"Dive into Deep Learning (D2L)" is an open source, interactive deep learning textbook that combines theoretical explanations, mathematical derivation and runnable code. It is available in Chinese and English versions and is used as a teaching material by a large number of universities around the world. It is suitable for learners and teachers who are getting started with deep learning in the system.

Hands-on learning of deep learning

Core parameters and statistics

"Deep Learning by Hands" (D2L) is an open source, interactive deep learning textbook. The core feature is "theory + mathematics + runnable code" in parallel - readers can not only understand the concepts, but also directly run the code verification in the book.

Projects	Public Information
Positioning	Open source interactive deep learning textbooks and courses
Content form	Text explanation + mathematical derivation + runnable code
Language version	Available in Chinese, English and other languages
Framework support	Multiple deep learning framework implementation
Topics covered	Neural network CNN, RNN, attention, etc.
How to use	Read online, the code can be run
Price	Free and Open Source

Positioning Interpretation: Many deep learning textbooks are theoretical, making it difficult for readers to translate concepts into code. The value of D2L is that each concept is supported by a runnable implementation, and learners can read, modify and run at the same time, forming a "understanding-implementation-verification" concept.

Boundary Note: It is a learning material rather than a tool product, and its value depends on whether learners can actually run and practice it.

User and market recognition

D2L’s recognition comes from its open source nature and widespread adoption by universities.

Education Adoption: As an open source textbook, it is used as teaching or reference material by a large number of universities and courses around the world, forming a strong reputation foundation.

Multi-framework and multi-language: Provides multiple deep learning framework implementations and multi-language versions, lowering the threshold for learners and teachers from different backgrounds.

Items to be verified: Specific data such as the number of universities and readerships will be updated periodically and should be subject to official real-time disclosure. It is not appropriate to treat approximate numbers as fixed numbers.

Cost advantage

D2L’s cost structure is minimalist: completely free and open source.

• C client/individual (learner): Online reading and running code is free, which is its biggest advantage.
• Developer/Teacher: It can be used for learning, lesson preparation and secondary use for free, following its open source agreement.
• Institution/University: As there is zero licensing fee for teaching materials, the hidden costs are mainly in course organization and localization.

True Cost Structure: The explicit cost is zero, the real investment is learning time and practice (such as the computing resources required to run the code). When evaluating, you should focus on whether the content depth and framework version match your needs, not the price.

Main functions

D2L’s capabilities are organized around “hands-on systematic learning”:

• Theory and code parallel: Each concept can be implemented in a complete package.
• Systematic chapter system: unfold step by step from basic mathematics to advanced models.
• Multi-framework implementation: Different deep learning framework versions are provided for the same content.
• Multi-language version: Chinese and English versions are convenient for different readers to use.

The key to functional value is "runnability": code that can be directly run through and modified is the core that distinguishes it from purely theoretical textbooks.

Model and version evolution

D2L is maintained in the form of continuously updated online teaching materials, and the version context is mainly content revision.

• First Edition (~2021): Form a more complete chapter system and launch the official version.
• Continuous Updates (~2026): Extended framework implementation, added chapters, and revised content.

It does not have a release version number of traditional software. The content is based on the online version. New topics (such as attention mechanism Transformer, etc.) are continuously added as the field develops.

Technical advantages

As a teaching material, the advantage of D2L lies in “instructional design” rather than software technology:

Mechanism: Decompose abstract concepts into a progressive structure of "Explanation → Mathematics → Code → Exercise" and host it in a runnable notebook.

Effectiveness: Learners can immediately verify each concept, reducing the common obstacle of "understanding but not being able to write".

Applicable scenarios: Most suitable for learners who need a systematic and solid introduction to deep learning, and want to master both theory and engineering implementation.

The price is that it requires a high level of initiative on the part of the learner, and it is difficult to realize its value by just reading without practicing.

How to use

D2L is open online through the official website:

Usage	Suitable objects	Features
Online reading	Self-learner	Systematic study by chapter
Running code	Practitioners	Run book notebooks locally or in the cloud
Course materials	Teachers/Universities	As teaching or reference materials

The typical process is "read chapters by chapter → run and modify the sample code → complete exercises to consolidate". Before starting, it is recommended to configure the executable code environment so that you can learn and practice at the same time.

Product Pricing

D2L is completely free and open source, and there is no charge for online reading and code running. The only possible cost to the learner is the computing resources (local GPU or cloud computing power) required to run the code, and has nothing to do with the textbook itself.

