import tensorflow as tf import numpy as np import matplotlib.pyplot as plt def load_weather(file_path, seq_length): weather = np.loadtxt(file_path, delimiter=' ', skiprows=1) weather = (weather[:,1:]) weather = weather.transpose() weather = weather.reshape(-1, 1) xx, yy = [], [] for i in range(len(weather)-seq_length): xx.append(weather[i:i+seq_length]) yy.append(weather[i+seq_length]) return np.float32(xx), np.float32(yy) # 날씨 데이타# http://www.weather.go.kr/weather/climate/past_table.jspdef rnn_weather(): batch_size = 1 seq_length = 7 n_classes = 1 hidden_size = 5 output_dim = 1 xx, yy = load_weather("Data/weather.txt", seq_length) xx_test, yy_test = load_weather("Data/weather_test.txt", seq_length) x = tf.placeholder(tf.float32, shape=[None, 7, 1]) cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size, activation=tf.tanh) output, state_ = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32) z = tf.contrib.layers.fully_connected(inputs=output[:, -1], num_outputs=1, activation_fn=None) loss = tf.reduce_mean((z-yy)**2) optimizer = tf.train.AdamOptimizer(0.01) train = optimizer.minimize(loss) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for i in range(5000): sess.run(train, {x:xx}) print(sess.run(loss , {x:xx})) p = sess.run(z, {x:xx_test[30:100]}) plt.plot(p, 'r') plt.plot(yy_test[30:100], 'g') plt.show() rnn_weather()
2018년 10월 20일 토요일
RNN 연습
2018년 10월 17일 수요일
RNN simple
import tensorflow as tf
import numpy as np
def rnn_1():
vocab = np.array(['e', 'n', 'o', 'r', 's', 't'])
x = [[0., 0., 0., 0., 0., 1.], # t
[1., 0., 0., 0., 0., 0.], # e
[0., 1., 0., 0., 0., 0.], # n
[0., 0., 0., 0., 1., 0.], # s
[0., 0., 1., 0., 0., 0.]] # o
y = [[1, 0, 0, 0, 0, 0], # e
[0, 1, 0, 0, 0, 0], # n
[0, 0, 0, 0, 1, 0], # s
[0, 0, 1, 0, 0, 0], # o
[0, 0, 0, 1, 0, 0]] # r
w = tf.Variable(tf.random_uniform([6,6]))
b = tf.Variable(tf.random_uniform([6]))
# (5,6) = (5,6) @ (6, 6)
z = tf.matmul(x, w) + b
h = tf.nn.softmax(z)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss=loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
print(sess.run(loss))
print(sess.run(h))
print(np.argmax(sess.run(h), axis=1 ))
print(vocab[np.argmax(sess.run(h), axis=1)])
rnn_1()
def rnn_2_2():
vocab = np.array(['e', 'n', 'o', 'r', 's', 't'])
# tenso
x = [[0., 0., 0., 0., 0, 1.],
[1., 0., 0., 0., 0, 0.],
[0., 1., 0., 0., 0, 0.],
[0., 0., 0., 0., 1, 0.],
[0., 0., 1., 0., 0, 0.]]
# ensor
y = [[1., 0., 0., 0., 0., 0.], # 클래스 6개. Dense 형태
[0., 1., 0., 0., 0., 0.],
[0., 0., 0., 0., 1., 0.],
[0., 0., 1., 0., 0., 0.],
[0., 0., 0., 1., 0., 0.]]
hidden_size = 2 # 전개량
x = np.float32([x]) # dynamic_rnn은 3차원 데이터를 입력으로 요구 함. x: 2차원 -> 3차원
# RNN layer
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, _states = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32)
print(outputs.shape)
# Softmax layer
w = tf.Variable(tf.random_uniform([hidden_size, 6]))
b = tf.Variable(tf.random_uniform([6]))
# y(5, 6) = x(5, hidden_size) @ w(hidden_size, 6) # w에서 x와 만나는 것은 피처의 개수. 뒤는 class의 개수
z = tf.matmul(outputs[0], w) + b
h = tf.nn.softmax(z)
loss_i = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
loss = tf.reduce_mean(loss_i)
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
pred = sess.run(z)
pred_arg = np.argmax(pred, axis=1)
print(i, sess.run(loss), pred_arg, ''.join(vocab[pred_arg])) #
print('-' * 50)
sess.close()
# one-hot encoding 유지한 상태에서 길이가 같은 여러 단어의 다음 글자 prediction
# 개인적으로 y(target, true value)도 one-hot encoding 유지한 것이 더 직관적이고 좋다.
# 아래 나오는 sparse_softmax_cross_entropy_with_logits와 sequence_loss는 과도하게 복잡하고, 최종 결과도 one-hot encoding된 결과이기 때문에, 불필요한 혼란 제거를 위해서는 그냥 y도 one-hot encoding 유지하는게 좋지 않을까.
