Welcome to my homepage

NN Minor Project 1차완본

2016-12-15 01:50:13

컴퓨터

## 분리수거 쓰레기와 음식물 를 버리러 가자 ㅠ 내일 어떻게 일어나나 ㅠ

##

# Tensorflow tutorial: https://www.tensorflow.org/versions/r0.12/tutorials/mnist/pros/index.html

#

import tensorflow as tf

import numpy as np

from sklearn.manifold import TSNE

from mpl_toolkits.mplot3d import Axes3D

import os

import matplotlib as mpl

if os.environ.get('DISPLAY','') == '':

print('no display found. Using non-interactive Agg backend')

mpl.use('Agg')

import matplotlib.pyplot as plt

#data setting

# Get MNIST data set from tensorflow

from tensorflow.examples.tutorials.mnist import input_data

mnist_set = input_data.read_data_sets('images/MNIST/', one_hot = True)

train_img = mnist_set.train.images

train_label = mnist_set.train.labels

test_img = mnist_set.test.images

test_label = mnist_set.test.labels

# Options for training

dim="2d"

stddev = 0.1

biasInit = 0.1

isDropout = False

training_epochs = 100

batch_size = 2000

point_N = 1000

display_step = 1

sum_level = "full" #full, pool

prefix = "sum_only" #total, sum_only

# Options for loading a pretrained model

load_dir = "models/"

load_keyword = "cnn_tanh_mnist_20000.epoch-"

load_epoch = "20000"

test_rand_a = np.random.randint(test_img.shape[0], size=test_img.shape[0])

test_rand_b = np.random.randint(test_img.shape[0], size=test_img.shape[0])

test_selt_a = []

test_selt_b = []

test_selt_c = []

for j in range(test_img.shape[0]):

a = np.where(test_label[test_rand_a[j]] == 1.)[0][0]

b = np.where(test_label[test_rand_b[j]] == 1.)[0][0]

if a + b >= 10:

j = j - 1

continue

test_rand_c = np.random.randint(test_img.shape[0], size = test_img.shape[0])

# find correct image

for k in range(test_img.shape[0]):

c = np.where(test_label[test_rand_c[k]] == 1.)[0][0]

if c == a + b:

test_selt_a.append(test_rand_a[j])

test_selt_b.append(test_rand_b[j])

test_selt_c.append(test_rand_c[k])

break

if len(test_selt_a) == test_img.shape[0]:

&nb\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0sp;break

test_batch_a = test_img[test_selt_a, :]

test_batch_b = test_img[test_selt_b, :]

test_batch_c = test_img[test_selt_c, :]

test_batch_a_label = test_label[test_selt_a, :]

test_batch_b_label = test_label[test_selt_b, :]

test_batch_c_label = test_label[test_selt_c, :]

train_img_num = []

for i in range(10):

train_img_num.append([])

for i in range(train_img.shape[0]):

train_img_num[np.where(train_label[i] == 1.)[0][0]].append(i)

print "Initialized"

##

# Build Graph

# The concept of graph is a sort of function and Placeholders are arguments for that function

#input & output

x = tf.placeholder(tf.float32, shape=[None, train_img.shape[1]])

x_image = tf.reshape(x, [-1,28,28,1])

#y is one-hot .. but it must be float32

y_ = tf.placeholder(tf.float32, shape=[None, train_label.shape[1]])

#Variables

# Initialization

# return variable in the specified shape with stddev 0.1

def weight_variable(shape):

initial = tf.truncated_normal(shape, stddev=stddev) # Initialize w/ stddev 0.1 normal distribution

return tf.Variable(initial, name='weight')

# return constant bias in the specified shape

def bias_variable(shape):

initial = tf.constant(biasInit, shape=shape) # Initilaize w/ 0.1

return tf.Variable(initial, name='biases')

# 2d cnn layer with stride 1

# padding SAME is valid ??(valid, full.. and .. what?? I can't remember ;;;)

def conv2d(x, W):

return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

# Max pooling

def max_pool_2x2(x):

return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

###

# 2 Convolution layers

###

# First Convolution Layer : Input to CNN(1 to 32)

# In order to build a deep network, we stack several layers of this type.

# The first layer will have 32 features for each 5x5 patch.

W_conv1 = weight_variable([5, 5, 1, 32])

b_conv1 = bias_variable([32])

