2. Deep learning of handwriting (2): Use Generative score-based models to generate handwritten digits
As score-based models have the same origin as DDPM, although they have not been so popular, many of their ideas have been borrowed by later researchers. This blog will explain the code of score-based models in detail and guide readers step by step to generate their own MNIST.
3. Deep learning of handwriting (3) - Theoretical chapter on automatically generating lyrics using RNN recurrent neural network
The emergence of RNN has solved the poor performance of CNN in processing spatiotemporal sequences. The so-called spatio-temporal sequence can be simply understood as the inputs before and after are connected. For example, in natural language processing, there is a connection between the previous sentence and the next sentence we say; in intelligent video processing, there is a connection between the before and after of each frame. ; Radar signals, there is a connection between each signal... In traditional image processing, there is no connection between each pixel value in the image. We use RNN to make an automatic lyrics generator. This article first introduces the theoretical basis of RNN.
4. Deep learning of handwriting (4) - using RNN recurrent neural network to automatically generate Zhang Liangying's lyrics
With the theoretical foundation in the previous article , we officially started to practice. As a die-hard fan of Jane Zhang for 60 years, we used "Little Dolphin" as an experiment this time to see if we can use RNN to write our own beautiful love songs.
5. Deep learning of handwriting (5) - Pytorch+RNN automatically generates Deng Ziqi style lyrics
The previous two articles talked about the basic theory of RNN and using mxnet to build an RNN network to automatically generate lyrics. This article is time to reveal my identity as a fan of Deng Ziqi for thirty years, and use the more widely used Pytorch framework to build an RNN model to automatically generate a Deng Ziqi-style lyrics generator.
6. Deep learning of handwriting (6): Use the simplest neural network multi-layer perceptron to recognize numbers
MLP multi-layer perceptron is the first love for many people who are introduced to neural networks . This lecture starts from the basic multi-layer perceptron principle to build an MLP that can recognize handwritten digits. As we all know, deep learning is an engineering discipline. We start from the basic MLP, adjust parameters step by step, experience the fun of parameter adjustment in practice, and optimize the effect of the baseline.
7. Hand-written deep learning (7): Use GANs to generate handwritten digits
GANs were proposed at the NPIS conference in 2014, but fell silent for two years due to various reasons. It was not until DCGANs came out that the perfect combination of GAN and CNN really opened up the GANs blowout era, and suddenly became the strongest outlet. Several years have passed, and the popularity continues unabated. This lecture starts from the principle of DCGAN and proceeds step by step to use DCGAN to generate handwritten digits.
8. Hand-written deep learning (8) - using LSTM to generate hand-written English articles
The previous article in this series introduced the use of GANs to generate handwritten digits. The task of generating handwritten digits is a very simple and introductory task, because the pixels provided by the MNIST data set are very low, and the final generated effect is also very blurry. You know, high-resolution generation has always been an eternal research hotspot in deep generation problems. In this lecture, handwritten English letters are used as vector expressions. LSTM, as a very excellent RNN network, is widely used in the processing of spatiotemporal sequences. This lecture is about using LSTM to generate handwritten English articles.
9. Hand-written deep learning (9): conditional GANs among the top ten GANs
Like DCGANs, conditional GANs also have epoch-making significance in the history of GAN development. CGAN is the first to allow the generation of images with specific conditions or attributes. Just like a nuclear reactor, it is remarkable that it can be controlled for the first time.
10. Hand-to-hand deep learning (10): Use Pix2Pix GANs to implement sketch-to-image cross-modal tasks (theoretical basis)
In 2017, GANs are developing in full swing. With the predecessors such as CGANs and DCGANs at the forefront, Pix2Pix GANs turned out to have outstanding performance in a variety of multi-modal tasks, and for the first time brought the U-net structure to deep-level generation. In the model, it is very meaningful. This article uses Pix2Pix GANs to implement the cross-modal task of sketch-image.
11. Hand-to-hand deep learning (11): Use Pix2Pix GANs to implement sketch-to-image cross-modal tasks (practical exercise)
The previous article talked a lot about the basics of Pix2Pix GANs. Unexpectedly, the more I wrote, the more I wrote. In the end, I wrote nearly 40,000 words, so I split it into two articles: the theoretical basis and practical exercises are introduced separately. This article talks about how to use pix2pix GANs to complete sketch-to-image tasks, focusing on code drills.