# 강사님은 왜 굳이 sparse 형태를 사용하려 하는 것일까...
import tensorflow as tf
import numpy as np
from sklearn import preprocessing
def make_onehot(words):
data = list(''.join(words))
lb = preprocessing.LabelBinarizer().fit(data)
xx, yy = [], []
for text in words:
onehot = lb.transform(list(text))
xx.append(onehot[:-1])
yy.append([onehot[i] for i in range(1,6)])
return np.float32(xx), np.int32(yy), lb.classes_
def rnn_words(words):
hidden_size = 5
x, y, vocab = make_onehot(words)
batch_size, seq_length, n_classes = x.shape
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, states_ = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32)
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
h = tf.nn.softmax(z)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(100):
sess.run(train)
print(sess.run(loss))
result = vocab[np.argmax(sess.run(z), axis=2)]
print([''.join(x) for x in result])
rnn_words(['tensor', 'houses', 'yellow'])
# 가변길이 단어들 RNN 처리. y에 대해 one-hot encoding 유지.
import tensorflow as tf
import numpy as np
from sklearn import preprocessing
def make_onehot(words):
same_length_words = []
m = max([len(w) for w in words])
for w in words:
w += ' '*(m-len(w))
same_length_words.append(w)
data = list(''.join(words))
lb = preprocessing.LabelBinarizer().fit(data)
print(lb.classes_)
xx, yy = [], []
for w in same_length_words:
onehot = lb.transform(list(w))
print("onehot", onehot.shape)
# (10 = 최대 단어 길이 seq_length, 12개 알파벳 n_classes)
xx.append(onehot[:-1])
yy.append(onehot[1:])
return np.float32(xx), np.int32(yy), lb.classes_
def rnn_multi_words_different_length(words):
word_length = [len(w) for w in words]
x, y, vocab = make_onehot(words)
print("x.shape", x.shape, "y.shape", y.shape, "\n", vocab)
# (4, 10, 12) (4, 10, 12)
# ['a' 'c' 'e' 'f' 'h' 'l' 'n' 'o' 'r' 's' 't' 'w']
hidden_size = 8
batch_size, seq_length, n_classes = x.shape
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, state_ = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32,
sequence_length=word_length)
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
print("z.shape", z.shape)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss=loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
print(sess.run(loss))
result = ([''.join(x) for x in vocab[np.argmax(sess.run(z) , axis=2)]])
for w, l in zip(result, word_length):
print("result ", w[:l-1])
rnn_multi_words_different_length (['sea', 'what', 'corw', 'tensorflow'])
# one-hot label 제거
def rnn_2_3():
vocab = np.array(['e', 'n', 'o', 'r', 's', 't'])
# tenso
x = [[0., 0., 0., 0., 0, 1.],
[1., 0., 0., 0., 0, 0.],
[0., 1., 0., 0., 0, 0.],
[0., 0., 0., 0., 1, 0.],
[0., 0., 1., 0., 0, 0.]]
# ensor
y = [0, 1, 4, 2, 3] # Sparse 형태
hidden_size = 5 # 전개량
x = np.float32([x])
# RNN layer
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, _states = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32)
print(outputs.shape)
# Softmax layer. Fully Connected Layer
w = tf.Variable(tf.random_uniform([hidden_size, 6]))
b = tf.Variable(tf.random_uniform([6]))
z = tf.matmul(outputs[0], w) + b
h = tf.nn.softmax(z)
# y(5, 6) = x(5, hidden_size) @ w(hidden_size, 6) # w에서 x와 만나는 것은 피처의 개수. 뒤는 class의 개수
loss_i = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=z, labels=y)
loss = tf.reduce_mean(loss_i)
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
pred = sess.run(z)
pred_arg = np.argmax(pred, axis=1)
print(i, sess.run(loss), pred_arg, ''.join(vocab[pred_arg])) #
print('-' * 50)
sess.close()
# Day_03_07_rnn_5_different.py
import numpy as np
import tensorflow as tf
from sklearn import preprocessing
def make_onehot_different(text_array):
data = list(''.join(text_array))
lb = preprocessing.LabelBinarizer().fit(data)
max_len = max([len(t) for t in text_array])
xx, yy = [], []
for text in text_array:
if len(text) < max_len:
text += '*' * (max_len - len(text))
onehot = lb.transform(list(text))
x = onehot[:-1]
x = np.float32(x)
y = onehot[1:]
y = np.argmax(y, axis=1)
xx.append(x)
yy.append(list(y))
return np.float32(xx), tf.constant(yy), lb.classes_
def rnn_5_different(text_array, iteration=100):
x, y, vocab = make_onehot_different(text_array)
batch_size, seq_length, n_classes = x.shape
hidden_size = 2
seq_len_array = [len(t) for t in text_array]
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, _states = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32,
