Topik ini menjelaskan cara menggunakan model remote MaxCompute untuk menghasilkan deskripsi produk E-dagang secara otomatis.
Skenario
Pertumbuhan pesat bisnis E-dagang dan siaran langsung telah memperpendek siklus dari pengambilan foto hingga penerbitan daftar produk. Metode anotasi manual tradisional bergantung pada pengalaman dan penilaian subjektif para anotator, sehingga mahal untuk memproses data dalam jumlah besar serta menimbulkan tantangan dalam hal biaya dan kontrol kualitas pada aplikasi skala besar jangka panjang. Penggunaan AI untuk memproses data multimodal merupakan tren yang terus berkembang. Topik ini menyajikan studi kasus tentang pembuatan tag citra dan ringkasan dalam skenario E-dagang, yang menunjukkan cara mengakses data tidak terstruktur yang disimpan di Object Storage Service (OSS) melalui Object Table, lalu memprosesnya menggunakan model remote MaxCompute dan fungsi AI.
Keunggulan
Akses langsung data yang disimpan di OSS menggunakan Object Table MaxCompute, sehingga menghilangkan kebutuhan migrasi data yang kompleks.
Gunakan langsung model yang telah Anda deploy di Elastic Algorithm Service (EAS) untuk memanfaatkan sumber daya layanan model secara efisien.
Ekstrak dan proses data citra menggunakan pernyataan SQL sederhana dengan fungsi AI MaxCompute, tanpa perlu enkapsulasi user-defined function (UDF) yang kompleks.
Prasyarat
MaxCompute dan DataWorks telah diaktifkan, serta proyek MaxCompute telah dibuat. Untuk informasi selengkapnya, lihat Aktifkan MaxCompute dan DataWorks dan Buat proyek MaxCompute.
Model bahasa besar multimodal (LLM) telah dideploy di Elastic Algorithm Service (EAS) melalui Konsol Platform for AI (PAI). Untuk informasi selengkapnya, lihat Ikhtisar EAS.
OSS telah diaktifkan, dan citra yang akan diproses telah disimpan di OSS. Untuk informasi selengkapnya, lihat Aktifkan OSS.
Titik akhir publik EAS telah ditetapkan sebagai alamat jaringan eksternal yang tersedia untuk proyek MaxCompute Anda. Untuk informasi selengkapnya, lihat Edit alamat jaringan eksternal melalui manajemen proyek.
Persiapkan data
Unggah data citra E-dagang yang akan diproses ke OSS. Contoh ini menggunakan 50 citra poster produk dari dataset publik Alibaba Cloud Tianchi: Poster Design Text and Image Dataset.
Buat Tabel Objek.
SET odps.namespace.schema=true; -- Aktifkan sintaks skema tingkat penyewa. CREATE OBJECT TABLE IF NOT EXISTS image_demo WITH SERDEPROPERTIES ('odps.properties.rolearn'='acs:ram::1393************:role/aliyunodpsdefaultrole') LOCATION 'oss://oss-cn-shanghai-internal.aliyuncs.com/mllm-demo/jpg-test/'; -- Segarkan tabel untuk menyimpan metadata citra di cache MaxCompute. ALTER TABLE image_demo REFRESH METADATA; SELECT COUNT(*) AS ROW_COUNT FROM image_demo; +------------+ | row_count | +------------+ | 50 | +------------+
Buat model remote
Deploy model multimodal Omni di Konsol PAI untuk EAS. Dapatkan nama layanan model, Titik akhir akses, dan token.
Daftarkan model remote multimodal Omni di MaxCompute.
