1. Studied AE, VAE, GAN, VQGAN, DragGan, GLIDE, DALLE2 MidJourney, Imagen and other models, and completed the first version of Stable Diffusion deployment using pytorch+ldm. On the 3060Ti, an image of size 512*512 can be generated in 2 seconds. 20 avatars of each user were collected, segmented using SAM, then replaced with backgrounds, and annotated through DeepBooru. The LoRA model is trained and can generate a variety of anime-style avatars.2. Completed the controlnet training of the fill50k data set, and used pytorch+diffuser to implement Stable Diffusion, which supports the image generation process of loading lora and controlnet, and can color animation line drawings. The principles of Locon, Loha, Lokr, Dylora and IA3 were also studied and their performances were compared.3. The background of the project is to detect the number of animals on the farm and conduct distributed node consensus algorithm recording to achieve transaction traceability. Using Pytorch's YOLOV8 for target detection, deployed through TensorRT, target detection was performed within 4 milliseconds of inference time, and a map value of 0.87 was obtained. In the OCR task, Text Renderer was used to expand the smart meter data set, and DBNet and CRNN were used to train the PP-OCRv3 model, and then pdmodel and pdiparams were used for inference, achieving an accuracy of up to 0.99.4. In the Hackathon competition, he led the team to develop WEB3.0 SMART INTERIOR DESIGN, which generated a 3D UE model from 2D architectural design drawings and deployed it on the company's internalvenachain, allowing the creators of blocks earlier in the timeline to Be able to enjoy the latter part of thebenefits.

1. Use ESRGAN to super-resolve 256*256 images to 1024*1024 large images to reduce amplification losses.2. Tuning the CycleGAN algorithm to complete model generation in an artistic painting style.3. Use the CLIP model to detect the correspondence between the video cover and ccompanying text, and provide modification suggestions for those that do not match.4. Use Swin-Transformer technology to classify videos to improve the event accuracy of videos in Douyin app.

1. In the video frame interpolation project, AdaCoF, CAIN, EDSC and RIFE models were compared and trained, and the PSNR value of the RIFE model was 28.2. In the image super-resolution project, the training of the Real-CUGAN image super-resolution model has been completed, and its PSNR value is 30.3. In the smart photo album project, through the face recognition algorithm, realize face special effects, such as beautification, stickers, etc., to increase the user's interactive experience.4. In the model conversion project, completed the PTQ INT8 quantization conversion of YOLO, RVM, NanoDet and Paddle-OCR models to MNN models.5. In the highlights project, based on ShuffleNet and OCR algorithm, identify the key frames in the LOL game, then generate the title, and use ffmpeg to synthesize a new video.

 Calculate the EAR by using the Dlib framework, and complete the fatigue driving algorithm based on the number of blink detections. Extract texture features through LBP and combine with SVM classifier to complete the cargo detection algorithm in the vehicle. Passenger flow prediction algorithm based on LSTM and ARIMA. In the Tianqing Phase I project in cooperation with Gaode, a variety of algorithms (including Mercator projection, computational geometry, polynomial fitting interpolation, genetic algorithm, A* algorithm, etc.) were integrated to complete the bus line network optimization algorithm: including bus line modification ,bus schedule calculation. The project involves a complex network scale of more than 300 lines and 4000 stations. This project is one of Tianmai's largest revenue businesses in 2020, with sales of more than 2 million. In this project, self-developed algorithms from scratch, including pandas, sklearn, numpy, gatb, neo4j, scipy, networkx, datetime, flask, pykafka, bokeh, seaborn, etc. In the Shenzhen Charging Scheduling Project, with the lowest charging cost and the best battery life as the optimization objectives and limiting conditions, the differential evolution algorithm is used to calculate the best combination of charging time periods. In the taxi supervision project, the convex hull algorithm, DBSCAN and A* search algorithm were used to complete the taxi aggregation detection and hotspot recommendation system. In the smart charging station project, the ant colony algorithm was used to realize the route planning algorithm of AGV, and a patent was applied for. In the smart charging station project, animation effects were realized through Unity, including scene switching, subtitles, Trigger and Colliders.

 Using MATLAB to realize the analysis of the abnormal behavior of electricity consumption, using the GMM model. In SIMULINK, the configuration of the charging station energy storage system and the simulation of the photovoltaic grid connection process are carried out. In the smart grid project, the degree distribution, centrality, connectivity, network elasticity, small-world properties, clustering coefficient and community discovery evaluation indicators of the power grid are studied based on complex network theory.