sequence_length=seq_len_array)
# xavier 초기화 사용
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
w = tf.ones([batch_size, seq_length])
loss = tf.contrib.seq2seq.sequence_loss(logits=z, targets=y, weights=w)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(iteration):
sess.run(train)
pred = sess.run(z)
pred_arg = np.argmax(pred, axis=2)
if i % 10 == 0:
print(i, end=' ')
for arg, length in zip(pred_arg, seq_len_array):
valid = vocab[arg]
print(''.join(valid[:length - 1]), end=' ')
print()
print('-' * 50)
print('ans :', text_array)
sess.close()
rnn_5_different(['tensor', 'sea', 'white'])
# Day_03_08_rnn_final.py
import numpy as np
import tensorflow as tf
from sklearn import preprocessing
def make_onehot_sentence(sentence, seq_length):
data = list(sentence)
lb = preprocessing.LabelBinarizer().fit(data)
print(lb.classes_)
onehot = lb.transform(data)
x = np.float32(onehot[:-1])
y = np.argmax(onehot[1:], axis=1)
# print(len(sentence))
# for i in range(len(sentence) - seq_length):
# print(i, i + seq_length, sentence[i:i+seq_length])
idx = [(i, i + seq_length) for i in range(len(sentence) - seq_length)]
xx = [x[s:e] for s, e in idx]
yy = [y[s:e] for s, e in idx]
print(*yy[:5], sep='\n')
return np.float32(xx), tf.constant(np.int32(yy)), lb.classes_
def rnn_final(sentence, iteration=100):
x, y, vocab = make_onehot_sentence(sentence, 20)
batch_size, seq_length, n_classes = x.shape
hidden_size = 7
cells = [tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size) for _ in range(2)]
multi = tf.nn.rnn_cell.MultiRNNCell(cells)
outputs, _states = tf.nn.dynamic_rnn(multi, x, dtype=tf.float32)
# xavier 초기화 사용
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
w = tf.ones([batch_size, seq_length])
loss = tf.contrib.seq2seq.sequence_loss(logits=z, targets=y, weights=w)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(iteration):
sess.run(train)
print(i, sess.run(loss))
print('-' * 50)
pred = sess.run(z)
print(pred.shape) # (151, 20, 25)
pred_arg = np.argmax(pred, axis=2)
print(pred_arg.shape) # (151, 20)
total = '*' + ''.join(vocab[pred_arg[0]])
for arg in pred_arg[1:]:
curr = ''.join(vocab[arg])
# print(curr)
total += curr[-1]
print('[{}]'.format(total))
print('[{}]'.format(sentence))
sess.close()
sentence = ("if you want to build a ship, "
"don't drum up people to collect wood and "
"don't assign them tasks and work, "
"but rather teach them to long for the endless immensity of the sea.")
rnn_final(sentence, 1000)
# if you want to build a ship
# ---------------
# 0 : if you want
# 1 : f you want
# 2 : you want t
# 3 : you want to
#
# 0 : if you want
# 1 : t to build
# Day_03_10_stock.py
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from sklearn import preprocessing, model_selection
np.set_printoptions(linewidth=1000)
def minmax_scale(data):
mx = np.max(data, axis=0)
mn = np.min(data, axis=0)
return (data - mn) / (mx - mn + 1e-7)
def rnn_stock_1():
stock = np.loadtxt('Data/stock_daily.csv',
delimiter=',', dtype=np.float32)
stock = stock[::-1]
# stock = minmax_scale(stock)
# stock = preprocessing.minmax_scale(stock)
scaler = preprocessing.MinMaxScaler().fit(stock)
stock = scaler.transform(stock)
print(stock[:3])
print(stock.shape, stock.dtype) # (732, 5) float32
x = stock
y = stock[:, -1:]
print(y.shape) # (732, 1)
seq_length = 7
n_features = 5 # n_classes
output_dim = 1
hidden_size = 10
x_data, y_data = [], []
for i in range(len(y) - seq_length):
x_data.append(x[i:i+seq_length]) # 0~6, 1~7
y_data.append(y[i+seq_length]) # 7 , 8
# print(x_data[-1], '->', y_data[-1])
x_data = np.float32(x_data)
y_data = np.float32(y_data)
print("SHAPE:", x_data.shape, y_data.shape)
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size, activation=tf.tanh)
outputs, _states = tf.nn.dynamic_rnn(cells, x_data, dtype=tf.float32)
print(outputs.shape, outputs[:, -1].shape)
z = tf.contrib.layers.fully_connected(inputs=outputs[:, -1],
num_outputs=output_dim,
activation_fn=None)
loss = tf.reduce_mean((z - y_data) ** 2)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(100):
sess.run(train)
print(i, sess.run(loss))
def rnn_stock_2():
stock = np.loadtxt('Data/stock_daily.csv',
delimiter=',', dtype=np.float32)
stock = stock[::-1]
# stock = minmax_scale(stock)
# stock = preprocessing.minmax_scale(stock)
scaler = preprocessing.MinMaxScaler().fit(stock)
stock = scaler.transform(stock)
