Yakha amabhodi wezindaba ahambisanayo avumelanayo usebenzisa i-Amazon Nova e-Amazon Bedrock – Ingxenye 2

nimda September 4, 2025

0 0 8 minutes read

Yakha amabhodi wezindaba ahambisanayo avumelanayo usebenzisa i-Amazon Nova e-Amazon Bedrock – Ingxenye 2

Yize ukuqamba ngokucophelela kungavumi imiphumela emihle, ukufeza ukungaguquguquki kwebanga lokubuka ibanga kudinga ukuvumelanisa imodeli eyisisekelo uqobo. Ukwakha e-Prompt Engineering kanye nendlela yokuthuthukisa umlingiswa embozwe engxenyeni 1 yalolu chungechunge lwezingxenye ezimbili, manje sicindezela izinga lokungaguquguquki kwezinhlamvu ezithile ngokuhlelela kahle imodeli ye-Amazon Nova Canvas Foundation (FM). Ngamasu wokuhlelela okuhle, abadali bangafundisa imodeli ukuthi balondoloze ukulawulwa okuqondile kokuvela kwezinhlamvu, izinkulumo, kanye nezinto ze-stylistic kuzo zonke izigcawu eziningi.

Kulokhu okuthunyelwe, sithatha ifilimu elifushane elifushane, i-picchu, elikhiqizwa yi-fuzyppixel kusuka kumasevisi weWebhu (ama-AWS), lungiselela idatha yokuqeqesha ngokukhipha umlingiswa okhiye wenhlamvu, ngakho-ke, imodeli ye-Storyboard, ukuze ikhiqize ngokushesha imiqondo yebhodi elifanelekile lezithombe ezilandelayo njengezithombe ezilandelayo.

Ukubuka konke

Ukuze usebenzise ukugeleza komsebenzi okuzenzakalelayo, siphakamisa ukuthi isakhiwo esilandelanayo esisebenzisa izinsizakalo ze-AWS zokusebenzisa okuphelele kokuphela.

Ukuchitheka komsebenzi kuqukethe lezi zinyathelo ezilandelayo:

Umsebenzisi ulayisha impahla yevidiyo kwi-Amazon Isitoreji Service (I-Amazon S3) Ibhakede.
I-Amazon Elastic Chontaer Service (ECS yase-Amazon ECS iyabangelwa ukucubungula impahla yevidiyo.
I-Amazon ECS yakha amafreyimu, ikhetha lezo eziqukethe umlingiswa, bese zibeka izilimo zikhiqiza izithombe zokugcina.
I-Amazon ECS inxusa imodeli ye-Amazon NOVA (amazon Nova Pro) kusuka e-Amazon Bedrock ukudala amazwibela ezithombeni.
I-Amazon ECS ibhala izihloko zesithombe neMetadata kwibhakede le-S3.
Umsebenzisi usebenzisa indawo yezincwadi zokubhalela e-Amazon Sagemaker AI ukucela umsebenzi wokuqeqesha omodeli.
Umsebenzisi ulungile – ifuna imodeli ye-Amazon Nova Canvas ejwayelekile ngokucela i-Amazon Bedrock create_model_customization_job na- create_model_provisioned_throughput I-API Izingcingo ukudala imodeli yangokwezifiso etholakalayo yokutholwa.

Lokhu kuhamba komsebenzi kuhlelwe ngezigaba ezimbili ezihlukile. Isigaba sokuqala, ngezinyathelo 1-5, sigxile ekulungiseleleni idatha yokuqeqesha. Kulokhu okuthunyelwe, sihamba ngepayipi elizenzakalelayo lokukhipha izithombe kusuka kuvidiyo yokufaka bese sikhiqiza idatha yokuqeqeshwa efakiwe. Isigaba sesibili, ngezinyathelo 6-7, sigxile ekuhleleni kahle imodeli ye-Amazon Nova Canvas futhi enze ukutholwa kokuhlolwa kusetshenziswa imodeli eqeqeshwe ngokwezifiso. Kulezi zinyathelo zokugcina, sinikezela ngemininingwane yesithombe esenziwe sokulungiswa kanye nekhodi eyisibonelo ephelele kwi-GitHub Repository elandelayo ukukuqondisa ngenqubo.

