# Deep Learning Workshop (@ FJU 2021)
本頁連結: https://bit.ly/32bYPUL

## 檔案下載
* [講義](https://www.dropbox.com/s/zsqa03lyp7kai8o/FJU_DL_intro_2021.pdf?dl=1)
* [範例程式 & 資料集](https://www.dropbox.com/s/udurwhwbxjeewz2/fju_workshop_20210414.tar.gz?dl=1)

## 環境
* 若為Windows環境 (無GPU)，請安裝[Anaconda](https://www.anaconda.com/products/individual)後，加裝如下套件:
* 基本套件

        pip install pandas matplotlib seaborn numpy scipy

* 機器學習套件

        pip install tensorflow==2.4.0 sklearn

* 其它套件

        pip install pillow mlxtend numexpr
        conda install -y shapely
        pip install imgaug opencv-python==4.4.0.46
        pip install jupyterlab

* 若為Linux (有GPU), 可考慮透過Docker啟動準備好的開發環境 (映像檔)。

    nvidia-docker run -v /data:/workspace \
                      -p 10003:22 \
                      -p 10002:8888 \
                      -p 10001:6006 \
                      --name tf \
                      --privileged \
                      --ipc=host \
                      moeaidb/aigo:cu11.1-dnn8.0.5-gpu-tf-cv-20.12

一些連結:
* [AIGO TensorFlow 映像檔說明](https://aigo.org.tw/zh-tw/hub/tool/2)
* [Docker使用簡介](https://aigo.org.tw/zh-tw/onlinecourse/467)
* [TensorFlow 映像檔 (NVIDIA 提供)](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow)
* [TensorFlow 映像檔 (TensorFlow 官方提供)]( https://hub.docker.com/r/tensorflow/tensorflow)

## 已學習內容
* Loss minimization equals to Likelihood maximization
* Understood Gradient Descent (GD) & Stochastic Gradient Descent (SGD)
* The meaning of Iteration, Batch size
* L1 and L2 can be used to against overfitting
* Definitions of Precision, Recall, F1
* Usage of TensorFlow Gradient Tape
* Backpropagation
* Understand how to build & train a Multi-Layer Perceptron (using Sequential API)
* History of Convolutional Neural Network
* Understand how to construct building blocks of Inception (Model API), and ResNet (subclassing API)
* BatchNorm, Dropout, Gaussian blur (data augmentation) can be used to against overfitting
* Adaptive methods (each weight now has its own learning rate)
#### 程式範例
* [[00]](
http://211.75.15.15:19201/00.html)
* [[01-GradientTapeForLinearRegression]](
http://211.75.15.15:19201/01-GradientTapeForLinearRegression.html)
* [[02-BasicLayersDenseConvMaxpooling]](
http://211.75.15.15:19201/02-BasicLayersDenseConvMaxpooling.html)
* [[04-InceptionAndVGG]](
http://211.75.15.15:19201/04-InceptionAndVGG.html)
* [[05-ResNetAndDenseNet]](
http://211.75.15.15:19201/05-ResNetAndDenseNet.html)
* [[07-FacialKeypointDetection]](
http://211.75.15.15:19201/07-FacialKeypointDetection.html)

# 補充連結
* Stanford CS231 [[Convolution layer]](https://cs231n.github.io/convolutional-networks/)
* Review Article [[Deep Learning; Nature, Yann LeCun 2015]](https://www.cs.toronto.edu/~hinton/absps/NatureDeepReview.pdf)
---
翁啟閎
[e-mail]