print(stock[:3])
print(stock.shape, stock.dtype) # (732, 5) float32
x = stock
y = stock[:, -1:]
print(y.shape) # (732, 1)
seq_length = 7
n_features = 5 # n_classes
output_dim = 1
hidden_size = 10
x_data, y_data = [], []
for i in range(len(y) - seq_length):
x_data.append(x[i:i + seq_length]) # 0~6, 1~7
y_data.append(y[i + seq_length]) # 7 , 8
# print(x_data[-1], '->', y_data[-1])
x_data = np.float32(x_data)
y_data = np.float32(y_data)
# 75:25
data = model_selection.train_test_split(x_data, y_data, shuffle=False)
x_train, x_test, y_train, y_test = data
x_holder = tf.placeholder(tf.float32, [None, seq_length, n_features])
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size, activation=tf.tanh)
outputs, _states = tf.nn.dynamic_rnn(cells, x_holder, dtype=tf.float32)
z = tf.contrib.layers.fully_connected(inputs=outputs[:, -1],
num_outputs=output_dim,
activation_fn=None)
loss = tf.reduce_mean((z - y_train) ** 2)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(100):
sess.run(train, {x_holder: x_train})
print(i, sess.run(loss, {x_holder: x_train}))
pred = sess.run(z, {x_holder: x_test})
plt.plot(y_test, 'r')
plt.plot(pred, 'g')
plt.show()
# rnn_stock_1()
rnn_stock_1()
print('\n\n\n\n\n\n\n')
import numpy as np
def rnn_1():
vocab = np.array(['e', 'n', 'o', 'r', 's', 't'])
x = [[0., 0., 0., 0., 0., 1.], # t
[1., 0., 0., 0., 0., 0.], # e
[0., 1., 0., 0., 0., 0.], # n
[0., 0., 0., 0., 1., 0.], # s
[0., 0., 1., 0., 0., 0.]] # o
y = [[1, 0, 0, 0, 0, 0], # e
[0, 1, 0, 0, 0, 0], # n
[0, 0, 0, 0, 1, 0], # s
[0, 0, 1, 0, 0, 0], # o
[0, 0, 0, 1, 0, 0]] # r
w = tf.Variable(tf.random_uniform([6,6]))
b = tf.Variable(tf.random_uniform([6]))
# (5,6) = (5,6) @ (6, 6)
z = tf.matmul(x, w) + b
h = tf.nn.softmax(z)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss=loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
print(sess.run(loss))
print(sess.run(h))
print(np.argmax(sess.run(h), axis=1 ))
print(vocab[np.argmax(sess.run(h), axis=1)])
rnn_1()
def rnn_2_2():
vocab = np.array(['e', 'n', 'o', 'r', 's', 't'])
# tenso
x = [[0., 0., 0., 0., 0, 1.],
[1., 0., 0., 0., 0, 0.],
[0., 1., 0., 0., 0, 0.],
[0., 0., 0., 0., 1, 0.],
[0., 0., 1., 0., 0, 0.]]
# ensor
y = [[1., 0., 0., 0., 0., 0.], # 클래스 6개. Dense 형태
[0., 1., 0., 0., 0., 0.],
[0., 0., 0., 0., 1., 0.],
[0., 0., 1., 0., 0., 0.],
[0., 0., 0., 1., 0., 0.]]
hidden_size = 2 # 전개량
x = np.float32([x]) # dynamic_rnn은 3차원 데이터를 입력으로 요구 함. x: 2차원 -> 3차원
# RNN layer
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, _states = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32)
print(outputs.shape)
# Softmax layer
w = tf.Variable(tf.random_uniform([hidden_size, 6]))
b = tf.Variable(tf.random_uniform([6]))
# y(5, 6) = x(5, hidden_size) @ w(hidden_size, 6) # w에서 x와 만나는 것은 피처의 개수. 뒤는 class의 개수
z = tf.matmul(outputs[0], w) + b
h = tf.nn.softmax(z)
loss_i = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
loss = tf.reduce_mean(loss_i)
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
pred = sess.run(z)
pred_arg = np.argmax(pred, axis=1)
print(i, sess.run(loss), pred_arg, ''.join(vocab[pred_arg])) #
print('-' * 50)
sess.close()
# one-hot encoding 유지한 상태에서 길이가 같은 여러 단어의 다음 글자 prediction
# 개인적으로 y(target, true value)도 one-hot encoding 유지한 것이 더 직관적이고 좋다.
# 아래 나오는 sparse_softmax_cross_entropy_with_logits와 sequence_loss는 과도하게 복잡하고, 최종 결과도 one-hot encoding된 결과이기 때문에, 불필요한 혼란 제거를 위해서는 그냥 y도 one-hot encoding 유지하는게 좋지 않을까.
# 강사님은 왜 굳이 sparse 형태를 사용하려 하는 것일까...
import tensorflow as tf
import numpy as np
from sklearn import preprocessing
def make_onehot(words):
data = list(''.join(words))
lb = preprocessing.LabelBinarizer().fit(data)
xx, yy = [], []
for text in words:
onehot = lb.transform(list(text))
xx.append(onehot[:-1])
yy.append([onehot[i] for i in range(1,6)])
return np.float32(xx), np.int32(yy), lb.classes_
def rnn_words(words):
hidden_size = 5
x, y, vocab = make_onehot(words)
batch_size, seq_length, n_classes = x.shape
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, states_ = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32)
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
h = tf.nn.softmax(z)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(100):
sess.run(train)
print(sess.run(loss))
result = vocab[np.argmax(sess.run(z), axis=2)]
print([''.join(x) for x in result])
rnn_words(['tensor', 'houses', 'yellow'])