Lungiselela idatha yokuqeqesha

Ake siqale ngesigaba sokuqala sokuhamba komsebenzi wethu. Esibonelweni sethu, sakha into yevidiyo ezenzakalelayo / ipayipi lokukhishwa komlingiswa lokukhipha izithombe eziphezulu ngamagama anezinkomba eziphezulu zisebenzisa lezi zinyathelo ezilandelayo.

Ukukhishwa kwezinhlamvu

Sincoma ozimele wokuqala we-sampling video ngezikhathi ezithile (ngokwesibonelo, uhlaka 1 ngomzuzwana). Ngemuva kwalokho, faka ilebula le-Amazon Rekognition Ilebula yokuthola nokusesha kobuso ukukhomba ozimele kanye nezinhlamvu zentshisekelo. Ukutholwa kwelebuli kungakhomba amalebula angaphezu kuka-2 000 ahlukile futhi athole izikhundla zawo ngaphakathi kwamafreyimu, okwenza kube kuhle ngokutholwa kokuqala kwezigaba zomlingiswa ojwayelekile noma izinhlamvu ezingezona ezesintu. Ukuze sihlukanise phakathi kwezinhlamvu ezahlukene, sibe sesisebenzisa isici sokuqalwa kwe-Amazon ukusesha ubuso eqoqweni. Lesi sici sikhomba futhi silandele abalingiswa ngokuqhathanisa ubuso babo ngokumelene nokuqoqwa kobuso bangaphambi kwangaphambili. Uma lezi zindlela ezimbili azinalo ngokwanele, singasebenzisa ama-Amazon Recononcy Labels ngokwezifiso ukuqeqesha imodeli yangokwezifiso yokuthola izinhlamvu ezithile. Umdwebo olandelayo ukhombisa lokhu kusebenza komsebenzi.

Ngemuva kokutholwa, silingisa umlingiswa ngamunye nge-Pixel padding efanelekile bese ugijimisa i-algorithm ye-ADUDUPTICT usebenzisa i-Amazon Titan Multimodal Embeddings Model ukususa izithombe ezifanayo ngenhla kwe-Humbary. Ukwenza lokhu kusisiza ukuba sakhe i-dataset ehlukahlukene ngoba amafreyimu angenakuphikwa noma acishe afane nawo angaholela ekutheni imodeli ngokweqile (lapho imodeli ifunda idatha yokuqeqeshwa ngokunembile, kufaka phakathi ukuguquguquka nokushintshashintsha kwayo, okwenza kube yinto engenzeki kahle kwidatha entsha, engabonakali). Singakwazi ukulinganisa umkhawulo wokufana ukuze sikwazi kahle lokho esikubheka njengezithombe ezifanayo, ngakho-ke singakwazi ukulawula kangcono ukulingana phakathi kokuhlukahluka kwesilinganiso kanye nokuqedwa kwe-redundancy.

Ukulebula kwedatha

Sikhiqiza izihloko zesithombe ngasinye sisebenzisa i-Amazon Nova Pro e-Amazon Bedrock bese silayisha isithombe bese ilebula ifayela lefayela le-Amazon S3. Le nqubo igxile ezicini ezimbili ezibucayi ze-Defect Engineering: Incazelo Yomlingiswa Ukusiza i-FM Khomba futhi babize izinhlamvu ngokuya ngezimpawu zabo ezihlukile, ngokwesibonelo, “umlingiswa opopayi”). Lokhu okulandelayo kuyisibonelo isifanekiso esisheshayo esisetshenziswa ngesikhathi senqubo yethu yokulebula idatha:

system_prompt = """ 
    You are an expert image description specialist who creates concise, natural alt
    text that makes visual content accessible while maintaining clarity and focus.
    Your task is to analyze the provided image and provide a creative description
    (20-30 words) that emphasizes the Three main characters, capturing the essential
    elements of their interaction while avoiding unnecessary details.
"""

prompt = """
    
