Deep Learning reflection.md

Tensorflow learning guide : http://www.yuchao.us/2017/01/tensorflow-my-learning-experience.html

Tensor board: http://www.yuchao.us/2017/02/tensorboard.html

It's time to reflection on what projects I have accomplished. The post date refers to the blog post date.

Postdate	dateset	records	in/out	Note
2016-10-25	MNIST	60 k/ 10 k	28^2 /10	NNDL1,sigmoid
2016-10-27	MNIST	60 k/ 10 k	28^2 /10	MNDL3,softmax, cross entropy
2016-10-31	MNIST	60 k/ 10 k	28^2 /10	MNDL4, CNN, theano
2017-2-2	notMNIST	520 k	28^2 /10	Tensorflow practice
2017-2-2	SVHN	99 k	1024/10	Tensorflow
2017-2-13	bikeshareDC	17.4 k	16/reg	from scratch, 56-30-1 net
2017-2-27	movieReview	25 k	10 k/ 2	from scratch, 10k-10-1 net
same	same	same	same	tflearn, 10k-200-25-2 softmax
same				miniflow
same	CIFAR-10	60 k	32^2/10	tensorflow, cnn(32)-1024-10
2017-3-30	movieReview	25 k	10 k/2	tf.contrib.rnn
same	TV script	11.5 k		tf.contrib.rnn
2017-6-6	flowers	3.67 k	224^2/5	vgg transfer(5 cnn)-256-5 tf.contrib.layers.fully_connected
2017-4-1	TrafficSign	50 k	36^2/43	tf 5 layer LeNet, no dropout
2017-4-7	ImageNet		227^2/1k	AlexNet: 5 cnn + 3 fc
2017-4-7	behaviorClone		50 k/reg	Nvidia: 5 cnn + 4 fc, 5 M para
2017-4-15	car image	18 k	64^2/2	LeNet: 2cnn + 3 fc

capstone video: https://youtu.be/w15GpupQusM

several tips:

Some suggestions from reviewer:

1. use TensorFlow to convert RGB images to grayscale
2. implement early termination, instead of a fixed number of epochs
3. use image augmentation technique to rebalance the number of examples for each class.Check out this article
4. use plt.bar(x,y) to visualize softmax probability distribution.

Appendix: the battle for deep learning frameworks