# 가변길이 단어들 RNN 처리. y에 대해 one-hot encoding 유지.
import tensorflow as tf
import numpy as np
from sklearn import preprocessing
def make_onehot(words):
same_length_words = []
m = max([len(w) for w in words])
for w in words:
w += ' '*(m-len(w))
same_length_words.append(w)
data = list(''.join(words))
lb = preprocessing.LabelBinarizer().fit(data)
print(lb.classes_)
xx, yy = [], []
for w in same_length_words:
onehot = lb.transform(list(w))
print("onehot", onehot.shape)
# (10 = 최대 단어 길이 seq_length, 12개 알파벳 n_classes)
xx.append(onehot[:-1])
yy.append(onehot[1:])
return np.float32(xx), np.int32(yy), lb.classes_
def rnn_multi_words_different_length(words):
word_length = [len(w) for w in words]
x, y, vocab = make_onehot(words)
print("x.shape", x.shape, "y.shape", y.shape, "\n", vocab)
# (4, 10, 12) (4, 10, 12)
# ['a' 'c' 'e' 'f' 'h' 'l' 'n' 'o' 'r' 's' 't' 'w']
hidden_size = 8
batch_size, seq_length, n_classes = x.shape
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, state_ = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32,
sequence_length=word_length)
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
print("z.shape", z.shape)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=z, labels=y)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss=loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
print(sess.run(loss))
result = ([''.join(x) for x in vocab[np.argmax(sess.run(z) , axis=2)]])
for w, l in zip(result, word_length):
print("result ", w[:l-1])
rnn_multi_words_different_length (['sea', 'what', 'corw', 'tensorflow'])
# one-hot label 제거
def rnn_2_3():
vocab = np.array(['e', 'n', 'o', 'r', 's', 't'])
# tenso
x = [[0., 0., 0., 0., 0, 1.],
[1., 0., 0., 0., 0, 0.],
[0., 1., 0., 0., 0, 0.],
[0., 0., 0., 0., 1, 0.],
[0., 0., 1., 0., 0, 0.]]
# ensor
y = [0, 1, 4, 2, 3] # Sparse 형태
hidden_size = 5 # 전개량
x = np.float32([x])
# RNN layer
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, _states = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32)
print(outputs.shape)
# Softmax layer. Fully Connected Layer
w = tf.Variable(tf.random_uniform([hidden_size, 6]))
b = tf.Variable(tf.random_uniform([6]))
z = tf.matmul(outputs[0], w) + b
h = tf.nn.softmax(z)
# y(5, 6) = x(5, hidden_size) @ w(hidden_size, 6) # w에서 x와 만나는 것은 피처의 개수. 뒤는 class의 개수
loss_i = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=z, labels=y)
loss = tf.reduce_mean(loss_i)
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(1000):
sess.run(train)
pred = sess.run(z)
pred_arg = np.argmax(pred, axis=1)
print(i, sess.run(loss), pred_arg, ''.join(vocab[pred_arg])) #
print('-' * 50)
sess.close()
# Day_03_07_rnn_5_different.py
import numpy as np
import tensorflow as tf
from sklearn import preprocessing
def make_onehot_different(text_array):
data = list(''.join(text_array))
lb = preprocessing.LabelBinarizer().fit(data)
max_len = max([len(t) for t in text_array])
xx, yy = [], []
for text in text_array:
if len(text) < max_len:
text += '*' * (max_len - len(text))
onehot = lb.transform(list(text))
x = onehot[:-1]
x = np.float32(x)
y = onehot[1:]
y = np.argmax(y, axis=1)
xx.append(x)
yy.append(list(y))
return np.float32(xx), tf.constant(yy), lb.classes_
def rnn_5_different(text_array, iteration=100):
x, y, vocab = make_onehot_different(text_array)
batch_size, seq_length, n_classes = x.shape
hidden_size = 2
seq_len_array = [len(t) for t in text_array]
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size)
outputs, _states = tf.nn.dynamic_rnn(cells, x, dtype=tf.float32,
sequence_length=seq_len_array)
# xavier 초기화 사용
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
w = tf.ones([batch_size, seq_length])
loss = tf.contrib.seq2seq.sequence_loss(logits=z, targets=y, weights=w)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(iteration):
sess.run(train)
pred = sess.run(z)
pred_arg = np.argmax(pred, axis=2)
if i % 10 == 0:
print(i, end=' ')
for arg, length in zip(pred_arg, seq_len_array):
valid = vocab[arg]
print(''.join(valid[:length - 1]), end=' ')
print()
print('-' * 50)
print('ans :', text_array)
sess.close()
rnn_5_different(['tensor', 'sea', 'white'])
# Day_03_08_rnn_final.py
import numpy as np
import tensorflow as tf
from sklearn import preprocessing
def make_onehot_sentence(sentence, seq_length):
data = list(sentence)
lb = preprocessing.LabelBinarizer().fit(data)
print(lb.classes_)
onehot = lb.transform(data)
x = np.float32(onehot[:-1])
y = np.argmax(onehot[1:], axis=1)
# print(len(sentence))
# for i in range(len(sentence) - seq_length):
# print(i, i + seq_length, sentence[i:i+seq_length])
idx = [(i, i + seq_length) for i in range(len(sentence) - seq_length)]
xx = [x[s:e] for s, e in idx]
yy = [y[s:e] for s, e in idx]
print(*yy[:5], sep='\n')