    1. Identify the main characters in the image: Character 1, Character 2, and
        Character 3 at least one will be in the picture so provide at a minimum a
        description with at least one character name.
      - "Character 1" describe the first character, key traits, background, attributes.
      - "Character 2" describe the second character, key traits, background, attributes.
      - "Character 3" describe the third character, key traits, background, attributes. 
    2. Just state their name WITHOUT adding any standard characteristics.
    3. Only capture visual element outside the standard characteristics
    4. Capture the core interaction between them
    5. Include only contextual details that are crucial for understanding the scene
    6. Create a natural, flowing description using everyday language
    
    Here are some examples
    
       ...
    
    
    
    [Identify the main characters]
    [Assessment of their primary interaction]
    [Selection of crucial contextual elements]
    [Crafting of concise, natural description]
    
    
    {
        "alt_text": "[Concise, natural description focusing on the main characters]"
    }
    
    
    Note: Provide only the JSON object as the final response.

Ukukhishwa kwelebula kufomethwe njengefayela le-json Leli fayela le-JSONL labe selilayishwa ku-Amazon S3 ngokuziqeqesha. Okulandelayo kuyisibonelo sefayela:

{"image_ref": "s3://media-ip-dataset/characters/blue_character_01.jpg", "alt_text": "This
    animated character features a round face with large expressive eyes. The character
    has a distinctive blue color scheme with a small tuft of hair on top. The design is
    stylized with clean lines and a minimalist approach typical of modern animation."}
{"image_ref": "s3://media-ip-dataset/props/iconic_prop_series1.jpg", "alt_text": "This
    object appears to be an iconic prop from the franchise. It has a metallic appearance
    with distinctive engravings and a unique shape that fans would immediately recognize.
    The lighting highlights its dimensional qualities and fine details that make it
    instantly identifiable."}

Ukuqinisekiswa Komuntu

Kwamacala okusebenzisa amabhizinisi, sincoma ukufaka inqubo yomuntu in-the-loop ukuqinisekisa imininingwane efakiwe ngaphambi kokuqhubeka nokuqeqeshwa ngemodeli. Lokhu kuqinisekiswa kungasetshenziswa kusetshenziswa i-Amazon augmented AI (Amazon A2i), insizakalo esiza abafaka ama-Anvotators baqinisekise bobabili isithombe nekhwalithi ye-ncazo. Ngemininingwane engaphezulu, bheka Ukuqala Nge-Amazon Agegmented AI.

I-Fine-Tune Amazon Nova Canvas

Manje njengoba sinedatha yokuqeqesha, singakwazi ukwenza kahle imodeli ye-Amazon Nova Canvas e-Amazon Bedrock. I-Amazon Bedrock idinga ubunikazi be-AWS ubunikazi kanye neqhaza lezokuphepha Ngemininingwane engaphezulu, bheka imodeli yokufinyelela nokuphepha. Ungenza umsebenzi wokuhlelela okuhle ngqo kwi-Amazon Bedrock Console noma usebenzise iBoto3 API. Sichaza zombili izindlela kulokhu okuthunyelwe, futhi ungathola isampula yekhodi yokugcina ePicchu-Fineting.ipynb.

Dala umsebenzi omuhle kakhulu e-Amazon Bedrock Console

Ake siqale ngokwakha i-Amazon Nova Canvas umsebenzi omuhle kakhulu wokuhlelela i-Amazon Bedrock Console:

E-Amazon Bedrock Console, kufasitelana lokuhambisa, khetha Amamodeli wangokwezifiso ngaphansi kwa- Amamodeli weSisekelo.
Qoka Yenza ngokwezifiso imodeli bese Dala umsebenzi omuhle.

