I-Deutsche Bahn redefines ebikezela ukubikezela usebenzisa amamodeli weChronon – manje ayatholakala emakethe ye-Amazon Bedrock

Lokhu okuthunyelwe kubhalwe ku-Kilian Zimmerer noDaniel Ringler abavela eDeutsche Bahn.

Nsuku zonke, i-Deutsche Bahn (DB) isusa abagibeli abangaphezu kwezigidi eziyi-6.6 ngaphesheya kweJalimane, edinga ukubikezela kwesikhathi esinembile ngezinhloso ezahlukahlukene. Kodwa-ke, ukwakha amamodeli abikezelayo abikezela ngokwesiko ngokwesiko ubuchwepheshe obukhulu namaviki wesikhathi sokuthuthuka.

Namuhla, sijabule kakhulu ukuhlola ukuthi isikhathi se-Time Series silungiselela kanjani imodeli ye-Chronos-Bolt, sisanda kwethulwa nge-Amazon SageMaker Cumpstart, siguqula ukubikezela kochungechunge lwe-Amazon ngokunika amandla ukubikezela okunembile ngomzamo omncane. Ngenkathi izindlela zokubikezela kwezibikezelo zendabuko zithembele kumamodeli wezibalo, idatha ye-Chronos iphatha idatha yesikhathi esizofakwanga futhi isebenzisa i-FM eqeqeshwe ngaphambi kokukhiqiza imibhalo emikhulu yolimi (i-LLMS) ikhiqiza imibhalo. I-Chronos ikusiza ukufezekisa ukubikezela okunembayo ngokushesha, ukunciphisa kakhulu isikhathi sokuthuthuka uma kuqhathaniswa nezindlela zendabuko.

Kulokhu okuthunyelwe, sabelana nge-Deutsche Bahn Ukuchaza Kanjani Ukubikezela Ukubikezela Kubonisa Amamodeli We-Chronos, futhi kunikeze icala lokusebenzisa isibonelo ukukhombisa ukuthi ungaqala kanjani ukusebenzisa i-Chronos.

Chronos: Ukufunda Ulimi Oluchungechunge Lwesikhathi

Umndeni wemodeli we-Chronos umele impumelelo ngokubikezela kochungechunge lwesikhathi ngokuthola izakhiwo zemodeli yolimi. Ngokungafani namamodeli wendabuko wokubikezela amamodeli adinga ukuqeqeshwa kuma-datasets athile, ama-mronos angasetshenziselwa ukubikezela ngokushesha. Imodeli yangempela yeChronos yasheshe yaba yinombolo # 1 ethandwa kakhulu imodeli ekugwingweni kobuso ngo-2024, ekhombisa isidingo esinamandla se-FMS ngesikhathi sokubikezela kochungechunge lwesikhathi.

Ukwakha kule mpumelelo, sisanda kwethula ukwethula i-Chronos-Bolt, eletha ukunemba okuphezulu kwe-zero-shot uma kuqhathaniswa namamodeli we-Chronor asekuqaleni. Inikeza ukuthuthuka okulandelayo:

• Kufika izikhathi ezingama-250 ezivutha ngokushesha
• Izikhathi ezingama-20 zokukhumbula okungcono
• Ukuxhaswa Kwe-CPU Ukusekelwa, ukwenza izindleko zokubamba kuze kube izikhathi eziyi-10 ezingabizi kakhulu

Manje, ungasebenzisa i-Amazon Bedrock Market Proleple ukuhambisa i-Chronos-Bolt. I-Amazon Bedrock MarketProce iyikhono elisha e-Amazon Bedrock elenza abathuthukisi bakwazi ukuthola, bahlole, futhi basebenzise ama-FM angaphezu kwe-100 athandwayo, aqhamukayo, futhi akhethekile eceleni kokukhethwa kwamanje kwamamodeli aholwa embonini e-Amazon Bedrock.

INSELELE

I-Deutsche Bahn, Inkampani kazwelonke kazwelonke yaseJalimane, isebenza abagibeli abayizigidi eziyizinkulungwane eziyi-1,8 minyaka yonke ebangeni elide kanye nokuhamba kwabagibeli bezitimela zesifunda, okwenza kwaba ngomunye wabasebenza ngojantshi abakhulu bezitimela emhlabeni. Isikhathi esingaphezu kweshumi leminyaka, i-deutsche bahn ibilokhu ishabala kahle nama-AWS. Ama-AWS angumhlinzeki wamafu okuyisisekelo we-Deutsche Bahn kanye nomlingani onamasu we-DB Sysk, umxhaso ophethwe ngokuphelele we-DB AG eshayela amadijithali kuzo zonke izinkampani zeqembu.