# Applying Max pooling Convolution

# We convolve x_image w/ the weight tensor, add the bias, apply the tanh function,and finally maxpool

h_conv1 = tf.nn.tan\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0h(conv2d(x_image, W_conv1) + b_conv1)

h_pool1 = max_pool_2x2(h_conv1)

##########################

##

# Second Convolutional Layer : CNN to CNN (32 to 64)

# In order to build a deep network, we stack several layers of this type.

# The second layer will have 64 features for each 5x5 patch.

W_conv2 = weight_variable([5, 5, 32, 64])

b_conv2 = bias_variable([64])

# Applying Max pooling Convolution

# We convolve x_image w/ the weight tensor, add the bias, apply the tanh function,and finally maxpool

h_conv2 = tf.nn.tanh(conv2d(h_pool1, W_conv2) + b_conv2)

h_pool2 = max_pool_2x2(h_conv2)

##########################

###

# 2 Fully Connected layers

###

##

# Densely Connected Layer : Fully connected layer w/ 1024 neurons (CNN to FNN (7 x 7 x 64 to 1024 ) )

# Now that the image size has been reduced to 7x7

W_fc1 = weight_variable([7 * 7 * 64, 1024])

# we add a fully-connected layer with 1024 neurons to allow processing on the entire image.

b_fc1 = bias_variable([1024])

# We reshape the tensor from the pooling layer into a batch of vectors,

h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])

# multiply by a weight matrix, add a bias, and a tanh

h_fc1 = tf.nn.tanh(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

##########################

##

# Dropout: To reduce overfitting, we will apply dropout before the readout layer.

# We create a placeholder for the probability that a neuron's output is kept during dropout.

# This allows us to turn dropout on during training, and turn it off during testing.

keep_prob = tf.placeholder(tf.float32)

# TensorFlow's tf.nn.dropout op automatically handles scaling neuron outputs in addition

# to masking them, so dropout just works without any additional scaling.

h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

# For this small convolutional network, performance is actually nearly identical with and

# without dropout. Dropout is often very effective at reducing overfitting, but it is most

# useful when training very large neural networks.

##

# Readout Layer : Fully Connected to Output Layer

# Finally, we add a layer, just like for the one layer softmax regression.

# The last layer shouldn't be activation function??

W_fc2 = weight_variable([1024, 10])

b_fc2 = bias_variable([10])

if isDropout :

&nb\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0sp; y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

else:

y_conv = tf.matmul(h_fc1, W_fc2) + b_fc2

################################

# Finally, we can do something with the layer built up

# To train and evaluate it we will use code that is nearly identical to that for the simple

# one layer SoftMax network above.

cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv, y_))

correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))

accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

# The differences are that:

# 1) We will replace the steepest gradient descent optimizer with the more sophisticated ADAM opti.

# 2) We will include the additional parameter keep_prob in feed_dict to control the dropout rate.

# 3) We will add logging to every 100th iteration in the training process.

# ADAM optimizer rather than the steepest gradient descent optimizer

train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

################################

# Load Model #

################################

# Load saver

saver = tf.train.Saver()

x0 = tf.placeholder(tf.float32, shape=[None, train_img.shape[1]])

x0_image = tf.reshape(x0, [-1,28,28,1])

x1 = tf.placeholder(tf.float32, shape=[None, train_img.shape[1]])

x1_image = tf.reshape(x1, [-1,28,28,1])

x2 = tf.placeholder(tf.float32, shape=[None, train_img.shape[1]])

x2_image = tf.reshape(x2, [-1,28,28,1])

y0_ = tf.placeholder(tf.float32, shape=[None, train_label.shape[1]])

y1_ = tf.placeholder(tf.float32, shape=[None, train_label.shape[1]])

y2_ = tf.placeholder(tf.float32, shape=[None, train_label.shape[1]])

ARW_fc1 = tf.identity(W_fc1)

ARb_fc1 = tf.identity(b_fc1)