12. Deep learning hand-in-hand (12): Basics of CycleGANs’ great style transfer theory
The previous blogs talked about pix2pix GANs, which can do point-to-point conversion and require paired data. This article introduces cycleGANs, which does not require paired data sets, can convert between domains, and can be used in many interesting applications: style transfer, object deformation, season conversion, image enhancement, color conversion, old photo restoration, day and night conversion, etc. wait. In this blog, we first learn the theoretical basis of recurrent generative adversarial networks.
13. Hand-written deep learning (13): Use CycleGAN to convert apples into oranges
The previous article introduced the theoretical basis related to CycleGAN . This time we put it into practice and used CycleGAN to turn apples into oranges. After learning, we can use the same method to turn day into night, wild horses into zebras, summer into autumn, oil paintings into photos, Van Gogh paintings into Monet paintings, etc.
14. Deep learning hand-in-hand (14): How to use the official pre-trained model for fine-tuning/transfer learning? (Take Resnet50 to extract image features as an example)
For models with deep networks and difficult to train, such as transformer and resnet, using the officially provided pre-training model and then fine-tuning it in your own data set/application scenario can also achieve great results and is also the most time-saving and labor-saving method. . This article takes training a resnet50 network to extract image features as an example to explain in detail how to use the official pre-trained model for transfer learning and how to use the fine-tuning model in actual application scenarios.
15. Hand-to-hand deep learning (15): Build your own corpus on Hugging Face
The tokenizer is one of the core components of the NLP pipeline. The core function is: convert text into data that the model can process . The model can only handle numbers, so the tokenizer needs to convert our text input into numeric data. This blog teaches you how to build your own corpus on Hugging Face.
16. Hand-written deep learning (16): Use CILP pre-training model to build image and text retrieval system/image search/keyword retrieval system
CLIP is all you need! CLIP has demonstrated strong dominance in the fields of text-to-image, image retrieval, video understanding, image editing, self-supervised learning, etc. This blog teaches you step by step how to build your own image and text retrieval system, which can improve the retrieval indicators. Kill all previous jobs in one fell swoop, demonstrating CLIP’s strong dominance!
17. Hand-to-hand deep learning (17): Use LSTM to generate text descriptions for images (Image-to-text task)
In recent years, the text-to-image task has been very popular. Large models such as Della, Imagen, GLIDE, etc. have achieved good results. However, there are relatively few studies related to the image-to-text mirroring task. This blog will guide you step by step. A simple LSTM network builds an image-to-text system.
18. Hands-on Deep Learning (18): Principles, codes, and parameter adjustment techniques of finetune fine-tuning CLIP models
In the previous blog "Hand-in-hand deep learning (16): Using CILP pre-training model to build a picture and text retrieval system/image search/keyword retrieval system" introduced how to perform image and text retrieval, image search , and keyword retrieval. CLIP is used in tasks such as retrieval. This blog focuses on explaining code and parameter adjustment techniques.
19. Deep learning step by step (19): Build Pytorch deep learning project architecture from scratch
In the industrial world, especially when building very large-scale distributed systems , TensorFlow has irreplaceable advantages. But in academia, pytorch is particularly preferred. The ease of learning and using pytorch is an important reason why it has become more and more popular in recent years. This article describes the overall architecture of a deep learning project and teaches you step by step how to build a deep learning project based on pytorch from scratch.
20. Hand-written deep learning (20): Building a sensitive word filtering system for LLM large language model
As more and more large language models such as Llama 2 and Tongyi Qianwen 7B are open sourced, developers can build their own dialogue systems, agents, etc. based on these open source models. However, due to our national conditions, developers need to make some special "security" considerations for these models to ensure that "harmful information" does not appear in interactions with users. This blog teaches you step by step how to build a sensitive word filtering system for a large language model.
21. Hand-written deep learning (21): fine-tuning Stable Diffusion with PEFT LoRA
PEFT is proudly produced by Hugging Face and is one of the most commonly used libraries for fine-tuning large models. This blog first introduces the principle of PEFT LoRA fine-tuning Stable Diffusion, then explains the code, and compiles a complete runnable script, which has been open sourced on GitHub.