Use Dala imininingwane yomsebenzi omuhle Ikhasi, khetha imodeli ofuna ukwenza ngayo ngokwezifiso bese ufaka igama lemodeli elungiselelwe kahle.
Ku Ukucushwa Kwabasebenzi Isigaba, faka igama lomsebenzi futhi ngokuzithandela engeza omaki ukuzihlanganisa nayo.
Ku Idatha yokufaka Ingxenye, faka indawo ye-Amazon S3 yefayela ledatha yokuqeqeshwa.
Ku IHyperameters Isigaba, faka amanani we-hyperparameters, njengoba kukhonjisiwe ku-skrini elandelayo.

Ku Idatha yokukhipha Ingxenye, faka indawo ye-Amazon S3 lapho i-Amazon Bedrock kufanele ilondoloze umphumela womsebenzi.
Qoka Umsebenzi omuhle wemodeli Ukuqala inqubo yokuhleleka okuhle.

Le nhlanganisela ye-hyperparameter yaveza imiphumela emihle ngesikhathi sokuhlola kwethu. Ngokuvamile, ukwandisa inani lokufunda kwenza ukuthi isitimela imodeli sibe nolaka ngokwengeziwe, esivame ukuletha ukuhweba okuthokozisayo: Singakwazi ukufeza ukuguquguquka komlingiswa ngokushesha, kepha kungathinta ikhwalithi yesithombe. Sincoma indlela ehlelekile yokuguqula ama-hyperparemeter. Qala ngosayizi we-batch ophakanyisiwe kanye nesilinganiso sokufunda, bese uzama ukwanda noma ukwehlisa inani lezinyathelo zokuqeqesha kuqala. Uma imodeli ilwela ukufundwa idatha yakho ngisho nangemva kwezinyathelo ezingama-20,000 (inani eliphakeme elivunyelwe e-Amazon Bedrock), bese siphakamisa ukwandisa ubukhulu be-batch noma ukuguqula isilinganiso sokufunda siye phezulu. Lokhu kulungiswa, ngokusebenzisa okucashile, kungenza umehluko omkhulu ekusebenzeni kwemodeli yethu. Ngemininingwane engaphezulu ngama-hyperparameters, bheka ama-hyperparemeters wemodeli yokuqukethwe yokuqukethwe yokuqukethwe.

Dala umsebenzi omuhle usebenzisa i-Python SDK

I-Python Code Snippet elandelayo idala umsebenzi ofanayo wokuhleleka kahle usebenzisa i-Cranse_model_customization_job API:

bedrock = boto3.client('bedrock')
jobName = "picchu-canvas-v0"
# Set parameters
hyperParameters = {
        "stepCount": "14000",
        "batchSize": "64",
        "learningRate": "0.000001",
    }

# Create job
response_ft = bedrock.create_model_customization_job(
    jobName=jobName,
    customModelName=jobName,
    roleArn=roleArn,
    baseModelIdentifier="amazon.nova-canvas-v1:0",
    hyperParameters=hyperParameters,
    trainingDataConfig={"s3Uri": training_path},
    outputDataConfig={"s3Uri": f"s3://{bucket}/{prefix}"}
)

jobArn = response_ft.get('jobArn')
print(jobArn)

Lapho umsebenzi usuqedile, ungabuyisa okusha customModelARN Kusetshenziswa ikhodi elandelayo:

custom_model_arn = bedrock.list_model_customization_jobs(
    nameContains=jobName
)["modelCustomizationJobSummaries"][0]["customModelArn"]

Sebenzisa imodeli ehlelwe kahle

Ngokucushwa kwe-hyperpareter eyandulele, lo msebenzi wokuhlenga kahle ungathatha amahora ayi-12 ukuqeda. Uma sekuqediwe, kufanele ubone imodeli entsha kuhlu lwamamodeli wangokwezifiso. Ungadala ukugcwaliseka okunikeziwe ukuze usingathe imodeli. Ngemininingwane engaphezulu kokufakwa okuhlinzekelwe kanye nezinhlelo ezahlukene zokuzibophezela, bheka Ukwandisa amandla okuhambisa amamodeli ngokufakwa okunikeziwe nge-Foodgetf

Sebenzisa imodeli e-Amazon Bedrock Console

Ukuze usebenzise imodeli e-Amazon Bedrock Console, uqedele lezi zinyathelo ezilandelayo:

E-Amazon Bedrock Console, khetha Amamodeli wangokwezifiso ngaphansi kwa- Amamodeli weSisekelo kufasitelana lokuzulazula.
Khetha imodeli entsha yangokwezifiso bese ukhetha Ukuthenga okuhlinzekiwe okuhlinzekiwe.