CREATE MODEL PAI_EAS_Qwen25_Omni_3B WITH VERSION v1 OPTIONS( MODEL_SOURCE_TYPE = 'REMOTE', MODEL_TYPE = 'MLLM', TASKS = 'text-generation', PAI_EAS_MODEL_NAME = 'Qwen2.5-Omni-3B', PAI_EAS_SERVICE_NAME = 'demo_remote_model', ENDPOINT = 'http://***********.cn-shanghai.pai-eas.aliyuncs.com', APIKEY = 'your-api-key', PAI_EAS_SYNC_MODE = 'true' ) COMMENT "PAI EAS remote model"; ALTER MODEL PAI_EAS_Qwen25_Omni_3B ADD VERSION v2 INPUT(data BINARY, promt STRING) WITH OPTIONS( MODEL_SOURCE_TYPE = 'REMOTE', MODEL_TYPE = 'MLLM', TASKS = 'text-generation', PAI_EAS_MODEL_NAME = 'Qwen2.5-Omni-3B', PAI_EAS_SERVICE_NAME = 'demo_remote_model', ENDPOINT = 'http://************.cn-shanghai.pai-eas.aliyuncs.com', APIKEY = 'your-api-key', PAI_EAS_SYNC_MODE = 'true' ) COMMENT "PAI EAS remote model binary input";Lihat model remote.
desc model PAI_EAS_Qwen25_Omni_3B; ID = 20250909024319251glhmocwqddl +------------------------------------------------------------------------------------+ | Informasi Model | +------------------------------------------------------------------------------------+ | Owner: ALIYUN$*********************** | | Project: pd_test_model | | Schema: default | | Nama Model: PAI_EAS_Qwen25_Omni_3B | | Jenis Model: MLLM | | Jenis Sumber: REMOTE | | Versi Default: v1 | | CreateTime: 2025-09-04 10:29:28 | | LastModifiedTime: 2025-09-09 10:43:11 | | Model ID: 46311b6397494a84ad23c5a4******** | | Komentar: PAI EAS remote model url input | +------------------------------------------------------------------------------------+ | Informasi Versi | +------------------------------------------------------------------------------------+ | Owner: ALIYUN$*********************** | | Project: pd_test_model | | Schema: default | | Nama Model: PAI_EAS_Qwen25_Omni_3B | | Jenis Model: MLLM | | Jenis Sumber: REMOTE | | Nama Versi: v1 | | ID Versi: c389823cd0324b72b1c0d55*********** | | Path: | | CreateTime: 2025-09-04 10:29:28 | | LastModifiedTime: 2025-09-04 10:29:28 | | apikey: Yzc4*******mNzFkMGM********4YzEwZjY1NTA*********NzkyNw== | | endpoint: http://13933481********.cn-shanghai.pai-eas.aliyuncs.com | | pai_eas_model_name: Qwen2.5-Omni-3B | | pai_eas_service_name: demo_eas_model | | pai_eas_sync_mode: true | | remote_service_type: PAI-EAS | +------------------------------------------------------------------------------------+ | Input | Type | Komentar | +------------------------------------------------------------------------------------+ +------------------------------------------------------------------------------------+ OK
Analisis data
Gunakan fungsi MaxCompute AI_GENERATE untuk mengklasifikasikan citra dan menghasilkan ringkasan deskriptif.
SET odps.namespace.schema=true;
SELECT
key,
AI_GENERATE(
PAI_EAS_Qwen25_Omni_3B, v1, image_url,
"Recognize and extract the product category from the E-commerce product sales poster. The result must be one of the following six options: Cosmetics, Apparel, Daily Necessities, Food, Other, Electronics. Do not include any other text or information."
) as item_catagory,
AI_GENERATE(
PAI_EAS_Qwen25_Omni_3B, v1, image_url,
"You are a professional E-commerce copywriter. Generate a product description summary based on the provided E-commerce product sales poster. Output requirements: 1. The summary must accurately reflect key information from the poster, such as product features and promotional details. 2. The language must be fluent and readable, without any special symbols or garbled characters, and no longer than 50 characters. 3. The description must be in Chinese, except for English brand names. Do not include any other English words. Example: Limited-time offer on pure cotton T-shirts, buy one get one free, only 3 days left in the event!"
