---
title: 'Tâm - Week 7'
tags: CoderSchool, Mariana
---

Week 7
===

## Table of Contents

[TOC]

## Monday

## Tuesday
## Wednesday
## Thursday
### Project presentation
Ly-Khoa:
- build model from scratch

Dinh-Long:
- use data augmentation: rotate, zoom, horizontal flip
- VGG16 pretrained on imagenet

Nhan-Sean:
- SGD, RMSProp, Adam
- Adam with MobileNetV2 works best 94%

Tuc-Felix:
- ImageDataGenerator()
- EarlyStopping
-
We will try to apply SQL and Base64

### Mechanics of Tensorflow
- Tensorflow converts protocols into graphs (which can be run on GPU and other devices or distributed systems)
- Tensorflow 2. is more pythonic
- AUTOTUNE helps find the best parameters for buffer_size
- `@tf.function` converts python code into tensorflow graph, can run on GPU
- DAG Directed Acyclic Graph
## Friday

### Dog-Cat classifier walkthru
#### preprocess and prepare train-test

#### evaluate and learn about the model
- List images with highest probability correct predictions/wrong predictions
- Using **TensorBoard** can see metrics of model. can see computational graphs of model.
#### build custom model
- We can use tf.keras layers to build a model
- activation functions
- GlorotNormal and GlorotUniform are best parameter initializer
- Optimizers update parameters. Adam optimizer uses moving average of dW to update parameters
- losses
- metrics
- applications: pretrained models

#### google image downloader
- require chromedriver and selenium to download more than 100 images
- I correctly predict for rose (88%) and sunflower (92%) https://colab.research.google.com/drive/1RkoWiQzPSkp_I_k_Gti6pz6fTJFJw_37

:::info
**Find this document incomplete?** Leave a comment!
:::

###### tags: `CoderSchool` `Mariana` `MachineLearning`