contributed by < weihsinyeh > 資訊所 葉惟欣 學號: R13922043 Problem 1: Zero-shot Image Captioning with LLaVA Zero-shot: You cannot do any finetuning for the pretrained model. In this problem, you only need to evaluate the pretrained LLaVA on image captioning task. Input: (1)image (2)language instruction (3)generation config Output: caption Please use the model “llava-hf/llava-1.5-7b-hf” in the transformers package

contributed by < weihsinyeh > 資訊所 113 姓名：葉惟欣 學號：R13922043 作業說明 Data please download dataset from the link below https://drive.google.com/file/d/1B2rY6cfCD-UbxBGM24sgxk5NWDgiXg8_/view?usp=drive_link

source code 請參考附件，展示類似 Cairo APIs [1] 風格的向量繪圖函式庫，後續的改進: 不依賴 libm，完全用 fixed point 建構 整合進 twin: https://github.com/Joouis/EVGUI_Viewer 以 perf 一類的工具找出效能瓶頸，並提出改進方案 在 Linux framebuffer 展現 [1] https://www.cairographics.org/manual/

contributer <weihsinyeh> 姓名 : 葉惟欣 學號 : R13922043 系級 : 資工所 113 作業要求投影片 Problem 1: Conditional Diffusion Models (30%) [digit dataset - MNIST-M & SVHN] evaluation (15%) epoch

contributer <weihsinyeh> 姓名 : 葉惟欣 學號 : R13922043 系級 : 資工所 113 Problem 1: Self-supervised pre-training for image classification (50%) (5%) Describe the implementation details of your SSL method for pre-training the ResNet50 backbone. (including but not limited to the name of the SSL method & data augmentation techniques you used, learning rate schedule, optimizer, and batch size setting for this pre-training phase) SSL method & data augmentation techniques

contributed by < weihsinyeh > 實驗環境 $ gcc --version gcc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 Copyright (C) 2019 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. $ lscpu Architecture: x86_64

contributed by < weihsinyeh > 從前 4 週的測驗題選出 3 題改進 2024q1 Homework4 (quiz3+4) 第四周 測驗 1 unsigned popcount(unsigned v) { int n; n = (v >> 1) & 0x77777777;

contributed by < weihsinyeh > 第一週測驗題 測驗一 參考 Optimized QuickSort — C Implementation (Non-Recursive)以實作作者 Darel Rex Finley 想法並提出改進。 不用遞迴方式原因是，遞迴執行效率比直接使用真正的陣列慢。同時因為遞迴其實用 stack 作為 their private array，可能會 stack overflow 導致無預警的 system crashing。為了避免無預警的狀況發生，因此改設定 MAX_LEVELS 為 1000 的陣列如果到達上限就 return NO 。 Wikipedia: C language implementation of QuickSort 上的 R 每次都只移動一個，而 L 則一次移到比 pivot 大的值上面然後將與 R 交換所以此時 R 有可能比 pivot 大但是也被交換到排序較小的地方。當 L 比 R 大的時候，再將 pivot 放到 L-1 的地方。

contributed by < weihsinyeh > 開發環境 $ gcc --version gcc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 Copyright (C) 2019 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. $ lscpu Architecture: x86_64

contributed by < weihsinyeh > 第三週測驗 測驗 1 改成版本二的原因是為了不使用 libm 的 log2，因為log 2 會導致錯誤。 Domain error: x is negative errno is set to EDOM. An invalid floating-point exception (FE_INVALID) is raised. Pole error: x is zero errno is set to ERANGE. A divide-by-zero floating-point exception (FE_DIVBYZERO) is raised. 但這裡個錯誤如果設定好 N 不會發生。接下來是精度問題。