learnware 0.4.0.post1

The learnware package supports the submission, usability testing, organization, identification, deployment, and reuse of learnware.

中文 | English

Introduction

Learnware paradigm was proposed by Professor Zhi-Hua Zhou in 2016 [1, 2]. In the learnware paradigm, developers worldwide can share models with the learnware dock system, which effectively searches for and reuse learnware(s) to help users solve machine learning tasks efficiently without starting from scratch.

The learnware package provides a fundamental implementation of the central concepts and procedures within the learnware paradigm. Its well-structured design ensures high scalability and facilitates the seamless integration of additional features and techniques in the future.

In addition, the learnware package serves as the engine for the Beimingwu System and can be effectively employed for conducting experiments related to learnware.

[1] Zhi-Hua Zhou. Learnware: on the future of machine learning. Frontiers of Computer Science, 2016, 10(4): 589–590
[2] Zhi-Hua Zhou. Machine Learning: Development and Future. Communications of CCF, 2017, vol.13, no.1 (2016 CNCC keynote)

Learnware Paradigm

A learnware consists of a high-performance machine learning model and specifications that characterize the model, i.e., "Learnware = Model + Specification". These specifications, encompassing both semantic and statistical aspects, detail the model's functionality and statistical information, making it easier for future users to identify and reuse these models.

The above diagram illustrates the learnware paradigm, which consists of two distinct stages:

• Submitting Stage: Developers voluntarily submit various learnwares to the learnware doc system, and the system conducts quality checks and further organization of these learnwares.
• Deploying Stage: When users submit task requirements, the learnware doc system automatically selects whether to recommend a single learnware or a combination of multiple learnwares and provides efficient deployment methods. Whether it’s a single learnware or a combination of multiple learnwares, the system offers convenient learnware reuse interfaces.

Framework and Infrastructure Design

The architecture is designed based on the guidelines including decoupling, autonomy, reusability, and scalability. The above diagram illustrates the framework from the perspectives of both modules and workflows.

• At the workflow level, the learnware package consists of Submitting Stage and Deploying Stage.

Module	Workflow
Submitting Stage	The developers submit learnwares to the learnware market, which conducts usability checks and further organization of these learnwares.
Deploying Stage	The learnware market recommends learnwares according to users’ task requirements and provides efficient reuse and deployment methods.

• At the module level, the learnware package is a platform that consists of Learnware, Market, Specification, Model, Reuse, and Interface modules.

Module	Description
Learnware	The specific learnware, consisting of specification module, and user model module.
Market	Designed for learnware organization, identification, and usability testing.
Specification	Generating and storing statistical and semantic information of learnware, which can be used for learnware search and reuse.
Model	Including the base model and the model container, which can provide unified interfaces and automatically create isolated runtime environments.
Reuse	Including the data-free reuser, data-dependent reuser, and aligner, which can deploy and reuse learnware for user tasks.
Interface	The interface for network communication with the Beimingwu backend.

Quick Start

Installation

Learnware is currently hosted on PyPI. You can easily install learnware by following these steps:

pip install learnware

In the learnware package, besides the base classes, many core functionalities such as "learnware specification generation" and "learnware deployment" rely on the torch library. Users have the option to manually install torch, or they can directly use the following command to install the learnware package:

pip install learnware[full]

Note: However, it's crucial to note that due to the potential complexity of the user's local environment, installing learnware[full] does not guarantee that torch will successfully invoke CUDA in the user's local setting.

Prepare Learnware

In the learnware package, each learnware is encapsulated in a zip package, which should contain at least the following four files:

• learnware.yaml: learnware configuration file.
• __init__.py: methods for using the model.
• stat.json: the statistical specification of the learnware. Its filename can be customized and recorded in learnware.yaml.
• environment.yaml or requirements.txt: specifies the environment for the model.

To facilitate the construction of a learnware, we provide a Learnware Template that users can use as a basis for building their own learnware. We've also detailed the format of the learnware zip package in Learnware Preparation.

Learnware Package Workflow

Users can start a learnware workflow according to the following steps:

Initialize a Learnware Market

You can initialize a basic Learnware Market named "demo" using the code snippet below:

from learnware.market import instantiate_learnware_market

# instantiate a demo market
demo_market = instantiate_learnware_market(market_id="demo", name="easy", rebuild=True)

Upload Learnware

Before uploading your learnware to the Learnware Market, you'll need to create a semantic specification, semantic_spec. This involves selecting or inputting values for semantic tags to describe the features of your task and model.

