# Deep-Learning-Workshop @ASE (10/23-25)
Link of this page: https://bit.ly/2p4Pq1k

---
* Day-1: [[Link]](https://www.dropbox.com/s/032a7y4vkd3drqy/20191023_ase_day1.pdf?dl=0) for slides
* Day-1: [[Link]](https://www.dropbox.com/s/k678yo1ylrnm4dy/ase_20191023.tar.gz?dl=0) for notebooks and datasets
* Day-2:
* [[06-FacialKeypointDetection.ipynb]](https://www.dropbox.com/s/6qwyhgcvu9qfcdd/06-FacialKeypointDetection.ipynb?dl=0)
* [[misc.py]](https://www.dropbox.com/s/c9d0k4rpt7znyod/misc.py?dl=0)
* Day-3:
* [[07-GroupNormalization.ipynb]](https://www.dropbox.com/s/q4hdoo9q3qbesaf/07-groupnorm.ipynb?dl=0) (put it in $ROOTDIR/notebooks)
* [[mrcnn_tutorial.tar.gz]](https://www.dropbox.com/s/n2wj0m7dahzcyba/mrcnn_tutorial.tar.gz?dl=0)(unzip and put it in $ROOTDIR/)
* [[Link]](https://www.dropbox.com/s/9tb5vinpo0hsjox/20191025_ase_day3.pdf?dl=0) for slides
* [[Link]](https://www.dropbox.com/s/gzhxibd6s260cb8/mrcnn_network.pdf?dl=0) for MRCNN network topology

We'll use AIGO's [TensorFlow container](https://aigo.org.tw/ai-plus/hub/tool_detail/3).

---
* 01-GradientTapeAndLinearRegression [[html]](http://211.75.15.16:60000/ase-01-GradientTapeAndLinearRegression.html)
* 02-LayersIO [[html]](http://211.75.15.16:60000/ase-02-LayersIO.html)
* 03-InceptionAndVGG [[html]](http://211.75.15.16:60000/ase-03-InceptionAndVGG.html)
* 04-ResNetAndDenseNet [[html]](http://211.75.15.16:60000/ase-04-ResNetAndDenseNet.html)
* 06-FacialKeypointDetection [[html]](http://211.75.15.16:60000/ase-06-FacialKeypointDetection.html)
* 07-GroupNorm [[html]](http://211.75.15.16:60000/ase-07-GroupNorm.html)
* 08-FacialKeypointDetection-MultiGPU [[html]](http://211.75.15.16:60000/ase-08-FacialKeypointDetection-MultiGPU.html)

---
#### Day 1: Getting started with TensorFlow 2.0

* Record gradients with ```Gradient Tape```
* *code: linear regression*
* Model construction with ```tf.keras```
* ```Sequential``` API, ```Model``` API, ```Sub-classing``` API
* *code: VGG, ResNet , DenseNet*

#### Day 2: Multi-GPU & mixed-precision training

* ```tf.strategy``` for Multi-GPU training
* Speed-up the training with ```Automatic Mixed-Precision```
* Introduction to keypoint detection
* *code: facial landmark detection (UNet + heatmap regression)*

#### Day 3: Integrate GN into Mask RCNN

* Group Normalization (GN)
* Mask RCNN code review
* ~~```SavedModel``` to ```TensorRT``` for faster model inference~~

---
翁啟閎
Data Scientest @ HonghuTech
[e-mail]