return np.float32(xx), tf.constant(np.int32(yy)), lb.classes_
def rnn_final(sentence, iteration=100):
x, y, vocab = make_onehot_sentence(sentence, 20)
batch_size, seq_length, n_classes = x.shape
hidden_size = 7
cells = [tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size) for _ in range(2)]
multi = tf.nn.rnn_cell.MultiRNNCell(cells)
outputs, _states = tf.nn.dynamic_rnn(multi, x, dtype=tf.float32)
# xavier 초기화 사용
z = tf.contrib.layers.fully_connected(inputs=outputs,
num_outputs=n_classes,
activation_fn=None)
w = tf.ones([batch_size, seq_length])
loss = tf.contrib.seq2seq.sequence_loss(logits=z, targets=y, weights=w)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(iteration):
sess.run(train)
print(i, sess.run(loss))
print('-' * 50)
pred = sess.run(z)
print(pred.shape) # (151, 20, 25)
pred_arg = np.argmax(pred, axis=2)
print(pred_arg.shape) # (151, 20)
total = '*' + ''.join(vocab[pred_arg[0]])
for arg in pred_arg[1:]:
curr = ''.join(vocab[arg])
# print(curr)
total += curr[-1]
print('[{}]'.format(total))
print('[{}]'.format(sentence))
sess.close()
sentence = ("if you want to build a ship, "
"don't drum up people to collect wood and "
"don't assign them tasks and work, "
"but rather teach them to long for the endless immensity of the sea.")
rnn_final(sentence, 1000)
# if you want to build a ship
# ---------------
# 0 : if you want
# 1 : f you want
# 2 : you want t
# 3 : you want to
#
# 0 : if you want
# 1 : t to build
# Day_03_10_stock.py
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from sklearn import preprocessing, model_selection
np.set_printoptions(linewidth=1000)
def minmax_scale(data):
mx = np.max(data, axis=0)
mn = np.min(data, axis=0)
return (data - mn) / (mx - mn + 1e-7)
def rnn_stock_1():
stock = np.loadtxt('Data/stock_daily.csv',
delimiter=',', dtype=np.float32)
stock = stock[::-1]
# stock = minmax_scale(stock)
# stock = preprocessing.minmax_scale(stock)
scaler = preprocessing.MinMaxScaler().fit(stock)
stock = scaler.transform(stock)
print(stock[:3])
print(stock.shape, stock.dtype) # (732, 5) float32
x = stock
y = stock[:, -1:]
print(y.shape) # (732, 1)
seq_length = 7
n_features = 5 # n_classes
output_dim = 1
hidden_size = 10
x_data, y_data = [], []
for i in range(len(y) - seq_length):
x_data.append(x[i:i+seq_length]) # 0~6, 1~7
y_data.append(y[i+seq_length]) # 7 , 8
# print(x_data[-1], '->', y_data[-1])
x_data = np.float32(x_data)
y_data = np.float32(y_data)
print("SHAPE:", x_data.shape, y_data.shape)
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size, activation=tf.tanh)
outputs, _states = tf.nn.dynamic_rnn(cells, x_data, dtype=tf.float32)
print(outputs.shape, outputs[:, -1].shape)
z = tf.contrib.layers.fully_connected(inputs=outputs[:, -1],
num_outputs=output_dim,
activation_fn=None)
loss = tf.reduce_mean((z - y_data) ** 2)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(100):
sess.run(train)
print(i, sess.run(loss))
def rnn_stock_2():
stock = np.loadtxt('Data/stock_daily.csv',
delimiter=',', dtype=np.float32)
stock = stock[::-1]
# stock = minmax_scale(stock)
# stock = preprocessing.minmax_scale(stock)
scaler = preprocessing.MinMaxScaler().fit(stock)
stock = scaler.transform(stock)
print(stock[:3])
print(stock.shape, stock.dtype) # (732, 5) float32
x = stock
y = stock[:, -1:]
print(y.shape) # (732, 1)
seq_length = 7
n_features = 5 # n_classes
output_dim = 1
hidden_size = 10
x_data, y_data = [], []
for i in range(len(y) - seq_length):
x_data.append(x[i:i + seq_length]) # 0~6, 1~7
y_data.append(y[i + seq_length]) # 7 , 8
# print(x_data[-1], '->', y_data[-1])
x_data = np.float32(x_data)
y_data = np.float32(y_data)
# 75:25
data = model_selection.train_test_split(x_data, y_data, shuffle=False)
x_train, x_test, y_train, y_test = data
x_holder = tf.placeholder(tf.float32, [None, seq_length, n_features])
cells = tf.nn.rnn_cell.BasicRNNCell(num_units=hidden_size, activation=tf.tanh)
outputs, _states = tf.nn.dynamic_rnn(cells, x_holder, dtype=tf.float32)
z = tf.contrib.layers.fully_connected(inputs=outputs[:, -1],
num_outputs=output_dim,
activation_fn=None)
loss = tf.reduce_mean((z - y_train) ** 2)
optimizer = tf.train.AdamOptimizer(0.1)
train = optimizer.minimize(loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(100):
sess.run(train, {x_holder: x_train})
print(i, sess.run(loss, {x_holder: x_train}))
pred = sess.run(z, {x_holder: x_test})
plt.plot(y_test, 'r')
plt.plot(pred, 'g')
plt.show()
# rnn_stock_1()
rnn_stock_1()
print('\n\n\n\n\n\n\n')
2018년 10월 15일 월요일
numpy
import numpy as np
def python_basie():
a = np.array([[1, 2, 3], [3, 4, 5]])
print(a[1])
a, b = 3, 5
c = 3, 5 # packing
print(a, b, c)
d, e = c # unpacking
print(d, e)
def f(a, b, c):
print(a, b, c)
