机器学习纳米学位毕业项目(Capstone project: MLND Udacity)

侦测走神司机(detect distracted drivers)

FloydHub

本项目主要在FloydHub的云计算服务器上完成。服务器配置Nvidia Tesla K80 GPU。环境为tensorflow。相对于 AWS 的 p2.xlarge 服务器，FloydHub 的 CPU 速度较慢，但 GPU 速度和内存应该比 p2.xlarge 大。相同的文本和代码情况下，FloydHub上batch_size=32毫无压力，而 p2.xlarge 只能 batch_size=8。

This project uses FloydHub cloud computational servers. The server is equipped with Nvidia Tesla K80 GPU. Environment: tensorflow. Comparing with AWS p2.xlarge instance, FloydHub's CPU seems slower, but GPU is faster and has more memories. With the same code, FloydHub can easily train models with batch_size = 32, while p2.xlarge can only do batch_size = 8.

原始图片数据通过终端上传至服务器，名称为 mingyi/datasets/drivers_original/2。原始计算文件保存于项目mingyi/projects/keras_test中。在Keras文件夹下进行floyd init命令后，用以下命令载入数据和文本

The original image dataset is uploaded to the server, with name mingyi/datasets/drivers_original/2. The raw notebooks of the projects are saved at mingyi/projects/keras_test. Under the directory Keras run floyd init and use the following command to run notebooks and load dataset.

floyd run --data mingyi/datasets/drivers_original/2:dataset_dir --mode jupyter --gpu --env tensorflow

库 (Modules)

主项目 (main project)：

Keras 2.0.6, TensorFlow backend
sklearn, numpy, pandas
matplotlib.pyplot, cv2, skimage.io
tqdm

基准模型 (baseline model)：TensorFlow-Slim image classification model library

系统 (System)

本地 (Local)：Windows 10, python 3.6.2

运行时间 (Operational hours)

训练每一个fold大约两个半小时。总共10-fold， 25小时。

Every fold: 2h30. 10 folds: 25h in total.

KNN计算大约8小时。

KNN: aroubd 8h.

文件目录(Content)

Keras: 主项目文件夹 (main project directory)

5-fold cross-validation

基础模型 (base_model)	运行文件 (operational notebook)	输出文件 (submission)	LB-Private	LB-Public
VGG16	`Keras_fine_tuning_aug_fold0.ipynb`	`submission_vgg16_ft0_aug.csv`	0.46152	0.54683
VGG16	`Keras_fine_tuning_aug_fold1.ipynb`	`submission_vgg16_ft1_aug.csv`	0.34059	0.34422
VGG16	`Keras_fine_tuning_aug_fold2.ipynb`	`submission_vgg16_ft2_aug.csv`	0.44347	0.63316
VGG16	`Keras_fine_tuning_aug_fold3.ipynb`	`submission_vgg16_ft3_aug.csv`	0.44602	0.36146
VGG16	`Keras_fine_tuning_aug_fold4.ipynb`	`submission_vgg16_ft4_aug.csv`	0.55130	0.52167
ResNet50	`Keras_fine_tuning_ResNet50_aug_fold0.ipynb`	`submission_resnet50_ft0_aug.csv`	0.35491	0.48216
ResNet50	`Keras_fine_tuning_ResNet50_aug_fold1.ipynb`	`submission_resnet50_ft1_aug.csv`	0.45376	0.42659
ResNet50	`Keras_fine_tuning_ResNet50_aug_fold2.ipynb`	`submission_resnet50_ft2_aug.csv`	0.26854	0.26618
ResNet50	`Keras_fine_tuning_ResNet50_aug_fold3.ipynb`	`submission_resnet50_ft3_aug.csv`	0.41992	0.49777
ResNet50	`Keras_fine_tuning_ResNet50_aug_fold4.ipynb`	`submission_resnet50_ft4_aug.csv`	0.37796	0.36269

集成学习 (Ensemble Learning)
Keras_5fold.ipynb

KNN
KNN.ipynb

baseline：基准模型

Resnet50_baseline_train_validation.ipynb
Resnet50_baseline_test.ipynb

MingyiZhang/mlnd_capstone