For instance, the following code illustrates the semantic specification for a Scikit-Learn type model. This model is tailored for education scenarios and performs classification tasks on tabular data:

from learnware.specification import generate_semantic_spec

semantic_spec = generate_semantic_spec(
    name="demo_learnware",
    data_type="Table",
    task_type="Classification",
    library_type="Scikit-learn",
    scenarios="Education",
    license="MIT",
)

After preparing the semantic specification, you can insert your learnware into the learnware market using a single line of code:

demo_market.add_learnware(zip_path, semantic_spec)

Here, zip_path is the file path of your learnware zip package.

Semantic Specification Search

To identify learnwares that align with your task's purpose, you'll need to provide a semantic specification, user_semantic, that outlines your task's characteristics. The Learnware Market will then perform an initial search based on user_semantic, which filters learnwares by considering the semantic information of your task.

# construct user_info, which includes a semantic specification
user_info = BaseUserInfo(id="user", semantic_spec=semantic_spec)

# search_learnware: performs semantic specification search when user_info doesn't include a statistical specification
search_result = demo_market.search_learnware(user_info) 
single_result = search_results.get_single_results()

# single_result: the List of Tuple[Score, Learnware] returned by semantic specification search
print(single_result)

Statistical Specification Search

If you generate and provide a statistical specification file rkme.json, the Learnware Market will conduct learnware identification based on statistical information, and return more targeted models. Using the API we provided, you can easily generate this statistical specification locally.

For example, the code below executes learnware search when using Reduced Kernel Mean Embedding (RKME) as the statistical specification:

import learnware.specification as specification

user_spec = specification.RKMETableSpecification()

# unzip_path: directory for unzipped learnware zipfile
user_spec.load(os.path.join(unzip_path, "rkme.json"))
user_info = BaseUserInfo(
    semantic_spec=user_semantic, stat_info={"RKMETableSpecification": user_spec}
)
search_result = demo_market.search_learnware(user_info)

single_result = search_results.get_single_results()
multiple_result = search_results.get_multiple_results()

# search_item.score: based on MMD distances, sorted in descending order
# search_item.learnware.id: id of learnwares, sorted by scores in descending order
for search_item in single_result:
    print(f"score: {search_item.score}, learnware_id: {search_item.learnware.id}")

# mixture_item.learnwares: collection of learnwares whose combined use is beneficial
# mixture_item.score: score assigned to the combined set of learnwares in `mixture_item.learnwares`
for mixture_item in multiple_result:
    print(f"mixture_score: {mixture_item.score}\n")
    mixture_id = " ".join([learnware.id for learnware in mixture_item.learnwares])
    print(f"mixture_learnware: {mixture_id}\n")

Reuse Learnwares

We offer two data-free methods JobSelectorReuser and AveragingReuser for reusing a given list of learnwares. Please substitute test_x in the code snippet below with your own testing data:

from learnware.reuse import JobSelectorReuser, AveragingReuser

# using job selector reuser to reuse the searched learnwares to make prediction
reuse_job_selector = JobSelectorReuser(learnware_list=mixture_item.learnwares)
job_selector_predict_y = reuse_job_selector.predict(user_data=test_x)

# using averaging ensemble reuser to reuse the searched learnwares to make prediction
reuse_ensemble = AveragingReuser(learnware_list=mixture_item.learnwares)
ensemble_predict_y = reuse_ensemble.predict(user_data=test_x)

We also provide two data-dependent methods: EnsemblePruningReuser and FeatureAugmentReuser, when the user has minor labeled data for refining a given list of learnwares. Here's an example for adopting multiple returned learnwares by labeled data to solve classification tasks:

from learnware.reuse import EnsemblePruningReuser, FeatureAugmentReuser

# Use ensemble pruning reuser to reuse the searched learnwares to make prediction
# (train_x, train_y) is the small amount of labeled data
# `mode` has two options "classification" and "regression"
reuse_ensemble = EnsemblePruningReuser(learnware_list=mixture_item.learnwares, mode="classification")
reuse_ensemble.fit(train_x, train_y)
ensemble_pruning_predict_y = reuse_ensemble.predict(user_data=test_x)

# Use feature augment reuser to reuse the searched learnwares to make prediction
reuse_feature_augment = FeatureAugmentReuser(learnware_list=mixture_item.learnwares, mode="classification")
reuse_feature_augment.fit(train_x, train_y)
feature_augment_predict_y = reuse_feature_augment.predict(user_data=test_x)

Auto Workflow Example

The learnware package also offers automated workflow examples. This includes preparing learnwares, uploading and deleting learnwares from the market, and searching for learnwares using both semantic and statistical specifications. To experience the basic workflow of the learnware package, the users can run test/test_workflow/test_workflow.py to try the basic workflow of learnware.