ARW_fc2 = tf.identity(W_fc2)

ARb_fc2 = tf.identity(b_fc2)

#ARW_fc3 = weight_variable([1024, 2])

#ARb_fc2 = bias_variable([2])

ARW_conv1 = tf.identity(W_conv1)

ARW_conv2 = tf.identity(W_conv2)

ARb_conv1 = tf.identity(b_conv1)

ARb_conv2 = tf.identity(b_conv2)

h_conv10 = tf.nn.tanh(conv2d(x0_image, ARW_conv1) + ARb_conv1)

h_conv11 = tf.nn.tanh(conv2d(x1_image, ARW_conv1) + ARb_conv1)

h_conv12 = tf.nn.tanh(conv2d(x2_image, ARW_conv1) + ARb_conv1)

h_pool10 = max_pool_2x2(h_conv10)

h_pool11 = max_pool_2x2(h_conv11)

\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0h_pool12 = max_pool_2x2(h_conv12)

h_conv20 = tf.nn.tanh(conv2d(h_pool10, ARW_conv2) + ARb_conv2)

h_conv21 = tf.nn.tanh(conv2d(h_pool11, ARW_conv2) + ARb_conv2)

h_conv22 = tf.nn.tanh(conv2d(h_pool12, ARW_conv2) + ARb_conv2)

h_pool20 = max_pool_2x2(h_conv20)

h_pool21 = max_pool_2x2(h_conv21)

h_pool22 = max_pool_2x2(h_conv22)

h_pool20_flat = tf.reshape(h_pool20, [-1, 7 * 7 * 64])

h_pool21_flat = tf.reshape(h_pool21, [-1, 7 * 7 * 64])

h_pool22_flat = tf.reshape(h_pool22, [-1, 7 * 7 * 64])

h_fc10 = tf.nn.tanh(tf.matmul(h_pool20_flat, ARW_fc1) + ARb_fc1)

h_fc11 = tf.nn.tanh(tf.matmul(h_pool21_flat, ARW_fc1) + ARb_fc1)

h_fc12 = tf.nn.tanh(tf.matmul(h_pool22_flat, ARW_fc1) + ARb_fc1)

h_fc10_drop = tf.nn.dropout(h_fc10, keep_prob)

h_fc11_drop = tf.nn.dropout(h_fc11, keep_prob)

h_fc12_drop = tf.nn.dropout(h_fc12, keep_prob)

if isDropout :

y_conv0 = tf.matmul(h_fc10_drop, ARW_fc2) + ARb_fc2

y_conv1 = tf.matmul(h_fc11_drop, ARW_fc2) + ARb_fc2

y_conv2 = tf.matmul(h_fc12_drop, ARW_fc2) + ARb_fc2

else:

y_conv0 = tf.matmul(h_fc10, ARW_fc2) + ARb_fc2

y_conv1 = tf.matmul(h_fc11, ARW_fc2) + ARb_fc2

y_conv2 = tf.matmul(h_fc12, ARW_fc2) + ARb_fc2

ce_vec0 = tf.nn.softmax_cross_entropy_with_logits(y_conv0, y0_)

ce_vec1 = tf.nn.softmax_cross_entropy_with_logits(y_conv1, y1_)

ce_vec2 = tf.nn.softmax_cross_entropy_with_logits(y_conv2, y2_)

cross_entropy0 = tf.reduce_mean(ce_vec0)

cross_entropy1 = tf.reduce_mean(ce_vec1)

cross_entropy2 = tf.reduce_mean(ce_vec2)

if sum_level == "pool":

sum_loss_vec = tf.add(h_pool21_flat, h_pool22_flat) - h_pool20_flat

#sum_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(tf.add(h_pool21_flat, h_pool22_flat), h_pool20_flat))

elif sum_level == "full":

sum_loss_vec = tf.add(h_fc11, h_fc12) - h_fc10

#sum_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(tf.add(h_fc11, h_fc12), h_fc10))

sum_loss_vec_sq = tf.square(sum_loss_vec)

sum_loss = tf.reduce_mean(sum_loss_vec_sq)

if prefix == "sum_only" :

total_optimizer = tf.train.AdamOptimizer(1e-3).minimize(sum_loss)

else :