Ku Imininingwane yokufaka okufakiwe Ingxenye, faka igama ngokutholwa okunikelwe.
Ngaphansi kwa- Khetha imodelikhetha imodeli yangokwezifiso osanda kwenza.
Bese ucacisa igama lokuzibophezela kanye namayunithi wemodeli.

Ngemuva kokuthenga okunikeziwe okunikezwayo, igama elisha le-Amazon Resource (ARN) lenziwe. Ungacela le arn lapho okunikezwayo okuhlinzekiwe kusebenza.

Sebenzisa imodeli usebenzisa i-Python SDK

Ikhodi elandelayo ye-Python Code idala ukufakwa kokudla okunikelwe kusetshenziswa i-Dracce_Provised_Model_throughtup:

custom_model_name = "picchu-canvas-v0"

# Create the provision throughput job and retrieve the provisioned model id
provisioned_model_id = bedrock.create_provisioned_model_throughput(
    modelUnits=1,
    # create a name for your provisioned throughput model
    provisionedModelName=custom_model_name, 
    modelId=custom_model_arn
)['provisionedModelArn']

Hlola imodeli ehlelwe kahle

Lapho ukugcwaliswa okuhlinzekwayo kuphila, singasebenzisa i-Snippet elandelayo yekhodi ukuvivinya imodeli yangokwezifiso futhi sihlole ukukhiqiza izithombe ezintsha ze-picchu:

import json
import io
from PIL import Image
import base64

def decode_base64_image(img_b64):
    return Image.open(io.BytesIO(base64.b64decode(img_b64)))
    
def generate_image(prompt,
                   negative_prompt="text, ugly, blurry, distorted, low
                       quality, pixelated, watermark, text, deformed", 
                   num_of_images=3,
                   seed=1):
    """
    Generate an image using Amazon Nova Canvas.
    """

    image_gen_config = {
            "numberOfImages": num_of_images,
            "quality": "premium",
            "width": 1024,  # Maximum resolution 2048 x 2048
            "height": 1024,  # 1:1 ratio
            "cfgScale": 8.0,
            "seed": seed,
        }

    # Prepare the request body
    request_body = {
        "taskType": "TEXT_IMAGE",
        "textToImageParams": {
            "text": prompt,
            "negativeText": negative_prompt,  # List things to avoid
        },
        "imageGenerationConfig": image_gen_config
    } 

    response = bedrock_runtime.invoke_model(
        modelId=provisioned_model_id,
        body=json.dumps(request_body)
    )

    # Parse the response
    response_body = json.loads(response['body'].read())

    if "images" in response_body:
        # Extract the image
        return [decode_base64_image(img) for img in response_body['images']]
    else:
        return
seed = random.randint(1, 858993459)
print(f"seed: {seed}")

images = generate_image(prompt=prompt, seed=seed)


Ubuso be-badlwana bukhombisa ukuhlanganiswa kovalo nokuzimisela. Umama uyaguqa eceleni kwakhe, umhawu wakhe. Isimo sendawo siyabonakala ngemuva.	Ubuso obuhlaziyi bemile enezitebhisi ezinde zokhuni ezinwebeka phansi. Halfway phansi i-ladder yi-mayu ngenkulumo enqunyiwe ebusweni bakhe. Izandla ezincane zikaMayu zibambe izinhlangothi zesitebhisi ngokuqinile njengoba ebeka izinyawo zakhe ngokucophelela ku-rug ngayinye. Imvelo ezungezile ikhombisa ukugqama komhlaba okumatasa, ezintabeni.	UMayu ume ngokuziqhenya emnyango wesakhiwo esilula sesikole. Ubuso bakhe bumamatheka ngokumamatheka okubanzi, okubonisa ukuziqhenya nokufeza okuthile.