f(1,2, 3) # positional
f(a=3, b=2, c=1) # keyword
# [] : List : 중간중간 쉼표가 있음.
# [] : numpy.ndarray : 중간중간 쉼표가 없음.
# slicing. List와 ndarray 둘다 동작
def slicing():
a = range(10)
print(a)
a = list(a)
print(a)
print(a[0])
print(a[0:3]) # range()와 동일
print(a[-1])
print(a[3:])
print(a[0:int(len(a) / 2)])
print(a[int(len(a) / 2):])
print(a[0:5])
print(a[5:])
print(a[:5])
print(a[5:])
print(a[:]) # 전체
print(a[::2]) # 짝수 번째
print(a[1::2]) # 홀수 번째
print(a[-1::-1])
print(a[::-1])
def np_basic():
a = np.arange(6)
print(a)
print(a.shape, a.dtype, a.size, a.ndim)
b = np.reshape(a, [2,3])
print(b)
b = a.reshape([2,3]) # OK
print(b)
b = a.reshape((2,3)) # OK
print(b)
b = a.reshape(2,3) # OK python에서 튜플의 ()의 괄호는 생략 가능
print(b)
# broadcast 연산
print(a+1)
print(a*2)
print(a/2)
print(a>2)
print(a ** 2)
print(a[a>2]) # a> 2 항목만 뽑아 줌
print(np.sin(a)) # universal function
print(b+1)
print(b*2)
print(b/2)
print(b>2)
print(b ** 2)
print(b[b>2]) # a> 2 항목만 뽑아 줌
print(np.sin(b)) # universal function
print(a+a) # vector operation
print(b+b)
a1 = np.arange(3)
a2 = np.arange(6)
a3 = np.arange(3).reshape(1,3)
a4 = np.arange(3).reshape(3,1)
a5 = np.arange(6).reshape(2,3)
print("-"*50)
print("a2", a2)
print("a3", a3)
print("a4", a4)
print("a5", a5)
print("-"*50)
# scaler + vector ok : broadcast
# shape이 다른 vector끼리는 NG !
# 같은 shape 끼리는 연산 가능 OK ! : vector op
# 1개 짜리 vector는 scale로 간주함. : broadcast
# a1+a2 # ng : shape이 다르면 error
print(a1+a3) # ok : vector op
print(a1+a4) # ok : broadcast + broadcast
print(a1+a5) # ok : broadcast + vector op
#print(a2+a3)
#print(a2+a4)
#print(a2+a5)
#
print(a3+a4)
# print(a3+a5)
#
#print(a4+a5)
c = np.arange(12)
#아래 3가지는 모두 동일
print(c.reshape(3,4).shape)
print(c.reshape(-1,4).shape)
print(c.reshape(3,-1).shape)
d = c.reshape(3,4)
# 아래 2가지는 모두 동일
print(d.reshape(12))
print(d.reshape(-1))
np.random.seed(1)
e = np.random.randint(100, size=12).reshape(3,4)
print(e)
print(np.max(e))
print("axis=0", np.max(e, axis=0)) # 수직
print("axis=1", np.max(e, axis=1)) # 수평
print(np.argmax(e))
print("axis=0", np.argmax(e, axis=0)) # 수직에서 가장 큰 위치 알려 줌
print("axis=1", np.argmax(e, axis=1)) # 수평에서 가장 큰 위치 알려 줌
import numpy as np
import tensorflow as tf
a = np.arange(10)
print(a)
# 여러 항목을 한번에 꺼내는 방법
print(a[3], a[9])
# 는 아래와 동일하다. numpy는 "Index 배열" 지원 함.
print(a[[3, 9]]) # []를 하나더 해줘야 한다. 배열 행/열 index와 충돌 피하기 위해.
print("-"*50)
b = np.arange(12).reshape(3,4)
print(b)
print(b[0])
print(b[[0, -1]]) # 2개의 행데이터를 뽑아서 결합하는 방법. Index 배열 사용
print(b[::-1]) # 행 순서 뒤집기
print(b[-1, -1]) # 마지막 데이터.