#total_loss = 4/(1/cross_entropy0 + 1/cross_entropy1 + 1/cross_entropy2 + 1/sum_loss)

total_loss = cross_entropy0 + cross_entropy1 + cross_entropy2 + sum_loss

total_optimizer = tf.train.AdamOptimizer(1e-3).minimize(total_loss)

#for checking classifiation task

correct_prediction0 = tf.equal(tf.argmax(y_conv0,1), tf.argmax(y0_,1))

correct_prediction1 = tf.equal(tf.argmax(y_conv1,1), tf.argmax(y1_,1))

correct_prediction2 = tf.equal(tf.argmax(y_conv2,1), tf.argmax(y2_,1))

accuracy0 = tf.reduce_mean(tf.cast(correct_prediction0, tf.float32))

accuracy1 = tf.red\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0uce_mean(tf.cast(correct_prediction1, tf.float32))

accuracy2 = tf.reduce_mean(tf.cast(correct_prediction2, tf.float32))

# Restore graph session

sess = tf.Session()

init = tf.initialize_all_variables()

sess.run(init)

ckpt = tf.train.get_checkpoint_state(load_dir)

if ckpt and ckpt.model_checkpoint_path:

saver.restore(sess, load_dir + load_keyword + load_epoch)

################################

# Train and Evaluate the Model #

################################

# Print out : I hope the final test set accruacy after runnign this code will be 99.2%

if dim == "2d":

# model = TSNE(n_components=2, random_state=0)

model = TSNE(perplexity=50, n_components=2, init='pca', n_iter=5000)

elif dim == "3d":

model = TSNE(perplexity=50, n_components=3, init='pca', n_iter=5000)

np.set_printoptions(suppress=True)

##

# Training

for epoch in range(training_epochs):

# Checking the classification task & vector summation

test_acc = sess.run(accuracy0, feed_dict={x0: mnist_set.test.images, y0_: mnist_set.test.labels, keep_prob: 1.0});

test_sum_cost = sess.run(sum_loss, feed_dict={x1: test_batch_a, x2: test_batch_b, x0: test_batch_c, y0_: test_batch_c_label, y1_: test_batch_a_label, y2_: test_batch_b_label})

print ("epoch: %03d/%03d, test image acc %.6f, test sum_loss: %.6f"% (epoch, training_epochs, test_acc, test_sum_cost))

# Checking the linear relation

if sum_level == "pool":

hpf = sess.run(h_pool20_flat, feed_dict={x0:mnist_set.test.images[0:point_N]})

elif sum_level == "full":

hpf = sess.run(h_fc10, feed_dict={x0:mnist_set.test.images[0:point_N]})

#save to image

tsne = model.fit_transform(hpf)

plt.figure(figsize=(36,36))

if dim == "3d":

fig = plt.figure(figsize=(36,36))

ax = fig.add_subplot(111, projection='3d')

for i in range(point_N):

if dim == "2d":

plt.scatter(tsne[i][0], tsne[i][1])

label = np.where(test\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0_label[i] == 1.)[0][0]

plt.annotate('%s' % str(label), xy=(tsne[i][0], tsne[i][1]), xytext=(5, 5), textcoords='offset points', ha='right', va='bottom')

elif dim == "3d":

ax.scatter(tsne[i][0], tsne[i][1], tsne[i][2])

avg_cost = 0.

filename = prefix + "_mnist_tanh_" + dim + "_" + sum_level + "_cnn_" + str(epoch) + "_summation_" + str("%.6f"%test_sum_cost) + "_classification_" + str("%.6f"%test_acc) + ".png"

plt.savefig(filename)

plt.close()

#Training

num_batch = int(train_img.shape[0] / batch_size )

for i in range(num_batch):

rand_a = np.random.randint(train_img.shape[0], size=train_img.shape[0])

rand_b = np.random.randint(train_img.shape[0], size=train_img.shape[0])

selt_a = []

selt_b = []

selt_c = []

for j in range(batch_size):

a = np.where(train_label[rand_a[j]] == 1.)[0][0]

b = np.where(train_label[rand_b[j]] == 1.)[0][0]

if a + b >= 10:

j = j - 1

continue

rand_c = np.random.randint(train_img.shape[0], size = train_img.shape[0])