) as item_description
from (
select GET_SIGNED_URL_FROM_OSS(
'pd_test_model.default.image_demo', key, 604800
) as image_url, key as key
from pd_test_model.default.image_demo
) Limit 10;
ID = 2025090909524183gsu********
Log view:
https://logview.alibaba-inc.com/logview/?h=http://server-ip:8008/daily&p=pd_test_model&i=2025********4183gsuq57sqddl&token=********
----------------------------------------------------------------------------------------------
STAGES STATUS TOTAL COMPLETED RUNNING PENDING BACKUP
M1_job_0 ................. TERMINATED 1 1 0 0 0
----------------------------------------------------------------------------------------------
STAGES: 01/01 [==========================>>] 100% ELAPSED TIME: 11.65 s
----------------------------------------------------------------------------------------------
+--------------------+-------------------+---------------------------------------------------------------------------------------------------------------------------------+
| key | item_catagory | item_description |
+--------------------+-------------------+---------------------------------------------------------------------------------------------------------------------------------+
| alimamazszw-1.jpg | Food | Limited-time offer on steakhouse beef. Place an order to receive a gift. Served with fresh basil. A delicious experience not to be missed. |
| alimamazszw-10.jpg | Electronics | New-feature fan on limited-time sale. Lowest price ever. Come and buy now! |
| alimamazszw-11.jpg | Electronics | Cool helmets for sale online with MIPS protection technology. Get an instant discount of CNY 270, bringing the price to CNY 729. Limited-time event. Don't miss out! |
| alimamazszw-12.jpg | Cosmetics | misspiggy "Ultra-fine Silky" eyeliner gel pencil. Soft, smooth, and long-lasting without smudging. Buy two and get CNY 10 off. Limited-time offer. Act now. |
| alimamazszw-13.jpg | Electronics | Special sale on IX35 wipers. Today's special offer. Buy now. |
| alimamazszw-14.jpg | Daily Necessities | Biological tissue glue from Beijing Yoshida Biological Protection Technology Co., Ltd. Can seal wounds. For animal experiments only. Valid for 1 year. |
| alimamazszw-15.jpg | Cosmetics | BodyAid hair growth serum. Awakens hair follicles, nurtures the scalp, and promotes new growth. |
| alimamazszw-16.jpg | Cosmetics | Limited-time offer on 3CE vitality blush. Buy one, get one free. Endless surprises on carnival day. |
| alimamazszw-17.jpg | Cosmetics | La Mer HR skin cream in black and white packaging. Six-piece set with six great gifts. Order now and get gifts worth CNY 2340. SF Express delivery. Limited-time offer. Limited quantity. |
| alimamazszw-18.jpg | Daily Necessities | Listerine mouthwash in various flavors on sale. A CNY 30 coupon is available. Limited quantity. Buy now! |
+--------------------+-------------------+---------------------------------------------------------------------------------------------------------------------------------+
-- Jika membaca data BINARY.
SELECT
key,
AI_GENERATE(
PAI_EAS_Qwen25_Omni_3B, v2, image_binary,
"Recognize and extract the product category from the E-commerce product sales poster. The result must be one of the following six options: Cosmetics, Apparel, Daily Necessities, Food, Other, Electronics. Do not include any other text or information."
) as item_catagory,
AI_GENERATE(
PAI_EAS_Qwen25_Omni_3B, v2, image_binary,
"You are a professional E-commerce copywriter. Generate a product description summary based on the provided E-commerce product sales poster. Output requirements: 1. The summary must accurately reflect key information from the poster, such as product features and promotional details. 2. The language must be fluent and readable, without any special symbols or garbled characters, and no longer than 50 characters. 3. The description must be in Chinese, except for English brand names. Do not include any other English words. Example: Limited-time offer on pure cotton T-shirts, buy one get one free, only 3 days left in the event!"
) as item_description
from (
select GET_DATA_FROM_OSS(
'pd_test_model.default.image_demo', key
) as image_binary, key as key
from pd_test_model.default.image_demo
) Limit 10;