• Individual/Learner: Free.
• Teachers/Universities: It can be used for teaching for free and follows the open source license.

Application scenarios

• System Introduction to Deep Learning: Establish theory and implementation capabilities from scratch, and the focus of verification is whether it can be run manually.
• College Course Teaching: As teaching materials or experimental materials, pay attention to the fit with the course progress.
• Engineer Skills Enhancement: Complete deep learning theory and code implementation, and the focus of verification is the matching of framework versions.

Applicable people

• Students and self-learners: Hope to systematically master the theory and practice of deep learning.
• College Teachers: Need ready-made, free, and operable teaching materials.
• Transformation Engineer: Moving from traditional development to deep learning requires a solid foundation.

Unsuitable situations are: learners who just want to quickly call ready-made models without caring about the principles, or who lack the context to run the code and the willingness to practice.

Summary and Outlook

The core value of "Hands-On Deep Learning" is to integrate theory, mathematics and runnable code, and to widely cover learners and universities in a free, open source, multi-language, and multi-framework manner. It is an introductory material with a solid reputation in the field of deep learning. Its advantages are systematicness and maneuverability.

The current limitation is that it requires high learner initiative and requires self-prepared operating conditions; the content will also be continuously revised as the field rapidly evolves. For learners, it is recommended to advance at the pace of "read one chapter, run one chapter, and practice one chapter" and keep the framework version consistent with the official online version to obtain the best learning effect.

Related tools: khanmigo, quizlet

Business process integration and ROI analysis

Hands-on learning of deep learning As a productivity tool for enterprises or professional positions, its true value depends on the depth of integration with existing workflows and the quantifiable efficiency improvement effect. The following is a systematic analysis from three core dimensions.

System integration and data interoperability The ability to interoperate with existing business systems is a key prerequisite for productivity tools to be integrated into workflows. It is recommended to focus on evaluating the following integration dimensions: the openness and documentation quality of the RESTful/GraphQL API (whether a complete API reference and SDK examples are provided), the support scope of Webhook event notifications (which business event types are supported for automatic push), the number and depth of pre-built integrations with common collaboration SaaS tools (WeChat Enterprise, DingTalk, Feishu, Slack, Notion, Jira, etc.), and enterprise-level identity authentication support (SSO/SAML/OAuth and LDAP/AD directory integration). Products that lack integration capabilities are easily isolated into information islands, which in turn increases the cognitive cost and operational friction for teams to switch between different tools.

Efficiency Quantification and ROI Estimation Methodology Before purchasing decisions, it is recommended to quantify the input-output ratio through a structured method: Step 1, choose 3-5 Standardized tasks that are frequently repeated and time-consuming in each team are used as test samples; in the second step, the average time consumption of a single task before and after tool intervention, first-time pass rate or error rate, and the number of links requiring manual intervention are recorded under controlled conditions; in the third step, the saved manpower time is converted according to the comprehensive cost of the position (salary, benefits, management sharing), and soft benefits (increased employee satisfaction, standardization of work quality, and improvement in response speed to core business) are superimposed to obtain a comprehensive ROI estimate. It is recommended to continue tracking ROI trends on a monthly basis, as the value of a tool usually increases over time as team proficiency increases and workflows are optimized.

Phase-based implementation strategy and risk control It is recommended to adopt a three-stage implementation path of "pilot verification-gradual promotion-continuous optimization". In the pilot stage (1-2 weeks), a single team or a single business scenario is selected for small-scale verification. The core goal is to verify technical feasibility and user acceptance, and establish preliminary usage specifications and success standards; in the promotion stage (2-4 weeks), after the pilot verification is passed, the coverage is gradually expanded, and standardized activation processes and training materials are developed; in the optimization stage (continuous), the workflow configuration is continuously adjusted based on actual usage data and user feedback, and more high-value application scenarios are explored. Clear quantitative key result indicators should be set at each stage to avoid blindly expanding the scope of use without data support.

Version Info

• Hands-on deep learning (current online version) ：The online teaching materials are continuously updated, covering deep learning content from basic to advanced, and provide runnable code implementations of multiple deep learning frameworks. For specific revisions, please refer to the official online version. There is currently no unified release version number. （~2026-06）
• The first version is officially released ：The textbook formed a relatively complete chapter system and launched the official version, and then continued to expand the framework implementation and new chapters. There is no official precise date yet, it is recorded according to public milestones. （~2021-01）