Phambilini, izinqubo zokubikezela ze-Deutsche Bahn zazinamandla amakhulu ama-heterioous amaqembu abheke emaqenjini, adinga umzamo omkhulu wecala ngalinye elisha lokusebenzisa. Imithombo yedatha ehlukene edingekayo kusetshenziswa izindlela eziningi zokubikezela ezikhethekile, okuholele ekutheni kube yimizamo yezandla- kanye nesikhathi esikhuthele. Inkampani-ububanzi, i-deutsche bahn ikhombe inqwaba yezinqubo zokubikezela ezihlukile nezizimele. Amaqembu amancane athole kunzima ukuzithethelela ekuthuthukiseni izixazululo ezihlelwe ngokwezifiso zezidingo zezidingo zabo ezithile.

Isibonelo, iplanethi yokuhlaziywa kwedatha yeziteshi zezitimela zabagibeli ze-DB Infrago AG ihlanganisa futhi ihlaziye imithombo ehlukahlukene yedatha, kusuka kudatha yesimo sezulu kanye nemininingwane yokulungiswa kwesitshalo ku-Video Analytics. Njengoba unikezwe imithombo yedatha ehlukahlukene, indlela yokubikezela eyenzelwe umthombo owodwa wedatha ngokuvamile ayidluliswanga kweminye imithombo yedatha.

Ukuze wenze amakhono okubikezela wentando yeningi enhlanganweni, uDeutsche Bahn wayedinga indlela esebenza kahle futhi enobuthayo yokuphatha izimo ezahlukahlukene zokubikezela. Usebenzisa i-Chronos, uDeutsche Bahn ukhombisa ukuthi ubuchwepheshe obunqunyelwe bungaguqula kanjani imisebenzi yokubikezela izimangaliso.

Ukubuka konke

Iqembu elibhalise kuhlelo lwe-Deutsche Bahn's Accelerator Programme, i-leb yelebhu ye-DBSAL, yathuthukisa uhlelo lokubikezela lwesikhathi se-FM usebenzisa i-Chronos njengemodeli engaphansi, ngokubambisana ne-DB Infrago AG. Lolu hlelo lunikeza i-API yangaphakathi evikelekile engasetshenziswa amaqembu e-Deutsche Bahn enhlanganweni ukuze abikezele futhi asebenzise isikhathi esilula, ngaphandle kwesidingo sokuthuthukisa isoftware eyenziwe ngezifiso.

Umdwebo olandelayo ukhombisa ukwakheka okwenziwe lula kwendlela deutsche bahn esebenzisa ngayo i-Chronos.

Esikhathini sokusebenza kwesixazululo, umsebenzisi angadlula idatha yamaphepha e-Amazon API Gateway esebenza njengomnyango wangaphambili wangaphambili wezingcingo ze-API, ukuphatha ubuqiniso kanye nokugunyazwa. Ukuthola imininingwane engaphezulu yokuthi ungalinganisela kanjani ukufinyelela kuma-API kubasebenzisi abagunyaziwe kuphela, bheka ukulawula nokuphatha ukufinyelela ukuze uthole ama-API e-API. Ngemuva kwalokho, umsebenzi we-AWS Lambda usetshenziswa njenge-compute engenasici yokucutshungulwa kanye nezicelo ezidlulayo kwimodeli ye-Chronos ukuze uthole ukutholwa. Indlela esheshayo yokusingatha imodeli ye-Chronos ukusebenzisa imakethe ye-Amazon Bedrock noma indawo ye-sagemaker jumpstart.

