_long_dataframe(df_cleaned, target=taget, item_id=id_column,\r\ntimestamp='date', freq='D',static_features=pd.concat(calframes, axis=1),\r\nfeat_dynamic_real = dyn_real_list)\r\n\r\n# unique_counts\u306b\u306f\u5404\u30ab\u30e9\u30e0\u3054\u3068\u306e (\u30ab\u30e9\u30e0\u540d, \u30e6\u30cb\u30fc\u30af\u306a\u30af\u30e9\u30b9\u6570) \u304c\u683c\u7d0d\u3055\u308c\u308b\r\nprint(\"Unique counts for each column:\", unique_counts)\r\n\r\nreturn result ,unique_counts\r\n\r\ndef do_deepar(df,unique_counts,list_pref,id_column):\r\n# \u958b\u59cb\u6642\u9593\u3092\u8a18\u9332\r\nstart_time = time.time()\r\n\r\n# DeepAREstimator\u306e\u4f5c\u6210\r\nestimator = DeepAREstimator(\r\nfreq='D',\r\nprediction_length=366,\r\nnum_layers=5,\r\ntrainer_kwargs={'max_epochs': 1},\r\nnonnegative_pred_samples=True,\r\ncontext_length=366,\r\ncardinality=unique_counts,\r\nnum_feat_dynamic_real=3\r\n)\r\n\r\npredictor = estimator.train(df)\r\nprint('\u5b66\u7fd2\u5b8c\u4e86')\r\nforecast_it, ts_it = make_evaluation_predictions(dataset=df, predictor=predictor, num_samples=100)\r\nforecast = list(forecast_it) # \u5c06\u6765\u4e88\u6e2c\u5024\u306e\u60c5\u5831\u304c\u5165\u3063\u305f\u30ea\u30b9\u30c8\r\ntss = list(ts_it) # \u30c6\u30b9\u30c8\u30c7\u30fc\u30bf\u306b\u95a2\u3059\u308b\u30ea\u30b9\u30c8\r\n# tss\u3092\u30c7\u30fc\u30bf\u30d5\u30ec\u30fc\u30e0\u5316\r\nts_entry = pd.concat([tss[i] for i in range(len(list_pref))], axis=1)\r\nprint(ts_entry)\r\nts_entry.columns = list_pref\r\n\r\n# forecast\u3092\u8f9e\u66f8\u306b\u683c\u7d0d\r\nforecast_entry = dict(zip(list_pref, forecast))\r\n\r\n# \u7d42\u4e86\u6642\u9593\u3092\u8a18\u9332\r\nend_time = time.time()\r\n\r\n# \u7d4c\u904e\u6642\u9593\u3092\u8868\u793a\r\nelapsed_time = end_time - start_time\r\nprint(f\"\u51e6\u7406\u306b\u304b\u304b\u3063\u305f\u6642\u9593: {elapsed_time}\u79d2\")\r\n\r\nreturn forecast,tss\r\n\r\ndef return_all_preds(pred,id_column):\r\nall_preds = list()\r\n\r\nfor item in pred:\r\nID = item.item_id\r\np = item.samples.mean(axis=0)\r\np10 = np.percentile(item.samples, 10, axis=0)\r\np90 = np.percentile(item.samples, 90, axis=0)\r\ndates = pd.date_range(start=item.start_date.to_timestamp(), periods=len(p), freq='D')\r\nID_pred = pd.DataFrame({'date': dates, id_column: ID, 'prediction': p, 'p10': p10, 'p90': p90})\r\nall_preds += [ID_pred]\r\n\r\nall_preds = pd.concat(all_preds, ignore_index=True)\r\nall_preds[id_column] = all_preds[id_column].astype(int)\r\n\r\nreturn all_preds\r\n\r\n#\u4e88\u6e2c\u5024\u306e\u52a0\u5de5\uff08\u203b\u5fc5\u8981\u3067\u3042\u308c\u3070\uff09\r\ndef process_all_preds(all_preds,df_original):\r\nreturn all_preds\r\n\r\n\r\ndef main():\r\n#\u30c7\u30fc\u30bf\u3092\u53d6\u5f97\r\ndf = pd.read_csv('\u30d5\u30a1\u30a4\u30eb\u540d')\r\ndf_original = df.copy()\r\n#\u6642\u7cfb\u5217\u306e\u8b58\u5225ID\u306e\u8a2d\u5b9a\r\nid_column = 'ID'\r\ntaget = 'sales'\r\n#\u9759\u7684\u30ab\u30c6\u30b4\u30ea\u578b\u306e\u30ab\u30e9\u30e0\r\nsta_cal_list = ['area','customers']\r\n#\u52d5\u7684\u9023\u7d9a\u578b\u306e\u30ab\u30e9\u30e0\r\ndyn_real_list = ['lag_8','lag_9','lag_10']\r\n#\u30c7\u30fc\u30bf\u306e\u52a0\u5de5\r\ndf,list_pref = deepar_process_data_1(df,id_column,taget)\r\n#pandas\u578b\u3092\u5b66\u7fd2\u7528\u306b\u5909\u63db\r\ndf,unique_counts= df_to_deepar(df,id_column,sta_cal_list,dyn_real_list,taget)\r\n#\u5b66\u7fd2\u3057\u3001\u4e88\u6e2c\u5024\u3092\u8fd4\u3059\r\nforecast,tss = do_deepar(df,unique_counts,list_pref,id_column)\r\n#\u4e88\u6e2c\u5024\u3092panda\u578b\u306b\u5909\u63db\r\nall_preds = return_all_preds(forecast,id_column)\r\n#\u4e88\u6e2c\u5024\u306e\u52a0\u5de5\uff08\u203b\u5fc5\u8981\u3067\u3042\u308c\u3070\u8ffd\u52a0\uff09\r\nresult = process_all_preds(all_preds,df_original)\r\n\r\nreturn result\r\n\r\nresult = main()<\/code><\/pre>\r\n<\/div>","protected":false},"excerpt":{"rendered":"
DeepAR\u3067\u8907\u6570\u6642\u7cfb\u5217\u4e88\u6e2c\u3059\u308b\u65b9\u6cd5\u3092\u89e3\u8aac\u3010\u5916\u90e8\u5909\u6570\u3042\u308a\u3011 \u306f\u3058\u3081\u306b DeepAR\u3067\u6642\u7cfb\u5217\u4e88\u6e2c\u3092\u3059\u308b\u8a18\u4e8b\u304c\u65e5\u672c\u3067\u306f\u3042\u307e\u308a\u306a\u3044\u3067\u3059\u306d\u3002\u3053\u306e\u8a18\u4e8b\u3067\u306f\u5c11\u3057\u3084\u3063\u3066\u307f\u305f\u3067\u306f\u306a\u304f\u3001\u5916\u90e8\u5909\u6570\u306e\u53d6\u308a\u7d44\u307f\u3084\u51fa\u529b\u7d50\u679c\u3092\u3069\u3046\u5909\u63db\u3059\u308b\u304b\u306a\u3069\uff11\u901a […]<\/p>\n","protected":false},"author":83,"featured_media":3123,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","swell_btn_cv_data":"","footnotes":"","_wp_rev_ctl_limit":""},"categories":[1249],"tags":[507],"class_list":["post-4679","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-knowledge","tag-deepar"],"_links":{"self":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts\/4679","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/users\/83"}],"replies":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/comments?post=4679"}],"version-history":[{"count":0,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts\/4679\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/media\/3123"}],"wp:attachment":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/media?parent=4679"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/categories?post=4679"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/tags?post=4679"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}