# Tiny Yolo ###### tags: `tinyML` ###### members: 李懿麒 ### Time: 7/12/2021 - 7/19/2021 ## 1. Successes Last Week * Succeed to run darknet and yolo-fastest. (李懿麒) * Run yolo-fastest-1.1_coco * Succeed to run darknet.(陳伯庭) ## 2. Progress and Problems Last Week * yolo-fastest 剛把dataset抓下來(陳伯庭) ## 3. Not Resolved Problems * 尚未理解cmd實際功能 (```./darknet detector test ./cfg/coco.data ./cfg/yolov3.cfg ./yolov3.weights data/dog.jpg -i 0 -thresh 0.25```) cfg是什麼? ## 4. Goals for next week * 去了解yolo model的架構,what is yolo ? What kind of layers/operations in yolo ? How does yolo work ? What are applications of yolo ? What are problems of yolo ? (李懿麒,陳伯庭) * 弄清楚如何使用 Netron (分析模型架構與weights),分析每一層weight size .... (李懿麒,陳伯庭) * 弄清楚如何把yolo fastest 轉成TFLite (李懿麒,陳伯庭) --- ### Time: 7/19/2021 - 7/26/2021 ## 1. Successes Last Week * [Yolo 整理](https://hackmd.io/-QMZCpneREK78Dc9wWR6Xw) * Succeed to train yolo/yolo-fastest by VOC dataset and test it.(陳伯庭) * 安裝yolo、整理關於底層技術的文章(陳暐誠) * 左邊是Yolo-fastest,右邊是yolov4-tiny <img src="https://i.imgur.com/zENUiXh.jpg" alt="yolo-fastest-1.1" width="100"/><img src="https://i.imgur.com/NiBW3VP.png" alt="yolov4-tiny" width="100"/> ## 2. Progress and Problems - 環境有問題,無法使用GPU、cuDNN與opencv (已解決) ![](https://i.imgur.com/qpGkytZ.png) - Yolo-fastest的模型層數比Yolov4-tiny多,但model-size卻較小 ## 3. Not Resolved Problems - compare different versions/variations of yolo. ## 4. Goals for next week - 把 yolo 官網的 publications 看過一遍、把 Mask RCNN看懂 (陳暐誠) - 整理Yolo各版本差異(李懿麒) - Yolo-v3-tiny、Yolo-v3、Yolo-fastest的差異(陳暐誠) - Yolo-v4-tiny、Yolo-v4、Yolo-fastest的差異(李懿麒) - Yolo-v5-tiny、Yolo-v5、Yolo-fastest的差異 (陳伯庭) - 拆解model結構,看哪一層是對到哪一層,省略掉的那層又是用來做甚麼 - 弄清楚如何把yolo fastest轉成TFLite (李懿麒,陳伯庭) ### Time: 7/26/2021 - 8/2/2021 ## 1. Successes Last Week - 整理Yolo各版本差異(李懿麒)[Yolo 各版本比較與整理](https://hackmd.io/iC1r9_LfQMebIWsdywuYSA) - Yolo-v5-tiny、Yolo-v5、Yolo-fastest的差異 (陳伯庭) [Yolo v5 整理](https://hackmd.io/faDFEnkzQ4yJXbeBCuZ-KA?both) ## 2. Progress and Problems - 把 yolo 官網的 publications 看過一遍、把 Mask RCNN看懂 (陳暐誠) - Yolo-v3-tiny、Yolo-v3、Yolo-fastest的差異(陳暐誠) - Yolo-v4-tiny、Yolo-v4、Yolo-fastest的差異(李懿麒) - 拆解model結構,看哪一層是對到哪一層,省略掉的那層又是用來做甚麼 - 弄清楚如何把yolo fastest轉成TFLite (李懿麒,陳伯庭) ## 3. Not Resolved Problems ## 4. Goals for next week 1. 確定各版本模型大小速度準度差異與各層用途(李懿麒) 2. 了解yolo-v5s跟yolo-fastest內層、大小、準確度差異 (陳伯庭) ### Time: 8/2/2021 - 8/9/2021 ## 1. Successes Last Week - Yolo-v3-tiny、Yolo-v3、Yolo-fastest的差異 (陳暐誠) [Yolov3 整理](/rK7HwUSgS4mc-Cavy4i8nw) ## 2. Progress and Problems 1. 確定各版本模型大小速度準度差異與各層用途(李懿麒) 2. 確定yolo-v5s跟yolo-fastest內層、大小、準確度差異 (陳伯庭) ## 3. Not Resolved Problems ## 4. Goals for next week - 嘗試砍層、減少深度 (拿掉一些不會影響精確度的層) - 找小的model (ResNet, DenseNet, EfficientNet, ...) 看能否拼出一個更好的 ### Time: 8/9/2021 - 8/16/2021 ## 1. Successes Last Week - EfficientNet 模型結構整理 (陳暐誠) [EfficientNet 模型結構](/aetm8iAPSZOHolr2YgSRnA) - [紀錄](https://hackmd.io/SNG0H3niQLaWbOZkfhW-WA?both)(陳伯庭) ## 2. Progress and Problems - EfficientNet 模型結構、TFlite api... (陳暐誠) - 看還有什麼小架構的模型 (像: EfficientNet) (李懿麒) - 看 Yolo-fastest 的結構以及它跟EfficientNet之間的關聯。如:為何使用、優缺點、如何使用...等。 (陳伯庭) ## 3. Not Resolved Problems ## 4. Goals for next week 尋找改造方向、合適的Dataset、完整的Train一遍 ### Time: 8/16/2021 - 8/23/2021 ## 1. Successes Last Week success in training a model by a [dataset](https://public.roboflow.com/object-detection/plantdoc)(陳伯庭) ## 2. Progress and Problems ## 3. Not Resolved Problems ## 4. Goals for next week 多丟幾個dataset進去EfficientNet-lite train,嘗試抽走其中的幾層並記錄model size與Accuracy。 ### Time: 8/23/2021 - 8/30/2021 ## 1. Successes Last Week ## 2. Progress and Problems ## 3. Not Resolved Problems ## 4. Goals for next week