Umthelela nezinhlelo zesikhathi esizayo

I-Deutsche Bahn ihlole insizakalo kumacala okusebenzisa amaningi, njengokubikezela izindleko zangempela zamaphrojekthi wokwakha kanye nokubikezela imali engenayo yanyangazonke yabasebenzi bezitolo abagibeli eziteshini zabagibeli. Ukuqaliswa kwamamodeli weChronos kwembula imiphumela ephoqelelayo. Ithebula elilandelayo libonisa imiphumela etholakele. Esimweni sokuqala sokusebenzisa, singabona ukuthi ezimweni ze-zero-vot (okusho ukuthi imodeli ayikaze ibone imininingwane ngaphambi), amamodeli we-Chronos angafinyelela ukunemba okuphakeme ngezindlela eziphakeme zezibalo njenge-autoarima kanye nama-autoets, yize lezi zindlela ziqeqeshwa ngokuqondile kwidatha. Ngaphezu kwalokho, kuzo zombili lezi zigameko, isikhathi sokuqashwa kwe-Chronos sifinyelela ezikhathini eziyi-100, futhi lapho sihlelwe kahle, amamodeli we-chronos ahlumela khona izindlela zendabuko kuzo zombili lezi zimo. Ngemininingwane engaphezulu ngomgwaqo olungisa kahle i-Chros, bheka ukubikezela okubikezelwa nge-Chronos – I-Autogluon.

Iphutha lilinganiswa ku-Smape. Ukuhlawuliswa kwamiswa ngemuva kwezinyathelo eziyi-10,000.

Ngokusekelwe ku-prototype ephumelelayo, i-Deutsche Bahn ithuthukisa insizakalo yokubikezela inkampani etholakala kuwo wonke amayunithi webhizinisi le-DB, asekela izimo ezihlukile zokubikezela. Okusemqoka, lokhu kuzokhunjulwa ngentando yeningi ukusetshenziswa kokubikezela kwenhlangano kuyo yonke inhlangano. Amaqembu acindezelwe izinsizakusebenza manje anikezwe amandla ukukhiqiza izibikezelo zawo, kanye nesikhathi sokulungiselela ukubikezela kungancishiswa kusuka emavikini kuya emahoreni.

Isibonelo Sokusetshenziswa Kwesibonelo

Ake sihambe ngesibonelo esisebenzayo sokusebenzisa i-Chronos-Bolt nge-Amazon Bedrock MarketPlace. Sizodlulisa amandla okusetshenziswa kwamandla abagibeli eziteshini zesitimela ezinde zaseJalimane neziteshi zesifunda usebenzisa idatha etholakala esidlangalaleni.

Izimfuneko

Kulokhu, uzosebenzisa ama-AWS SDK wePython (BOTO3) ukuxhumanisa ngokuhlelekile nge-Amazon Bedrock. Njengezimfuneko, udinga ukuba nemitapo yolwazi yePython boto3, pandasfuthi matplotlib Kufakiwe. Ngaphezu kwalokho, lungiselela ukuxhumeka kwi-akhawunti ye-AWS enjalo iBoto3 ingasebenzisa i-Amazon Bedrock. Ukuthola eminye imininingwane yokuthi ungayisetha kanjani iBoto3, bheka i-QuickStart – Boto3. Uma usebenzisa i-Python ngaphakathi kwe-Amazon Sagemaker Notebook, amaphakheji adingekayo asevele efakiwe.

Isimo sezulu sabagibeli

Okokuqala, layisha imininingwane ngokusetshenziswa kwamandla abagibeli bomlando. Okwalesi sibonelo, gxila esiteshini sesitimela 239:

import pandas as pd

# Load data
df = pd.read_csv(
    "
)
df_train_station = df[df["train_station"] == 239].reset_index(drop=True)

Okulandelayo, faka isiphetho seMamazon Bedrock MarketProke equkethe i-Chronos-Bolt. Lesi siphelo sisebenza njengensizakalo ebanjelwe, okusho ukuthi singathola izicelo eziqukethe imininingwane yochungechunge kanye nokubuyisa ukubikezela ukuphendula.

I-Amazon Bedrock izothatha ubunikazi be-AWS ubunikazi kanye neqhaza lokuphatha (i-IAM) ukuhlinzekwa isiphetho. Shintsha le khodi elandelayo ukuze ubhekise indima yakho. Ngokokufundisa ekwakheni indima yengozi, bhekisa kwindlela yokusebenzisa i-Sagemaker AI Premetion Roles.

import boto3
import time

def describe_endpoint(bedrock_client, endpoint_arn):
    return bedrock_client.get_marketplace_model_endpoint(endpointArn=endpoint_arn)[
        "marketplaceModelEndpoint"
    ]

def wait_for_endpoint(bedrock_client, endpoint_arn):
    endpoint = describe_endpoint(bedrock_client, endpoint_arn)
    while endpoint["endpointStatus"] in ["Creating", "Updating"]:
        print(
            f"Endpoint {endpoint_arn} status is still {endpoint['endpointStatus']}."
            "Waiting 10 seconds before continuing..."
        )
        time.sleep(10)
        endpoint = describe_endpoint(bedrock_client, endpoint_arn)
    print(f"Endpoint status: {endpoint['status']}")

bedrock_client = boto3.client(service_name="bedrock")
region_name = bedrock_client.meta.region_name
executionRole = "arn:aws:iam::account-id:role/ExecutionRole" # Change to your role

