心有所向，日复一日，必有精进

摘要： Pre title: Replacing softmax with ReLU in Vision Transformers accepted: Arxiv 2023 paper: https://export.arxiv.org/abs/2309.08586 code: None 关键词：atten 阅读全文

摘要： tl;dr: pytorch的 torch.optim.lr_scheduler.OneCycleLR 就很不错，能兼顾warmup和余弦学习率，也不用下载额外的包 import torch from torch.optim.lr_scheduler import CosineAnnealingLR 阅读全文

摘要： Pre title: Randomized Quantization: A Generic Augmentation for Data Agnostic Self-supervised Learning accepted: ICCV 2023 paper: https://arxiv.org/abs 阅读全文

摘要： Pre title: R-Drop: Regularized Dropout for Neural Networks accepted: NeurIPS 2021 paper: https://arxiv.org/abs/2106.14448 code: https://github.com/dro 阅读全文

摘要： 1. Pre title: Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference accepted: arXiv 2023 (ICLR 2024 Submission) paper 阅读全文

摘要： Pre title: EMO: Earth Mover Distance Optimization for Auto-Regressive Language Modeling accepted: arXiv2023 paper: https://arxiv.org/abs/2310.04691 co 阅读全文

摘要： Pre title: Painterly Image Harmonization using Diffusion Model accepted: AAAI2023 paper: https://arxiv.org/abs/2212.08846 code: https://github.com/bcm 阅读全文

摘要： Pre title: Disentangling Writer and Character Styles for Handwriting Generation accepted: CVPR2023 paper: https://arxiv.org/abs/2303.14736 code: https 阅读全文

摘要： Pre title: SimCSE: Simple Contrastive Learning of Sentence Embeddings accepted: EMNLP 2021 paper: https://arxiv.org/abs/2104.08821 code: https://githu 阅读全文

摘要： prologue title: [pytorch] 训练时冻结一部分模型的参数 —— module.requires_grad_(False) 代码用到一个解码器\(dec\)，希望用它预测生成结果\(g\)的counting encode并用以计算损失，以此约束生成器生成合理的结果（能解码出正确的 阅读全文