# "Fancy index"이라 부름. 대괄호 안에 ,를 사용하는 문법. 차원마다 ,가 늘어남.
# 정수, slicing,
print(b[::-1, ::-1]) # 행과 열을 모두 뒤집기.
print(b.T) # transpose
# 문제 : 2차원 배열을 전치 형태로 출력하세요 (반복문, fancy index 사용)
for i in range(b.shape[-1]):
print(b[:, i])
# 행을 가져오는 것도 사실은 열 인덱싱을 생략한 것.
for i in range(b.shape[0]):
print(b[i, :])
def show_matmul(s1, s2):
a = tf.constant(s1, shape=s1)
b = tf.constant(s2, shape=s2)
result = tf.matmul(a, b).shape
print(" {} @ {} = {} ".format(s1, s2, result))
show_matmul([2,3], [3,2]) # 데이터가 아니라 shape을 넘긴 것임.
show_matmul([3,4], [4,2])
show_matmul([7, 2, 3], [7, 3, 2]) # tensorflow의 행렬 2차원까지밖에 없음.
# 앞의 7은 7번 2차원행렬 연산을 할거란 의미. 2, 3과 3, 2는 매칭되어야 함.
show_matmul([2, 7, 3], [2, 3, 2])
# np squeeze : 의미 없는 차원을 없애 줌. 1차원만 없앨 수 있음. 2차원은 못 없앰.
x = np.array([[[[[0], [1], [2]], [[3], [4], [5]]]]])
print(x)
print(x.shape)
print(x.squeeze(axis=0).shape)
print(x.squeeze(axis=1).shape)
print(x.squeeze(axis=4).shape)
print(x.squeeze().shape) # 모두 한꺼번에 제거
np.random.seed(123)
t = np.random.random_integers(0, 100, 24).reshape(2, 3, 4) # 0~100에서 24개 생성
print("\n")
print(t)
print("\n")
print(np.argmax(t, axis=0), end='\n\n') # 0차원 2개 끼리 비교
print(np.argmax(t, axis=1), end='\n\n') # 1차원 3개 끼리 비교
print(np.argmax(t, axis=2), end='\n\n') # 2차원 4개 끼리 비교
gradient descent
Single Variable
Multi Variables
x1, x2, x3는 각각 피처. w1, w2, w3는 각각의 비중.
y는 실측
import matplotlib.pyplot as plt x = [1, 2, 3] y = [1, 2, 3] # y = x : True function# h(x) = wx : Hyperthesisdef cost(x, y, w): sum = 0 for a, b in zip(x, y): sum += (w*a - b)**2 return sum/len(x) def show_cost(): print(cost(x, y, 0)) print(cost(x, y, 1)) print(cost(x, y, 2)) c = [] for w in range(-10, 11): c.append(cost(x, y, w)) print(c) plt.plot(range(-10, 11), c) plt.show() # 미분 : 기울기. 순간 변화량 # x축으로 1만큼 움직일 때, y축으로 움직인 거리 # y = x 1 = 1, 2 = 2, 3 = 3 # y = 2x 2 = 1, 4 = 2, 6 = 3 def gradient_descent(x, y, w): gd = 0 for a, b in zip(x, y): gd += (w*a - b)*a # MSE cost를 w에 대해 미분시 x값, 여기서는 a를 한번 더 곱해 주는 식이 된다. return gd/len(x) def update_gradient_descent(init=3, learning_rate=0.2, epoch=20): x = [1, 2, 3] y = [1, 2, 3] w = init for i in range(epoch): c = cost(x, y, w) g = gradient_descent(x, y, w) w -= learning_rate * g print(i, c) return w w = update_gradient_descent() print (5*w) print (7*w) # 문제 w를 1.0으로 만드는 방법. Hyper parameter# 1. step을 올린다.# 2. learning late를# 3. 초기값 # x가 5와 6일 대의 결과를 예측
Multi Variables
x1, x2, x3는 각각 피처. w1, w2, w3는 각각의 비중.
y는 실측
def loss(x1, x2, x3, y, w1, w2, w3): c = 0 for x_1, x_2, x_3, y_1 in zip(x1, x2, x3, y): c += (x_1*w1 + x_2*w2 + x_3*w3 - y_1) ** 2
return c / len(y) def update_gradient_decent_one(x, y, w): d = 0 for x_1, y_1 in zip(x, y): d += (x_1*w - y_1)*x_1 return d def update_gradient_decent(x1, x2, x3, y, w1, w2, w3): w1 = w1 - 0.2 * update_gradient_decent_one(x1, y, w1) w2 = w2 - 0.2 * update_gradient_decent_one(x2, y, w2) w3 = w2 - 0.2 * update_gradient_decent_one(x3, y, w3) return w1, w2, w3 x1 = [1, 4, 5, 8] # 공부 시간 x2 = [3, 4, 7, 9] # 출석 일 수 x3 = [4, 6, 1, 3] # 학습 열정 y = [3, 5, 4, 7] # 성적
w1 = 3 w2 = 4 w3 = 5
for i in range(50): w1, w2, w3 = update_gradient_decent(x1, x2, x3, y, w1, w2, w3) print(loss(x1, x2, x3, y, w1, w2, w3))
피드 구독하기:
글 (Atom)