# Deploy Endpoint
body = {
        "modelSourceIdentifier": f"arn:aws:sagemaker:{region_name}:aws:hub-content/SageMakerPublicHub/Model/autogluon-forecasting-chronos-bolt-base/2.0.0",
        "endpointConfig": {
            "sageMaker": {
                "initialInstanceCount": 1,
                "instanceType": "ml.m5.xlarge",
                "executionRole": executionRole,
        }
    },
    "endpointName": "brmp-chronos-endpoint",
    "acceptEula": True,
 }
response = bedrock_client.create_marketplace_model_endpoint(**body)
endpoint_arn = response["marketplaceModelEndpoint"]["endpointArn"]

# Wait until the endpoint is created. This will take a few minutes.
wait_for_endpoint(bedrock_client, endpoint_arn)

Ngemuva kwalokho, cela isiphetho sokwenza isibikezelo. Thumela i-Payload kwi-Endpoint, okubandakanya amanani wezochungechunge lomlando kanye namapharamitha wokucushwa, njengezinga lokubikezela kanye namazinga anaga. I-Endpoint icubungula lokhu okufakwayo futhi ibuye impendulo equkethe amanani abikezelwe ngokusekelwe kwimininingwane enikeziwe.

import json

# Query endpoint
bedrock_runtime_client = boto3.client(service_name="bedrock-runtime")
body = json.dumps(
    {
        "inputs": [
            {"target": df_train_station["capacity"].values.tolist()},
        ],
        "parameters": {
            "prediction_length": 64,
            "quantile_levels": [0.1, 0.5, 0.9],
        }
    }
)
response = bedrock_runtime_client.invoke_model(modelId=endpoint_arn, body=body)
response_body = json.loads(response["body"].read())  

Manje ungabona ngeso lengqondo izibikezelo ezikhiqizwe yiChronos-Bolt.

import matplotlib.pyplot as plt

# Plot forecast
forecast_index = range(len(df_train_station), len(df_train_station) + 64)
low = response_body["predictions"][0]["0.1"]
median = response_body["predictions"][0]["0.5"]
high = response_body["predictions"][0]["0.9"]

plt.figure(figsize=(8, 4))
plt.plot(df_train_station["capacity"], color="royalblue", label="historical data")
plt.plot(forecast_index, median, color="tomato", label="median forecast")
plt.fill_between(
    forecast_index,
    low,
    high,
    color="tomato",
    alpha=0.3,
    label="80% prediction interval",
)
plt.legend(loc="upper left")
plt.grid()
plt.show()

Isibalo esilandelayo sibonisa umphumela.

Njengoba sikwazi ukubona ohlangothini olungakwesokudla lwegrafu eyandulele ebomvu, imodeli iyakwazi ukukhetha iphethini esingabona ngayo engxenyeni yesobunxele yesakhiwo (ngombala oluhlaza okwesibhakabhaka). Imodeli yeChronos ibikezela ukwehla okugcwele kulandelwa ama-spikes amabili amancane. Kuhle ukugqanyisisa ukuthi imodeli yabikezela ngempumelelo le mali isebenzisa ukubonwa okungu-zero, okungukuthi, ngaphandle kokuqeqeshwa kwimininingwane. Ukubuyela emuva emsebenzini wokuqala wokubikezela, singatolika ukuthi lesi siteshi sesitimela sisetshenziswe ngezimpelasonto.

