ModelScope 실전 가이드: 검색부터 파인튜닝, 배포까지

Artificial Intelligence AI Infrastructure Technology Editors Pick Staff Tutorials In this tutorial, we explore ModelScope through a practical, end-to-end workflow that runs smoothly on Colab. We begin by setting up the environment, verifying dependencies, and confirming GPU availability so we can work with the framework reliably from the start. From there, we interact with the ModelScope Hub to search for models, download snapshots, load datasets, and understand how its ecosystem connects with familiar tools such as Hugging Face Transformers. As we move forward, we apply pretrained pipelines across NLP and computer vision tasks, then fine-tune a sentiment classifier on IMDB, evaluate its performance, and export it for deployment. Through this process, we build not only a working implementation but also a clear understanding of how ModelScope can support research, experimentation, and production-oriented AI workflows. Copy Code Copied Use a different Browser !pip install -q addict simplejson yapf gast oss2 sortedcontainers requests !pip install -q modelscope transformers>=4.37.0 datasets torch torchvision \ accelerate scikit-learn sentencepiece Pillow matplotlib evaluate optimum[exporters] import torch, os, sys, json, warnings, numpy as np warnings.filterwarnings("ignore") import addict; print("✅ addict OK") print(f"PyTorch: {torch.__version__}") print(f"CUDA available: {torch.cuda.is_available()}") if torch.cuda.is_available(): print(f"GPU: {torch.cuda.get_device_name(0)}") import modelscope print(f"ModelScope: {modelscope.__version__}") DEVICE = 0 if torch.cuda.is_available() else -1 from modelscope import snapshot_download from modelscope.hub.api import HubApi api = HubApi() print("\n🔍 Searching ModelScope Hub for 'bert' models...\n") try: models = api.list_models(filter_dict={"Search": "bert"}, sort="StarCount") for i, m in enumerate(models): if i >= 5: break print(f" • {m.get('Name', m.get('id', 'N/A'))}") except Exception as e: print(f" (Hub search may be unavailable outside China — {e})") model_dir = snapshot_download( "AI-ModelScope/bert-base-uncased", cache_dir="./ms_cache", ) print(f"\n✅ Model downloaded to: {model_dir}") print(" Files:", os.listdir(model_dir)[:8]) from modelscope.msdatasets import MsDataset print("\n📦 Loading 'imdb' dataset...\n") try: ds = MsDataset.load("imdb", split="train") print(f" Dataset size: {len(ds)} samples") sample = next(iter(ds)) print(f" Keys: {list(sample.keys())}") print(f" Text preview: {sample['text'][:120]}...") print(f" Label: {sample['label']} (0=neg, 1=pos)") except Exception as e: print(f" Falling back to HuggingFace datasets: {e}") from datasets import load_dataset ds = load_dataset("imdb", split="train") print(f" Dataset size: {len(ds)} samples") labels = [row["label"] for row in ds] print("\n Label distribution:") for label in sorted(set(labels)): count = labels.count(label) print(f" Label {label}: {count} ({count/len(labels)*100:.1f}%)") We set up the complete Colab environment and install all the libraries required for the tutorial. We verify important dependencies such as addict, check the PyTorch and CUDA setup, and confirm that ModelScope is installed correctly before moving forward. We then begin working with the ModelScope ecosystem by searching the hub for BERT models, downloading a model snapshot locally, loading the IMDB dataset, and examining its label distribution to understand the data we will use later. Copy Code Copied Use a different Browser from transformers import pipeline as hf_pipeline print("\n🧠 NLP PIPELINES\n") print("── 4a. Sentiment Analysis ──") sentiment = hf_pipeline( "sentiment-analysis", model="distilbert-base-uncased-finetuned-sst-2-english", device=DEVICE, ) test_texts = [ "ModelScope makes AI model access incredibly easy and intuitive!", "The documentation was confusing and the API kept returning errors.", "The weather today is partly cloudy with a slight breeze.", ] for text in test_texts: result = sentiment(text)[0] emoji = "🟢" if result["label"] == "POSITIVE" else "🔴" print(f' {emoji} {result["label"]} ({result["score"]:.4f}): "{text[:60]}..."') print("\n── 4b. Named Entity Recognition ──") ner = hf_pipeline( "ner", model="dbmdz/bert-large-cased-finetuned-conll03-english", aggregation_strategy="simple", device=DEVICE, ) ner_text = "Alibaba's ModelScope platform was developed in Hangzhou, China and competes with Hugging Face." entities = ner(ner_text) for ent in entities: print(f' 🏷️ {ent["word"]} → {ent["entity_group"]} (score: {ent["score"]:.3f})') print("\n── 4c. Zero-Shot Classification ──") zsc = hf_pipeline( "zero-shot-classification", model="facebook/bart-large-mnli", device=DEVICE, ) zsc_result = zsc( "ModelScope provides pretrained models for NLP, CV, and audio tasks.", candidate_labels=["technology", "sports", "politics", "science"], ) for label, score in zip(zsc_result["labels"], zsc_result["scores"]): bar = "█" * int(score * 30) print(f" {label:<12} {score:.3f} {bar}") print("\n── 4d. Text Generation (GPT-2) ──") generator = hf_pipeline( "text-generation", model="gpt2", device=DEVICE, ) gen_output = generator( "The future of open-source AI is", max_new_tokens=60, do_sample=True, temperature=0.8, top_p=0.9, num_return_sequences=1, ) print(f" 📝 {gen_output[0]['generated_text']}") print("\n── 4e. Fill-Mask (BERT) ──") fill_mask = hf_pipeline( "fill-mask", model=model_dir, device=DEVICE, ) mask_results = fill_mask("ModelScope is an open-source [MASK] for AI models.") for r in mask_results[:5]: print(f" ✏️ [MASK] → '{r['token_str']}' (score: {r['score']:.4f})") We focus on natural language processing pipelines and explore how easily we can run multiple tasks with pretrained models. We perform sentiment analysis, named entity recognition, zero-shot classification, text generation, and fill-mask prediction, providing a broad view of ModelScope-compatible inference workflows. As we test these tasks on sample inputs, we see how quickly we can move from raw text to meaningful model outputs in a unified pipeline. Copy Code Copied Use a different Browser print("\n👁️ COMPUTER VISION PIPELINES\n") print("── 5a. Image Classification (ViT) ──") img_classifier = hf_pipeline( "image-classification", model="google/vit-base-patch16-224", device=DEVICE, ) img_url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg" img_results = img_classifier(img_url) for r in img_results[:5]: print(f" 🖼️ {r['label']:<30} ({r['score']:.4f})") print("\n── 5b. Object Detection (DETR) ──") detector = hf_pipeline( "object-detection", model="facebook/detr-resnet-50", device=DEVICE, ) detections = detector(img_url) for d in detections[:5]: box = d["box"] print(f" 📦 {d['label']:<15} score={d['score']:.3f} box=({box['xmin']:.0f},{box['ymin']:.0f},{box['xmax']:.0f},{box['ymax']:.0f})") print("\n── 5c. Visualising Detections ──") from PIL import Image, ImageDraw import requests, matplotlib.pyplot as plt from io import BytesIO img = Image.open(BytesIO(requests.get(img_url).content)) draw = ImageDraw.Draw(img) colors = ["#58a6ff", "#3fb950", "#d2a8ff", "#f78166", "#ff7b72"] for i, d in enumerate(detections[:5]): box = d["box"] color = colors[i % len(colors)] draw.rectangle([box["xmin"], box["ymin"], box["xmax"], box["ymax"]], outline=color, width=3) draw.text((box["xmin"]+4, box["ymin"]+2), f"{d['label']} {d['score']:.2f}", fill=color) plt.figure(figsize=(10, 7)) plt.imshow(img) plt.axis("off") plt.title("DETR Object Detection") plt.tight_layout() plt.savefig("detection_result.png", dpi=150, bbox_inches="tight") plt.show() print(" ✅ Saved detection_result.png") print("\n🔄 HUGGINGFACE INTEROP\n") from transformers import AutoTokenizer, AutoModelForSequenceClassification print("── Approach A: snapshot_download (works for models on ModelScope Hub) ──") print(f" We already downloaded bert-base-uncased in Section 2: {model_dir}") print("\n── Approach B: Direct HF loading (works globally for any HF model) ──") hf_model_name = "distilbert-ba