AlphaTree: A Curated Roadmap of Deep Learning Model History with Linked Source Code

The Gap AlphaTree Fills: Model Lineage Without a Paper Trail

Deep learning papers describe similar architectures with different notation, different diagrams and different emphasis on what improved. The README states this directly: researchers write in different styles, and similar models are drawn differently depending on which improvement the authors want to highlight. For an application engineer who has to touch several AI subfields rather than specialize in one, that fragmentation makes it hard to see how LeNet leads to AlexNet, how AlexNet leads to VGG and GoogLeNet, and how those lead to ResNet. AlphaTree's stated goal is to publish one article per model, pair it with code, and produce a diagram, so that the progression is visible in one place. The repository also names a second audience: interview candidates who can derive formulas on a whiteboard but cannot explain engineering details. The README describes interviewing people who claim they want to move into deep learning yet freeze when the conversation moves past equations. AlphaTree is aimed at that gap between textbook derivation and applied understanding.

What Is Actually in the Repository: Diagrams, Tables and Branches

The visible structure is a set of Chinese-named directories, with the main one being DNN深度学习神经网络 (DNN deep neural network). Inside it, model pages are organized by task, starting with object classification 物体分类, and the README's index table links to individual pages for LeNet, AlexNet, GoogLeNet, Inception V3, VGG, ResNet and ResNeXt, Inception-Resnet-V2 and others. Each model page is expected to combine prose, a diagram and a link to source code. The README also embeds a comparison table across AlexNet, ZFNet, VGG, GoogLeNet and ResNet that lists year, layer count, Top-5 error, data augmentation, inception modules, convolution layer count, kernel sizes, fully connected layer sizes, dropout, local response normalization and batch normalization. That table is the clearest single artifact in the material: it lets you compare architectural decisions side by side rather than reading five papers. A separate image, allmodel.png, is described as an overview of all models. The README also contains a section on ILSVRC competition results from 2016 and 2017, naming the winning teams and their error rates.

How You Use It: Clone, Open, Read

There is no installation step, no package to build and no configuration file. The material describes a documentation repository, so the workflow is to clone it and read it in a browser or editor. The command is git clone https://github.com/weslynn/graphic-deep-neural-network.git, followed by opening the DNN深度学习神经网络 directory. The README's model index links point to files such as DNN深度学习神经网络/object classification 物体分类/LeNet.md and DNN深度学习神经网络/object classification 物体分类/ResNet.md. Note that some links in the README use the older repository name weslynn/graphic-deep-neural-network while others use weslynn/AlphaTree-graphic-deep-neural-network, so a link that 404s in the browser may still exist under the other path. The code that accompanies the model pages is described as Python and PyTorch, with TensorFlow and Caffe official source code attached to the deep learning paper sections. The README does not give a single command to run any of that code, and no requirements.txt or environment file is visible in the supplied material. Treat the code as reference reading material rather than a runnable project.

The Comparison Table Is the Most Useful Artifact, and Also the Most Dated

The architecture comparison table is worth examining on its own. It shows AlexNet at 8 layers and 16.4 percent Top-5 error, ZFNet at 8 layers and 11.2 percent, VGG at 19 layers and 7.3 percent, GoogLeNet at 22 layers and 6.7 percent, and ResNet at 152 layers and 3.57 percent. It also marks which models introduced inception modules, which used local response normalization, and which used batch normalization. That last column is a compact illustration of a real shift: ResNet is the only model in the table marked with batch normalization, and it is also the only one without local response normalization. For someone preparing for an interview, that table is a better starting point than five separate paper abstracts. The limitation is that the table stops at ResNet. Everything after it (DenseNet, MobileNet, EfficientNet, transformer-based vision models) is outside this table, and the README mentions DenseNet and DPN only in passing prose. If your work involves architectures from the last several years, this table will not cover them.

Where AlphaTree Breaks Down: Language, Scope and Maintenance Signals

The documentation is written in Chinese. The README, the directory names and the model pages are all in Chinese, with English model names embedded. An engineer who cannot read Chinese will get the diagrams, the comparison table and the linked source code, but not the explanatory prose, which is the part that connects a model to its motivation. The second limitation is scope drift. The README's top section is largely about AI千集 (aiqianji.com), a community site, a PPT generation service, a Feishu document link with a password, an AI video product price comparison table, and a link to a separate project called awesome-human-skill. None of that is deep learning curriculum. The reader has to scroll past promotional and unrelated material to reach the model index. The third issue is maintenance. The repository was last pushed on 2026-05-11 according to the metadata, but the content describes ILSVRC 2017 as the final ImageNet competition and does not cover architectures after ResNet in the comparison table. The README also states that the navigation website was removed and replaced with a Feishu document, which suggests the project's supporting infrastructure has changed over time. No releases were retrieved, so there is no versioned snapshot to pin against.

Alternatives: Papers With Code and Framework Model Zoos

Papers With Code takes a different approach. It indexes papers and links each one to implementations, datasets and leaderboards, and it is updated continuously by the research community. The difference in approach is that Papers With Code is a database you search, while AlphaTree is a curated narrative you read in order. Papers With Code will tell you the current state of the art on a benchmark; AlphaTree will tell you why VGG used only 3x3 kernels and how that contrasts with GoogLeNet's mixed kernel sizes. If you want the latest model, Papers With Code is the better tool. If you want the historical thread that explains why models look the way they do, AlphaTree's diagram-and-table format is more direct. A second alternative is a framework model zoo, such as torchvision's model collection, which gives you importable model definitions with pretrained weights. That is a different job entirely: torchvision gives you something to run, AlphaTree gives you something to understand. The two are complementary, not substitutes, and AlphaTree does not claim to provide pretrained weights or a training pipeline.

Licence, Reuse and What to Verify Before You Rely on It

The README states the licence as CC-BY-NC-SA, which is Creative Commons Attribution-NonCommercial-ShareAlike. The repository metadata supplied does not list a licence, so the README statement is the only licence information available. The practical implication is that the non-commercial clause restricts using the material in a commercial training product or paid course, and the share-alike clause requires derivative works to carry the same licence. This is not legal advice; if you plan to reuse the diagrams or text in a commercial context, check the licence terms directly. The code linked from the model pages may carry different licences depending on its origin, since the README describes it as TensorFlow and Caffe official source code. Before relying on AlphaTree for interview preparation, verify three things: that the model pages you need are present under the current directory names, that the linked source code repositories still exist, and that the comparison table's numbers match the papers you are citing, since the table is a secondary source. The repository is a map, not a primary reference.

Editorial conclusion

AlphaTree suits engineers and interview candidates who want a chronological map of deep learning architectures with pointers to reference code, and who are comfortable reading Chinese-language documentation. It is the wrong tool for anyone who needs a runnable library, a maintained tutorial series or English-only material. Before committing study time, open the DNN深度学习神经网络 directory on the master branch and confirm that the model pages you care about (LeNet, AlexNet, VGG, GoogLeNet, ResNet) still resolve, since several links in the README point to a renamed repository path.

Official sources

1. Issues
2. README
3. weslynn/AlphaTree-graphic-deep-neural-network on GitHub