v2.13.1 has breaking change in types. "Type 'FeatureExtractionPipeline' is missing the following properties from type 'Pipeline': task, model, tokenizer, processor, dispose"

[vickkhera]

System Info

Transformers.js version 2.13.1 or newer
Next.js version 13.5.6
typescript 5.3.3
macOS 14.2.1

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

I'm using the singleton pattern from the examples in the documentation:

class SentimentSingleton {
    static task = 'sentiment-analysis';
    static model = 'Xenova/distilbert-base-uncased-finetuned-sst-2-english';    // this is the default model for sentiment
    static instance : Promise<Pipeline>;
    static getInstance () {
        this.instance ??= pipeline(this.task, this.model);
        return this.instance;
    }
}

Starting with 2.13.1, it the next.js build step throws this error:

Type error: Type 'Promise<FeatureExtractionPipeline | TextClassificationPipeline | TokenClassificationPipeline | ... 19 more ... | DepthEstimationPipeline>' is not assignable to type 'Promise<Pipeline>'.
  Type 'FeatureExtractionPipeline | TextClassificationPipeline | TokenClassificationPipeline | ... 19 more ... | DepthEstimationPipeline' is not assignable to type 'Pipeline'.
    Type 'FeatureExtractionPipeline' is missing the following properties from type 'Pipeline': task, model, tokenizer, processor, dispose

  19 |     static instance : Promise<Pipeline>;
  20 |     static getInstance () {
> 21 |         this.instance ??= pipeline(this.task, this.model);
     |         ^
  22 |         return this.instance;
  23 |     }
  24 | }

Up to and including transformers.js version 2.13.0 this compiles and runs cleanly. The pipeline() function interface is changed between 2.13.0 and 2.13.1. A patch release should not change the interfaces. A breaking change like this should bump the minor version number at least.

Reproduction

Save this typescript file and open it in VS Code. Install transformers 2.13.1 or newer (tested through 2.13.4) Typescript will complain with the same error above at this.instance = assignment and the type of the first argument to pipeline().

By moving the string 'sentiment-analysis' into the call to pipeline() as a constant instead of using this.task, and using the TextClassificationPipeline type for the Promise, the typing works.

import { pipeline, env, Tensor, Pipeline } from '@xenova/transformers';

class SentimentSingleton {
    static task = 'sentiment-analysis';
    static model = 'Xenova/distilbert-base-uncased-finetuned-sst-2-english';    // this is the default model for sentiment
    static instance : Promise<Pipeline>;
    static getInstance () {
        this.instance ??= pipeline(this.task, this.model);
        return this.instance;
    }
}

[vickkhera]

It also changes the return type of the function generated by pipeline from Tensor to TextClassificationOutput | TextClassificationOutput [] and I have to cast the result output[0] as TextClassificationSingle in order to reference the .label and .score elements without TS errors.

[vickkhera]

The 'feature-extraction' / FeatureExtractionPipeline generated function still returns Tensor. 🤷🏽

[xenova]

Type 'FeatureExtractionPipeline' is missing the following properties from type 'Pipeline': task, model, tokenizer, processor, dispose

Hi there 👋 This was fixed in #500 and released in version 2.13.3:

could you please upgrade and try again?

using the TextClassificationPipeline type for the Promise

The Pipeline type was originally meant to be used as an abstract base class (and not to be used directly), and the correct type is indeed TextClassificationPipeline. Perhaps a follow-up PR can be made to allow Pipeline to be used as a type here too.

It also changes the return type of the function generated by pipeline from Tensor to TextClassificationOutput | TextClassificationOutput [] and I have to cast the result output[0] as TextClassificationSingle in order to reference the .label and .score elements without TS errors.

This is a tricky thing about the pipelines (which we tried to replicate from the python library): the type of the output depends on the input parameters. For example, with text classification, it changes depending on whether the input is provided as an array of strings, or if topk > 1. If you know how to add conditional types based on these parameters, I'd be happy to review a PR for it. Also, the return type of the text-classification pipeline has always been an array of objects (or arrays of objects) with the results.

The 'feature-extraction' / FeatureExtractionPipeline generated function still returns Tensor. 🤷🏽

The type of this pipeline didn't change :)

A patch release should not change the interfaces. A breaking change like this should bump the minor version number at least.

Yes, my apologies - this was an oversight on my part. Typing for the library is still a work-in-progress 😇 (and is an additional challenge since the python library doesn't type results from pipelines).

[kungfooman]

I believe part of the problem is that we don't allow the first pipeline argument to be an unspecified non-literal string aswell (side effect of the singleton pattern...), which should probably simply return a generic Pipeline.

I added it to my JSDoc fixes here if someone wants to test/review:

ec577d9

Example:

import {pipeline} from './pipelines.js';
class SentimentSingleton {
    static task = 'sentiment-analysis';
    static model = 'Xenova/distilbert-base-uncased-finetuned-sst-2-english';    // this is the default model for sentiment
    /** @type {Promise<import('./pipelines.js').Pipeline>} */
    static instance;
    static getInstance () {
        const ret = pipeline(this.task, this.model); // Promise<Pipeline>
        const retGivenButUnknown = pipeline("unknown-task-asd-123", this.model); // Promise<Pipeline>
        const retGiven = pipeline("text-generation", this.model); // Promise<TextGenerationPipeline>
        this.instance ??= pipeline(this.task, this.model);
        return this.instance;
    }
}

Output without any errors:

[kungfooman]

Without the fix in PR, you can also just keep the literal strings as-is, so you don't have to mess with the "type upgrade" to string:

class SentimentSingleton {
    /** @type {'sentiment-analysis'} */
    static task = 'sentiment-analysis';
    static model = 'Xenova/distilbert-base-uncased-finetuned-sst-2-english';    // this is the default model for sentiment
    /** @type {Promise<import('./pipelines.js').AllTasks['sentiment-analysis']>} */
    static instance;
    static getInstance () {
        this.instance ??= pipeline(this.task, this.model);
        return this.instance;
    }
}