# find correct image

for k in range(train_img.shape[0]):

c = np.where(train_label[rand_c[k]] == 1.)[0][0]

if c == a + b:

selt_a.append(rand_a[j])

selt_b.append(rand_b[j])

selt_c.append(rand_c[k])

break

if len(selt_a) == batch_size:

&\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0nbsp; break

if len(selt_a) != len(selt_b) or len(selt_a) != len(selt_c):

print("Something wrong!!", len(selt_a)," ",len(selt_b)," " ,len(selt_c))

batch_a = train_img[selt_a, :]

batch_b = train_img[selt_b, :]

batch_c = train_img[selt_c, :]

batch_a_label = train_label[selt_a, :]

batch_b_label = train_label[selt_b, :]

batch_c_label = train_label[selt_c, :]

sess.run(total_optimizer, feed_dict={x1: batch_a, x2: batch_b, x0: batch_c, y0_: batch_c_label, y1_: batch_a_label, y2_:batch_b_label})

avg_cost += sess.run(sum_loss, feed_dict={x1: batch_a, x2: batch_b, x0: batch_c, y0_: batch_c_label, y1_: batch_a_label, y2_: batch_b_label})/num_batch

print ("epoch: %03d/%03d, train sum_loss: %.6f"% (epoch, training_epochs, avg_cost))

▼ more

선형성, 분류력을 가지는 네트워크

2016-12-15 01:46:19

컴퓨터

그것은 데이터의 문제였다.

train_subset_mnist_tanh_full_cnn_256_sum_0.297016_imgcls_1.000000

test_subset_mnist_tanh_full_cnn_254_sum_0.988055_imgcls_0.969500

#Tags TensorFlow, 텐저플로우, 텐서플로우, 텐져플로우, tensor

▼ more

TensorFlow 소스들

2016-12-07 08:39:14

컴퓨터

음..

▼ more

TensorFlow 실전 1 Vector Summation & Shared Network

2016-12-05 02:29:42

컴퓨터

왜일까 아래 코딩을 할때는 노래를 듣지도 않았는데 "금지된 경호" 와 "She's gone" 이 내머릿속에 맴돌던 이유는..;

# Tags TensorFlow, 텐저플로우, 텐서플로우, 텐져플로우, tensor

# how to visualize a 2-D or 3-D scatter plot of weights in TensorFlow done.

# coding: utf-8

# In[1]:

import tensorflow as tf

import numpy as np

import matplotlib.pyplot as plt

get_ipython().magic('matplotlib inline')

#######################

# data setting #

#######################

train_input1 = []

train_input2 = []

train_target = []

for i in range(50000):

x1 = np.random.randint(10)

x2 = np.random.randint(10)

if(x1 + x2 < 10):

train_input1.append(np.eye(10)[x1])

train_input2.append(np.eye(10)[x2])

train_target.append(np.eye(10)[x1 + x2])

#print(x1," + ", x2," = ", x1+x2)

if(len(train_input1) == 100):

break

print("Training Data: ",len(train_input1),",",len(train_input2),"Target",len(train_target));

train_input1 = np.float32(train_input1)

train_input2 = np.float32(train_input2)

train_target = np.float32(train_target)

test_input1 = []

test_input2 = []

test_target = []

check = []

for i in range(10):

check.append(np.eye(10)[i])

for j in range (10):

if(i + j < 10):

test_input1.append(np.eye(10)[i])

test_input2.append(np.eye(10)[j])

test_target.append(np.eye(10)[i + j])

print("Test Data: ",len(test_input1),",",len(test_input2),"Target",len(test_target));

test_input1 = np.float32(test_input1)

test_input2 = np.float32(test_input2)

test_target = np.float32(test_target)

# In[2]:

########################

# Graph Building #

########################

#placeholder : Graph Argument

x1 = tf.placeholder(tf.float32, shape=[None, 10])

x2 = tf.placeholder(tf.float32, shape=[None, 10])

y = tf.placeholder(tf.float32, shape=[None, 10])

check_y = tf.placeholder(tf.float32, shape=[None, 10])