Experiments and Examples

We build various types of experimental scenarios and conduct extensive empirical study to evaluate the baseline algorithms for specification generation, learnware identification, and reuse on tabular, image, and text data.

Environment

For all experiments, we used a single Linux server. Details on the specifications are listed in the table below. All processors were used for training and evaluating.

System	GPU	CPU
Ubuntu 20.04.4 LTS	Nvidia Tesla V100S	Intel(R) Xeon(R) Gold 6240R

Tabular Scenario Experiments

On various tabular datasets, we initially evaluate the performance of identifying and reusing learnwares from the learnware market that share the same feature space as the user's tasks. Additionally, since tabular tasks often come from heterogeneous feature spaces, we also assess the identification and reuse of learnwares from different feature spaces.

Settings

Our study utilize three public datasets in the field of sales forecasting: Predict Future Sales (PFS), M5 Forecasting (M5), and Corporacion. To enrich the data, we apply diverse feature engineering methods to these datasets. Then we divide each dataset by store and further split the data for each store into training and test sets. A LightGBM is trained on each Corporacion and PFS training set, while the test sets and M5 datasets are reversed to construct user tasks. This results in an experimental market consisting of 265 learnwares, encompassing five types of feature spaces and two types of label spaces. All these learnwares have been uploaded to the Beimingwu system.

Baseline algorithms

The most basic way to reuse a learnware is Top-1 reuser, which directly uses the single learnware chosen by RKME specification. Besides, we implement two data-free reusers and two data-dependent reusers that works on single or multiple helpful learnwares identified from the market. When users have no labeled data, JobSelector reuser selects different learnwares for different samples by training a job selector classifier; AverageEnsemble reuser uses an ensemble method to make predictions. In cases where users possess both test data and limited labeled training data, EnsemblePruning reuser selectively ensembles a subset of learnwares to choose the ones that are most suitable for the user’s task; FeatureAugment reuser regards each received learnware as a feature augmentor, taking its output as a new feature and then builds a simple model on the augmented feature set. JobSelector and FeatureAugment are only effective for tabular data, while others are also useful for text and image data.

Homogeneous Cases

In the homogeneous cases, the 53 stores within the PFS dataset function as 53 individual users. Each store utilizes its own test data as user data and applies the same feature engineering approach used in the learnware market. These users could subsequently search for homogeneous learnwares within the market that possessed the same feature spaces as their tasks.

We conduct a comparison among different baseline algorithms when the users have no labeled data or limited amounts of labeled data. The average losses over all users are illustrated in the table below. It shows that unlabeled methods are much better than random choosing and deploying one learnware from the market.

Setting	MSE
Mean in Market (Single)	0.897
Best in Market (Single)	0.756
Top-1 Reuse (Single)	0.830
Job Selector Reuse (Multiple)	0.848
Average Ensemble Reuse (Multiple)	0.816

The figure below showcases the results for different amounts of labeled data provided by the user; for each user, we conducted multiple experiments repeatedly and calculated the mean and standard deviation of the losses; the average losses over all users are illustrated in the figure. It illustrates that when users have limited training data, identifying and reusing single or multiple learnwares yields superior performance compared to user's self-trained models.

Heterogeneous Cases

Based on the similarity of tasks between the market's learnwares and the users, the heterogeneous cases can be further categorized into different feature engineering and different task scenarios.

Different Feature Engineering Scenarios

We consider the 41 stores within the PFS dataset as users, generating their user data using a unique feature engineering approach that differ from the methods employed by the learnwares in the market. As a result, while some learnwares in the market are also designed for the PFS dataset, the feature spaces do not align exactly.

