Google tensorflow certification 09

February 17, 2022 2 분 소요

Google Tensorflow Certification 09

Category 5 - 시퀀스와 시계열 (Sequence, Time Series) - 1

Windowd Dataset

시계열 데이터에서 데이터를 예측하기 위해 참고할 데이터를 세팅하는 것

window_size, shift, drop_remainder # 실습 참조

Conv1D(Convolution Neural Network 1-Dimention)

효율적인 연산을 위해 특성을 추출하는 필터

kernel_size, strides, padding # 실습 참조

Optimizer 튜닝

기본적인 최적화(Optimizer) 알고리즘인 SGD, ADAM 등을 튜닝할 수 있다.

momentum, Learning Rate 등.. # 실습 참조

Huber Loss

MSE와 MAE 중에 낮은 값을 따라가는 오차 함수

실습(Sunspots)

Step 1. Import

import csv # csv(comma separated value) import
import urllib
import tensorflow_datasets as tfds 
import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Dense, LSTM, Lambda, Conv1D
from tensorflow.keras.models import Sequential 
from tensorflow.keras.callbacks import ModelCheckpoint 

from tensorflow.keras.optimizers import SGD 
from tensorflow.keras.losses import Huber

Step 2. Preprocessing

# 전처리할 데이터 로드
url = 'https://storage.googleapis.com/download.tensorflow.org/data/Sunspots.csv'
urllib.request.urlretrieve(url, 'sunspots.csv')

with open('sunspots.csv') as csvfile:
    reader = csv.reader(csvfile, delimiter=',')
    next(reader)
    i = 0
    for row in reader:
        print(row)
        i+=1
        if i > 10:
            break

전처리(preprocessing) 요구 조건

sunspots(x), time_step(y)를 분할

# 빈 리스트에 분할시키기
sunspots = []
time_step = []
with open('sunspots.csv') as csvfile:
    reader = csv.reader(csvfile, delimiter=',')
    next(reader) # 첫 줄은 header이므로 skip.
    for row in reader:
        sunspots.append(float(row[2]))
        time_step.append(int(row[0]))
        
# List to np.ndarray
series = np.array(sunspots)
time = np.array(time_step)

# Train data set / Validation data set 분할
split_time = 3000

time_train = time[:split_time]
time_valid = time[split_time:]

x_train = series[:split_time]
x_valid = series[split_time:]

# Window Dataset Setting
window_size=30
batch_size = 32
shuffle_size = 1000

def windowed_dataset(series, window_size, batch_size, shuffle_buffer):
    series = tf.expand_dims(series, axis=-1)
    ds = tf.data.Dataset.from_tensor_slices(series)
    ds = ds.window(window_size + 1, shift=1, drop_remainder=True)
    ds = ds.flat_map(lambda w: w.batch(window_size + 1))
    ds = ds.shuffle(shuffle_buffer)
    ds = ds.map(lambda w: (w[:-1], w[1:]))
    return ds.batch(batch_size).prefetch(1)

# window dataset 적용
train_set = windowed_dataset(x_train, 
                             window_size=window_size, 
                             batch_size=batch_size,
                             shuffle_buffer=shuffle_size)

validation_set = windowed_dataset(x_valid, 
                                  window_size=window_size,
                                  batch_size=batch_size,
                                  shuffle_buffer=shuffle_size)

Step 3. Modeling

model = Sequential([
    # Conv1D(몇 개의 데이터를 가지고 특성을 추출할건지(커널사이즈), padding = causal(사이즈 그대로))
    # input_shape에 None = 어떤 모양이 오든 무시
    Conv1D(60, kernel_size=5, padding="causal", activation="relu",input_shape=[None, 1]),
    LSTM(60, return_sequences=True),
    LSTM(60, return_sequences=True), # many to many 기법 쓰기 때문에 return)sequence = True
    Dense(30, activation="relu"),
    Dense(10, activation="relu"),
    Dense(1),
    Lambda(lambda x: x * 400) #Lambda Layer
])

#####

Step 4. Compile

# SGD Custom
optimizer = SGD(lr=0.0001, momentum=0.9) # lr = 학습률, momentum = 관성(가중치)
# Huber
loss= Huber()
# Compile
model.compile(loss=loss, optimizer=optimizer, metrics=["mae"])

Step 4.5. ModelCheckpoint 생성

checkpoint_path = "my_checkpoint.ckpt" # 체크포인트 위치는 로컬, 이름.ckpt or 이름.m5
checkpoint = ModelCheckpoint(filepath=checkpoint_path, 
                             save_weights_only=True, # 가중치만 저장
                             save_best_only=True, # 가장 좋은 결과만 저장
                             monitor='val_mae',  # 기준 = 'validation_MAE가 가장 낮은 것'
                             verbose=1) # 출력

Step 5. Fit

# 학습(train data, Validation_data, epochs, callbacks[ckpt])
epochs=100
history = model.fit(train_set, 
                    validation_data=(validation_set), 
                    epochs=epochs, 
                    callbacks=[checkpoint],
                   )

Step 5.5. Ckpt Load Weight

# 이 코드가 없다면, Ckpt 만드는 이유가 없음(가중치 저장만 해두고 사용 안하는 것)
model.load_weights(checkpoint_path)

Twitter Facebook LinkedIn