ModelScope

1.35.3 · active · verified Fri Apr 10

ModelScope is an open-source model-as-a-service (MaaS) platform from Alibaba Damo Academy, providing a wide range of AI models (vision, NLP, audio, multimodal) for easy deployment and use. It abstracts complex AI model inference into a simple API and offers functionalities for model discovery, download, and fine-tuning. The library is actively developed, with its current version at 1.35.3 and frequent releases.

Warnings

breaking The `ms_dataset` module underwent significant refactoring to align with `datasets` library 4.x. If you're using dataset functionalities, ensure your `datasets` library is updated to version 4.0 or above, or fix potential breaking changes in your code.
Fix: Upgrade `datasets` to >=4.0.0 and review usage of `modelscope.msdatasets` for API changes. For example, `from modelscope.datasets` changed to `from modelscope.msdatasets`.
deprecated For security reasons, `delete_repo`, `delete_model`, and `delete_dataset` methods within the `HubApi` have been temporarily deprecated. Calling these methods will now issue a `DeprecationWarning`.
Fix: Avoid using these methods. If repository deletion is critical, consult ModelScope documentation for alternative, authenticated methods or use the web interface.
gotcha Hub API authentication token priority changed in v1.34.0. The token priority is now: function parameter > instance attribute > environment variable (`MS_TOKEN`). Be explicit if your application relies on a specific token source.
Fix: Explicitly pass the `token` parameter to `HubApi` methods or `pipeline` initializations if you wish to override tokens set in environment variables or instance attributes.
gotcha Users encountering `ModuleNotFoundError` for `packaging` or issues with `oss2` (e.g., global import conflicts) in versions prior to 1.35.1 might experience unexpected runtime errors or installation failures.
Fix: Upgrade to ModelScope v1.35.1 or later, which includes fixes for these dependency issues. Ensure `packaging` and `oss2` are correctly installed if needed for specific functionalities (e.g., cloud storage interaction).
gotcha Long-running dataset downloads using OSS (Object Storage Service) previously suffered from STS token expiration issues. This has been addressed, but older versions might still be susceptible.
Fix: Upgrade to ModelScope v1.34.0 or later, which includes a refactored OSS utility with automatic temporary credential refresh. This ensures dataset downloads can complete without interruption due to token expiration.

Install

pip install modelscope Base Installation
pip install 'modelscope[cv]' # For computer vision tasks pip install 'modelscope[nlp]' # For natural language processing tasks pip install 'modelscope[audio]' # For audio tasks With Task-Specific Dependencies (Optional)

Imports

pipeline

from modelscope.pipelines import pipeline

Tasks

from modelscope.utils.constant import Tasks

snapshot_download

from modelscope.hub.snapshot_download import snapshot_download

Model
```
from modelscope.models import Model
```
MsDataset
```
from modelscope.msdatasets import MsDataset
```
The dataset module was refactored, use `msdatasets`.

Quickstart

This quickstart demonstrates how to use ModelScope to perform image classification using a pre-trained model. It initializes a `pipeline` for a specific task and model, then processes an input image from a URL. Models are automatically downloaded and cached locally upon first use. For tasks requiring authentication (e.g., uploading to the Hub or accessing private models), ensure your ModelScope token is set via `os.environ['MS_TOKEN']` or passed explicitly.

import os
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

# Initialize an image classification pipeline
classifier = pipeline(task=Tasks.image_classification, model='damo/cv_resnest50_image-classification_damo')

# Example input image URL
image_url = 'https://modelscope.cn/api/v1/models/damo/cv_resnest50_image-classification_damo/repo/files/animal.JPEG'

# Perform inference
result = classifier(image_url)
print(f"Image classification result: {result}")

# You can also download a model first
# from modelscope.hub.snapshot_download import snapshot_download
# model_dir = snapshot_download('damo/cv_resnest50_image-classification_damo')
# local_classifier = pipeline(task=Tasks.image_classification, model=model_dir)
# print(f"Local inference result: {local_classifier(image_url)}")

view raw JSON →