Hlanza

Ukuze ugweme ukufukula izindleko ezingadingekile, sebenzisa ikhodi elandelayo ukususa isiphetho semodeli:

bedrock_client.delete_marketplace_model_endpoint(endpointArn=endpoint_arn)

# Confirm that endpoint is deleted
time.sleep(5)
try:
    endpoint = describe_endpoint(bedrock_client, endpoint_arn=endpoint_arn)
    print(endpoint["endpointStatus"])
except ClientError as err:
    assert err.response['Error']['Code'] =='ResourceNotFoundException'
    print(f"Confirmed that endpoint {endpoint_arn} was deleted")

Ukugcina

Umndeni wama-Chroses wamamodeli, ikakhulukazi imodeli entsha yeChronos-Bolt, imele ukuthuthuka okukhulu ekwenzeni ukubikezela okuchungechunge kwesikhathi okunembile. Ngokusebenzisa izinketho zokuthunyelwa ezilula nge-Amazon Bedrock MarketPlace kanye ne-SageMaker Jumpstart, izinhlangano manje zingasebenzisa izixazululo zokubikezela eziphakeme ngamahora amahora hhayi amasonto, ngenkathi kufeza ukunemba kombuso.

Noma ngabe ubikezela ukubikezela okuthengisayo, ukwenza imisebenzi elungiselelwe, noma ukuhlela ukwabiwa kwezinsizakusebenza, amamodeli weChronos ahlinzeka ngesisombululo esinamandla nesisebenzayo esingakwazi ukukala nezidingo zakho.

Mayelana nababhali

I-Kilian Zimmerer Unjiniyela we-AI noDeveps ku-DB Sysk GmbH eBerlin. Ngobuchwepheshe bakhe ekufundeni komshini wombuso kanye nokufunda okujulile, kanye nokuphathwa kwengqalasizinda yengqalasizinda, ushayela amaphrojekthi, achaze umbono wabo ngempumelelo, futhi asekela ukuqaliswa kwawo ngempumelelo ngaphakathi kwe-deutsche bahn.

UDaniel Ringler Ingabe unjiniyela wesoftware obhekele ekufundeni komshini ku-DB Sysk GmbH eBerlin. Ngaphezu komsebenzi wakhe wobungcweti, ungumhleli wokuzithandela wePydata Berlin, efaka isandla emphakathini wezemininingwane wesayensi kanye nePython Programming Community.

UPedro Eduardo Mercado Lopez Ungusosayensi osetshenzisiwe e-Amazon Web Services, lapho asebenza khona ngokubikezela kochungechunge lwezobasebenzi kanye nokuhlelwa kwamandla okugxilwe ku-Hierarchical Isikhathi Series namamodeli wesisekelo. Uthole i-PhD evela eSAarand University, eJalimane, enze ucwaningo ekuhlolweni okubonakalayo kwamagrafu asayiniwe nawokusayina.

USimeon Brüggenjürgen Ingabe ukwakhiwa kwezixazululo e-Amazon Web Services kususelwa eMunich, eJalimane. Ngasemuva ekucwaningweni kokufunda komshini, uSimeon osekelwe deutsche bahn kule phrojekthi.

UJohn Liu Ineminyaka engu-15 yesipiliyoni njengomphathi womkhiqizo kanye neminyaka engu-9 yesipiliyoni njenge-Portfolio Manager. E-AWS, uJohn ungumphathi womkhiqizo oyinhloko we-Amazon Bedrock. Phambilini, wayeyinhloko yomkhiqizo we-AWS Web3 / Blockchain. Ngaphambi kwe-AWS, uJohn wabamba izindima ezihlukene zobuholi bomkhiqizo ku-Protocols ye-Blockchain, izinkampani ze-Fintech futhi wasebenzisa iminyaka engu-9 njengemenenja yephothifoliyo kwizikhwama zezimali ezahlukahlukene.

UMichael Bohlke-Schneider ngumphathi wesayensi osetshenzisiwe e-Amazon Web Services. E-AWS, uMichael usebenza ekufundeni komshini nokubikezela, ngokugxila kumamodeli wesisekelo ukuthola idatha ehlelekile ne-thmol. Uthole i-PhD yakhe eBerlin yezobuchwepheshe, lapho asebenza khona ekubikezelweni kwesahlulelo yamaprotheni.

UFlorian Saupe Imenenja yomkhiqizo oyinhloko wezobuchwepheshe e-AWS AI / ML Ucwaningo olwesekele amaqembu wesayensi afana neqembu lomshini wegrafu, kanye namaqembu emishini ye-ML asebenza ekuqeqesheni okukhulu okusatshalaliswa, ukuqina kwephutha. Ngaphambi kokujoyina ama-AWS, iFlorian Holy Manageget ukuphathwa komkhiqizo okuzenzakalelayo eBosch, bekungumeluleki wecebo eMcKinsey & Company, futhi wasebenza njengesayensi yokulawula kanye ne-robotic scientist – inkambu lapho ephethe khona i-PhD.