# layer

def one_layer(_x,_W,_b):

return tf.add(tf.matmul(_x,_W),_b)

## Image to Shared Network

w_dev = 0.3

hl_size = 10

learning_rate = 0.05

Image_Vector_Size = 2

# Shared Wegiths : _W1, _b1, _W2, _b2

# Image(10) to Hidden Lay\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0er(hl_size) including shared weights

_W1 = tf.Variable(tf.random_normal([train_input1.shape[1], hl_size], stddev=w_dev))

_b1 = tf.Variable(tf.random_normal([hl_size]))

# Hidden Layer(hl_size) to Image Vector(2)

_W2 = tf.Variable(tf.random_normal([hl_size, Image_Vector_Size], stddev=w_dev))

_b2 = tf.Variable(tf.random_normal([Image_Vector_Size]))

# Layers

# Inputs to Hidden

x1h = one_layer(x1, _W1, _b1)

x2h = one_layer(x2, _W1, _b1)

# Hidden to Input Image Vectors

v1 = one_layer(x1h, _W2, _b2)

v2 = one_layer(x2h, _W2, _b2)

#summation

out = tf.add(v1, v2) #Returns x + y element-wise

# Target to Hidden

yh = one_layer(y, _W1, _b1)

# Hidden to Target Image Vector

vy = one_layer(yh, _W2, _b2)

loss = tf.reduce_mean(tf.square(vy - out))

#Gradient Descent Methods

#optm = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss)

optm = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss)

#Get point.x, point.y

xy = one_layer(one_layer(check_y, _W1, _b1), _W2, _b2)

#Init Graph

init = tf.global_variables_initializer()

print ("Graph build")

# In[3]:

#Start Session

sess = tf.Session()

sess.run(init)

training_epochs = 100

batch_size = 50

display_step = 20

# In[4]:

for epoch in range(training_epochs):

avg_cost = 0.

num_batch = int(train_input1.shape[0] / batch_size )

for i in range(num_batch):

randidx = np.random.randint(train_input1.shape[0], size=batch_size)

batch_x1 = train_input1[randidx, :]

batch_x2 = train_input2[randidx, :]

batch_ys = train_target[randidx, :]

sess.run(optm, feed_dict={x1: batch_x1, x2: batch_x2, y: batch_ys})

avg_cost += sess.run(loss, feed_dict={x1: batch_x1, x2: batch_x2, y: batch_ys})/num_batch

if epoch % display_step == 0:

print ("epoch: %03d/%03d , cost: %.6f"% (epoch, training_epochs,avg_cost))

result_xy = sess.run(vy, feed_dict={y: check})

&nb\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0sp; point_x = []

point_y = []

point_z = []

#drawing scatter plots of image vector

for j in range(len(check)):

print(j, result_xy[j])

point_x.append(result_xy[j][0])

point_y.append(result_xy[j][1])

plt.plot(point_x, point_y, "o")

plt.show()

# In[5]:

avg_loss = 0.

if batch_size > test_input1.shape[0]:

num_batch_test = 1

else:

num_batch_test = int(test_input1.shape[0]/batch_size)

for i in range(num_batch_test):

batch_x1 = test_input1[i * batch_size : (i+1) * batch_size ]

batch_x2 = test_input2[i * batch_size : (i+1) * batch_size ]

batch_y = test_target[i * batch_size : (i+1) * batch_size ]

# it gives cost of batch

avg_loss += sess.run(loss, feed_dict={x1 : batch_x1, x2: batch_x2, y : batch_y})

print ("Training complete, Test Average Loss : %.6f" % (avg_loss / num_batch_test,))

result_xy = sess.run(vy, feed_dict={y: check})

x = []

y = []

z = []

for j in range(len(check)):

print(j, result_xy[j])

x.append(result_xy[j][0])

y.append(result_xy[j][1])

plt.plot(x, y, "o")

# In[ ]:

▼ more

Lee, Kyungmin'sPersonal Blog

Lee, Kyungmin's
Personal Blog