In this experimental setup, we examine various data-free reusers. The results in the following table indicate that even when users lack labeled data, the market exhibits strong performance, particularly with the AverageEnsemble method that reuses multiple learnwares.

Setting	MSE
Mean in Market (Single)	1.149
Best in Market (Single)	1.038
Top-1 Reuse (Single)	1.075
Average Ensemble Reuse (Multiple)	1.064

Different Task Scenarios

We employ three distinct feature engineering methods on all the ten stores from the M5 dataset, resulting in a total of 30 users. Although the overall task of sales forecasting aligns with the tasks addressed by the learnwares in the market, there are no learnwares specifically designed to satisfy the M5 sales forecasting requirements.

In the following figure, we present the loss curves for the user's self-trained model and several learnware reuse methods. It is evident that heterogeneous learnwares prove beneficial with a limited amount of the user's labeled data, facilitating better alignment with the user's specific task.

Image Scenario Experiment

Second, we assess our algorithms on image datasets. It is worth noting that images of different sizes could be standardized through resizing, eliminating the need to consider heterogeneous feature cases.

Settings

We choose the famous image classification dataset CIFAR-10, which consists of 60000 32x32 color images in 10 classes. A total of 50 learnwares are uploaded: each learnware contains a convolutional neural network trained on an unbalanced subset that includs 12000 samples from four categories with a sampling ratio of 0.4:0.4:0.1:0.1. A total of 100 user tasks are tested and each user task consists of 3000 samples of CIFAR-10 with six categories with a sampling ratio of 0.3:0.3:0.1:0.1:0.1:0.1.

Results

We assess the average performance of various methods using 1 - Accuracy as the loss metric. The following table and figure show that when users face a scarcity of labeled data or possess only a limited amount of it (less than 2000 instances), leveraging the learnware market can yield good performances.

Setting	Accuracy
Mean in Market (Single)	0.655
Best in Market (Single)	0.304
Top-1 Reuse (Single)	0.406
Job Selector Reuse (Multiple)	0.406
Average Ensemble Reuse (Multiple)	0.310

Text Scenario Experiment

Finally, we evaluate our algorithms on text datasets. Text data naturally exhibit feature heterogeneity, but this issue can be addressed by applying a sentence embedding extractor.

Settings

We conduct experiments on the well-known text classification dataset: 20-newsgroup, which consists approximately 20000 newsgroup documents partitioned across 20 different newsgroups. Similar to the image experiments, a total of 50 learnwares are uploaded. Each learnware is trained on a subset that includes only half of the samples from three superclasses and the model in it is a tf-idf feature extractor combined with a naive Bayes classifier. We define 10 user tasks, and each of them encompasses two superclasses.

Results

The results are depicted in the following table and figure. Similarly, even when no labeled data is provided, the performance achieved through learnware identification and reuse can match that of the best learnware in the market. Additionally, utilizing the learnware market allows for a reduction of approximately 2000 samples compared to training models from scratch.

Setting	Accuracy
Mean in Market (Single)	0.507
Best in Market (Single)	0.859
Top-1 Reuse (Single)	0.846
Job Selector Reuse (Multiple)	0.845
Average Ensemble Reuse (Multiple)	0.862

LLM Experimental Results (New)

This section refers to Section 4 of our paper Learnware of Language Models: Specialized Small Language Models Can Do Big. We simulate a learnware system comprising approximately 100 learnwares of specialized SLMs with 8B parameters, fine-tuned across finance, healthcare, and mathematics domains.

Experimental results demonstrate promising performance: by selecting one suitable learnware for each task-specific inference, the system outperforms the base SLMs on all benchmarks. Compared to LLMs, the system outperforms Qwen1.5-110B, Qwen2.5-72B, and Llama3.1-70B-Instruct by at least 14% in finance domain tasks. Additionally, it surpasses Flan-PaLM-540B (ranked 7th on the Open Medical LLM Leaderboard) in medical domain tasks.

The figure and table below show the performance value in finance scenario.