Hlanza

Ukugwema ukukhokhiswa kwe-AWS ngemuva kokuqedwa kokuhlolwa, gcwalisa izinyathelo zokuhlanza ePicchu-FinTuning.ipynb bese ususa izinsiza ezilandelayo:

I-Amazon SageMaker Studio Domain
Imodeli enhle ye-Amazon Nova ephelele kanye nokunikezwa kokuphela kokuphela

Ukugcina

Kulokhu okuthunyelwe, sikhombise ukuthi ungakuphakamisa kanjani ukuguquguquka kwesimo kanye nesitayela ebhokisini lezindaba kusuka engxenyeni 1 ngocingo oluhle lwe-Amazon Nova Canvas e-Amazon Bedrock. Ukuqhekeka kwethu kokusebenza okuphelele kuhlanganisa ukucubungula kwevidiyo okuzenzakalelayo, ukukhishwa kwezinhlamvu ezihlakaniphile usebenzisa ukubhekisisa i-Amazon, kanye nokwenza ngokwezifiso okunembile usebenzisa i-Amazon Bedrock ukudala isisombululo esigcina ukwethembana okubonakalayo. Ngokulungiselela kahle imodeli ye-Amazon Nova canvas kuzinhlamvu ezithile nezitayela, sifinyelele ezingeni lokuvumelana okudlula ubunjiniyela obujwayelekile, ngakho-ke amaqembu e-Creative angakhiqiza amabhodi wezindaba ezisezingeni eliphakeme emahoreni hhayi amasonto. Qala ukuzama ngeNova Canvas Ukuhleleka okuhle namhlanje, ngakho-ke ungaphakamisa ukuphendula kwakho izindaba ngokuvumelana okungcono nokuvumelana okungcono.

Mayelana nababhali

UDkt Achin Jain Ingabe i-Selight yasebenzisa usosayensi e-Amazon Agi, lapho asebenza khona ekwakheni amamodeli wesisekelo ahlukahlukene. Uletha iminyaka eyi-10 + yemboni ehlangene nolwazi lokucwaninga ngezifundo. Uhole ukuthuthukiswa kwamamojula ambalwa we-Amazon Nova Canvas kanye ne-Amazon Titan Image Generator, kufaka phakathi ukuqondiswa okuhle okuhle (SFT), imodeli ngokwezifiso, ukwenza ngokwezifiso okusheshayo, nokuholwa nge-Colour Palette.

James Wu Ungumakhi wesixazululo esiphezulu se-AI / ML kuma-AWS. Ukusiza amakhasimende aklanywe futhi akhe izixazululo ze-AI / ML. Umsebenzi kaJakobe uhlanganisa amacala anhlobonhlobo we-ML ahlukahlukene, enentshisekelo eyinhloko embonweni wekhompyutha, ukufunda okujulile, kanye nokulinganisa ml ngaphesheya kwebhizinisi. Ngaphambi kokujoyina ama-AWS, uJames wayengumakhi, unjiniyela, kanye nobuholi bezobuchwepheshe iminyaka engaphezu kwengu-10, kubandakanya iminyaka engu-6 enjiniyela kanye neminyaka emi-4 kwezimboni zokukhangisa nokukhangisa.

URandy Ridgley Ingabe ukwakhiwa kwesoftware kuyagxila kuma-analytics wesikhathi sangempela ne-AI. Ngobuchwepheshe ekwakheni amachibi edatha namapayipi. URandy usiza izinhlangano ziguqula ukusakazwa kwedatha okuhlukahlukene kube yimininingwane esebenzayo. Ubheka izixazululo ze-IOT, ama-analytics, kanye nokusetshenziswa kwengqalasizinda. Njengomhlinzeki womthombo ovulekile kanye nomholi wezobuchwepheshe, iRandy ihlinzeka ngolwazi olujulile lobuchwepheshe ukuletha izixazululo zedatha ezikali ezindaweni zonke.