This repository is what I developed in AI lab.

GOAL : Parallel Training in PipeDream

(Please compile in Linux)

Two computers connect with socket.

• HOW TO EXECUTE

  1. go to aio folder
  2. add code "#define NOT_AIO" in both files.
  3. gcc -o server server.c -lrt
  4. ./server
  5. gcc -o client client.c -lrt
  6. ./client output

Two computers connect with socket and communicates in the way of Asynchronous non-blocking I/O instead of Synchronous blocking I/O.

• HOW TO EXECUTE

  1. go to aio folder
  2. gcc -o server server.c aiocb.c -lrt
  3. ./server <ip> <port>
  4. gcc -o client client.c aiocb.c -lrt
  5. ./client <ip> <port>

• HOW TO EXECUTE

  1. go to DL folder
  2. g++ -o step(n) step(n).cpp (ex : g++ -o step1 step1.cpp)
  3. ./step(n) (ex : ./step1)

Simple Linear Regression with neuron class

env [ learning_rate : 0.01, epoch : 2000]

Training data

x = [1, 2, 3]
y = [1, 2, 3]

goal

w = 1, b = 0

AILAB

training 1960
w = 0.999898 b = 0.000232593
training 1980
w = 0.999905 b = 0.00021603
training 2000
w = 0.999912 b = 0.000200646

DeepLearningZeroToAll : lab-02-1-linear_regression.py

1960 1.4711059e-05 [1.004444] [-0.01010205]
1980 1.3360998e-05 [1.0042351] [-0.00962736]
2000 1.21343355e-05 [1.0040361] [-0.00917497]

(My program is more correct because it executes forward propagation 4 times in one training.)

Multi Variable Linear Regression with neuron class

env [ learning_rate : 0.00001, epoch : 2000]

Training data

x = [[73,80,75],
    [93,88,93],
    [89,91,90],
    [96,98,100],
    [73,66,70]]
y = [152,185,180,196,142]

AILAB

training 2000
w = 1.03983 0.544093 0.429563 b = 0.00476187
COST : 0.437949
Prediction: 151.657 184.539 180.723 196.106 141.892 

DeepLearningZeroToAll : lab-04-1-multi_variable_linear_regression.py

2000 Cost:  4.897013 
Prediction:
 [148.15846 186.88112 179.63167 195.81955 144.45374]

(My program is more correct because it executes forward propagation 4 times in one training.)

basic neural network(use softmax cross-entropy)

env [ learning_rate : 0.1, epoch : 2000]

Training data

x = [[1, 2, 1, 1],
      [2, 1, 3, 2],
      [3, 1, 3, 4],
      [4, 1, 5, 5],
      [1, 7, 5, 5],
      [1, 2, 5, 6],
      [1, 6, 6, 6],
      [1, 7, 7, 7]]
y = [[0, 0, 1],
      [0, 0, 1],
      [0, 0, 1],
      [0, 1, 0],
      [0, 1, 0],
      [0, 1, 0],
      [1, 0, 0],
      [1, 0, 0]]

AILAB

training 2000
COST : 0.15368

Prediction: 
4.24636e-06 0.00123228 0.998763 
0.00329391 0.0805825 0.916124 
8.80453e-08 0.162293 0.837707 
3.38348e-06 0.850948 0.149049 
0.251397 0.736676 0.0119277 
0.129234 0.870734 3.18653e-05 
0.760781 0.239184 3.54718e-05 
0.918814 0.0811854 6.45421e-07

DeepLearningZeroToAll : lab-06-1-softmax_classifier.py

2000 0.15216155
--------------
[[4.5496481e-06 1.1123288e-03 9.9888307e-01]
 [2.4959880e-03 8.2680933e-02 9.1482306e-01]
 [2.0493591e-08 1.6134089e-01 8.3865911e-01]
 [4.7109361e-07 8.5135329e-01 1.4864619e-01]
 [2.4689783e-01 7.4094397e-01 1.2158173e-02]
 [1.2933041e-01 8.7065351e-01 1.6084563e-05]
 [7.6624662e-01 2.3372723e-01 2.6257312e-05]
 [9.1828001e-01 8.1719600e-02 4.6712012e-07]]

adapt basic neural network to mnist

The correct solution was fitted with a probability of 93%.

optimize : Mini-batch Gradient Descent(MGD)

you can change LEARNING_RATE, DATA_SET, batch_size

env : [learning_rate : 0.01, epoch : 50, batch_size : 1]

Training data : mnist

AILAB

training 48
cost : 0.24433
training 49
cost : 0.244146
training 50
cost : 0.243966
Correct Rate : 0.931633

DeepLearningZeroToAll : lab-10-1-mnist_softmax.py

Epoch: 0048, Cost: 0.257301215
Epoch: 0049, Cost: 0.256664445
Epoch: 0050, Cost: 0.256630104
Learning Finished!
Accuracy: 0.9178

adapt basic multi layer neural network to mnist

The correct solution was fitted with a probability of 92%.

Hidden layer uses ReLU activation, output layer uses softmax activation.

optimize : Mini-batch Gradient Descent(MGD)

env : [epoch : 15, layer : (784,256),(256,10)]

Training data : mnist

AILAB (learning_rate : 0.001, batch_size : 1)

training 8
COST : 0.0731284
Correct Rate : 0.93985

...

training 14
COST : 0.0813918
Correct Rate : 0.934033

training 15
COST : 0.0881044
Correct Rate : 0.931167

DeepLearningZeroToAll : lab-10-2-mnist_nn.py (learning_rate : 0.01, batch_size : 100)

Epoch: 0013 cost = 1.956830746
Epoch: 0014 cost = 1.851161093
Epoch: 0015 cost = 1.748512294
Learning Finished!
Accuracy: 0.9106

adapt advanced multi layer neural network to mnist

The correct solution was fitted with a probability of 92%.

initializer : Xavier(softmax), He(ReLU)

optimize : Adam

env : [learning_rate : 0.001, epoch : 15, batch_size : 100, layer : (784,256),(256,10)]

Training data : mnist

AILAB

DeepLearningZeroToAll : lab-10-3-mnist_xavier.py

Epoch: 0013 cost = 0.048236459
Epoch: 0014 cost = 0.053944308
Epoch: 0015 cost = 0.047278727
Learning Finished!
Accuracy: 0.9659

(http://shuuki4.github.io/deep%20learning/2016/05/20/Gradient-Descent-Algorithm-Overview.html)

Modify linux type in aiocb.h (ex : #define LINUX ubuntu)

• HOW TO EXECUTE

  1. g++ -o step1 step1.cpp aiocb.c -lrt
  2. ./step1 -i 127.0.0.1 -p 9999 -l output
  3. open other terminal, and ./step1 -i 127.0.0.1 -p 9998 -l input
  4. Then, this program will ask next layer's ip & port. (Ex : 127.0.0.1, 9999)