User	Qwen2.5-7B	Llama3.1-8B-Instruct	Llama3.1-8B	Qwen1.5-110B	Qwen2.5-72B	Llama3.1-70B-Instruct	Random	Learnware	Best-single	Oracle
australian	43.17	44.6	43.17	43.17	43.17	47.48	44.45	56.83	42.21	56.83
cra_lendingclub	80.82	76.33	57.34	80.82	47.01	53.07	81.52	92.07	80.82	92.07
fiqasa	38.3	40.43	56.17	63.4	64.26	68.51	46.53	76.38	32.06	76.38
fpb	76.08	32.78	30.72	70.72	78.35	78.04	67.95	84.25	77.73	84.25
german	65.0	49.5	66.0	66.0	66.5	43.5	51.5	67.06	65.33	67.06
headlines	74.81	59.95	59.95	62.96	77.84	77.53	72.43	95.61	95.61	95.61
ner	21.75	0.62	9.01	17.89	9.36	9.52	24.99	52.79	23.98	52.79
sm_acl	51.1	51.4	51.34	49.3	51.56	49.38	51.42	52.82	50.71	53.63
sm_bigdata	55.3	55.57	52.79	51.02	50.27	47.76	53.86	52.4	55.52	55.88
sm_cikm	58.44	54.24	54.07	44.01	58.27	47.86	55.89	55.99	57.98	58.52
causal20_sc	65.14	88.48	79.45	83.75	76.17	87.16	74.71	84.17	88.61	88.61
finarg_ecc_arc	64.78	46.67	60.0	62.32	63.04	44.64	62.27	64.31	57.87	68.36
finarg_ecc_auc	48.3	51.81	49.85	55.01	61.71	65.02	52.08	58.08	48.68	58.08
fomc	60.48	29.44	34.68	58.47	57.66	66.13	56.05	62.7	61.36	62.7
ma	79.2	56.4	51.0	81.4	84.6	83.2	73.64	79.81	79.27	79.81
mlesg	35.67	32.67	20.0	34.67	38.67	42.33	31.99	33.42	38.33	38.33
multifin_en	60.99	31.32	28.39	65.38	63.55	68.5	54.96	63.46	58.61	63.46
Avg.	57.61	47.19	47.29	58.25	58.35	57.63	56.25	66.6	59.69	67.79
Avg. rank	5.94	7.35	7.82	5.94	4.71	5.24	6.47	2.88	5.47	1.65
Learnware (win/tie/loss)	13/0/4	15/0/2	16/0/1	14/0/3	12/0/5	11/0/6	16/0/1	nan	12/1/4	0/11/6
Oracle (win/tie/loss)	17/0/0	17/0/0	17/0/0	15/0/2	13/0/4	12/0/5	17/0/0	6/11/0	14/3/0	nan

Our system demonstrates strong performance across financial tasks, achieving the highest average score among all methods, delivering an nearly 14% improvement compared with the best large-scale model Qwen2.5-72B. It ranks first strategies utilizing specialized SLMs except Oracle in 13 out of 17 tasks, identifies the optimal learnware (tied with Oracle) on 11 and outperforms all contenders in 8.

These results shows that our system can match or surpass large-scale models with over 70B parameters under the Task-Level evaluation setting, while requiring only the memory for models under 8B efficiently.

For more scenarios (medical and math) and details, please see here.

Citation

If you use our project in your research or work, we kindly request that you cite the following papers:

@article{zhou2024learnware,
  title = {Learnware: Small models do big},
  author = {Zhou, Zhi-Hua and Tan, Zhi-Hao},
  journal = {Science China Information Sciences},
  volume = {67},
  number = {1},
  pages = {112102},
  year = {2024}
}

@article{tan2024beimingwu,
  title = {Beimingwu: A learnware dock system}, 
  author = {Tan, Zhi-Hao and Liu, Jian-Dong and Bi, Xiao-Dong and Tan, Peng and Zheng, Qin-Cheng and Liu, Hai-Tian and Xie, Yi and Zou, Xiao-Chuan and Yu, Yang and Zhou, Zhi-Hua},
  journal = {arXiv preprint arXiv:2401.14427},
  year = {2024}
}

About

How to Contribute

learnware is still in its early stages and may contain bugs and issues. We highly value and encourage contributions from the community, and we sincerely invite everyone to research and improve learnware algorithms, develop and explore within the learnware paradigm. For detailed development guidelines, please consult our Developer Guide. We kindly request that contributors adhere to the provided commit format and pre-commit configuration when participating in the project. Your valuable contributions are greatly appreciated.

About Us

The learnware repository is developed and maintained by the LAMDA Beimingwu R&D (Research and Development) Team. To learn more about our team, please visit the Team Overview.