Naver Project (Semiconductor_thin_film_layer_prediction_dnn_baseline)

Jacob Kim·2024년 1월 30일

Naver Project Week 2

목록 보기

6/20

반도체 박막 두께 분석

배경

최근 고사양 반도체 수요가 많아지면서 반도체를 수직으로 적층하는 3차원 공정이 많이 연구되고 있습니다. 반도체 박막을 수십 ~ 수백 층 쌓아 올리는 공정에서는 박막의 결함으로 인한 두께와 균일도가 저하되는 문제가 있습니다. 이는 소자 구조의 변형을 야기하며 성능 하락의 주요 요인이 됩니다. 이를 사전에 방지하기 위해서는 박막의 두께를 빠르면서도 정확히 측정하는 것이 중요합니다.

박막의 두께를 측정하기 위해 광스펙트럼 분석이 널리 사용되고 있습니다. 하지만 광 스펙트럼을 분석하기 위해서는 관련 지식을 많이 가진 전문가가 필요하며 분석과정에 많은 컴퓨팅자원이 필요합니다. 빅데이터 분석을 통해 이를 해결하고자 반도체 소자의 두께 분석 알고리즘 경진대회를 개최합니다.

배경 자료

반도체 박막은 얇은 반도체 막으로 박막의 종류와 두께는 반도체 소자의 특성을 결정짓는 중요한 요소 중 하나입니다. 박막의 두께를 측정하는 방법으로 반사율 측정이 널리 사용되며 반사율은 입사광 세기에 대한 반사광 세기의 비율로 정해집니다. (반사율 = 반사광/입사광) 반사율은 빛의 파장에 따라 변하며 파장에 따른 반사율의 분포를 반사율 스펙트럼이라고 합니다.

구조 설명

이번 대회에서 분석할 소자는 질화규소(layer_1)/이산화규소(layer_2)/질화규소(layer_3)/이산화규소(layer_4)/규소(기판) 총 5층 구조로 되어 있습니다. 대회의 목적은 기판인 규소를 제외한 layer_1 ~ layer_4의 두께를 예측하는 것으로 학습 데이터 파일에는 각 층의 두께와 반사율 스펙트럼이 포함되어 있습니다

데이터 설명

train.csv 파일에는 4층 박막의 두께와 파장에 따른 반사율 스펙트럼이 주어집니다. 헤더의 이름에 따라 layer_1 ~ 4는 해당 박막의 두께,

데이터값인 0 ~ 225은 빛의 파장에 해당하는 반사율이 됩니다. 헤더 이름인 0~225은 파장을 뜻하며 비식별화 처리가 되어있어 실제 값과는 다릅니다.

import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras import layers
import numpy as np
import matplotlib.pyplot as plt
import os

use_colab = True
assert use_colab in [True, False]

from google.colab import drive
drive.mount('/content/drive')

#Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).

데이터 로드

새로운 버전의 Colab파일에선 왼쪽의 폴더 tree에서 직접 드라이브 마운트를 진행해야합니다.

train_d = np.load("/content/drive/MyDrive/Dataset/semicon_train_data.npy")
test_d = np.load("/content/drive/MyDrive/Dataset/semicon_test_data.npy")

# train data path
# train_d = np.load("./semicon_train_data.npy")

# test data path
# train_d = np.load("./semicon_test_data.npy")
print(train_d.shape, test_d.shape)

(729000, 230) (81000, 230)

print(train_d[1].shape)

# (230,)

print(test_d[0].shape)

# (230,)

# Raw Data 확인
print(train_d[1000])
# print(test_d[0])

[7.00000000e+01 4.00000000e+01 1.50000000e+02 1.10000000e+02
 4.02068880e-01 4.05173400e-01 4.22107550e-01 4.35197980e-01
 4.66261860e-01 4.67785980e-01 4.76300870e-01 5.07203700e-01
 5.23296600e-01 5.31541170e-01 5.23319200e-01 5.32340650e-01
 5.38513840e-01 5.62478700e-01 5.74338900e-01 5.62083540e-01
 5.83467800e-01 5.88462350e-01 5.89606940e-01 5.93186560e-01
 5.93241930e-01 5.98241150e-01 5.87171000e-01 5.92055500e-01
 6.06527860e-01 5.87087870e-01 5.94492100e-01 5.73233500e-01
 5.94926300e-01 5.76777600e-01 5.70286300e-01 5.76913000e-01
 5.52754400e-01 5.61735150e-01 5.52500000e-01 5.23800200e-01
 5.11056840e-01 5.11732600e-01 4.92525460e-01 4.73373200e-01
 4.57444160e-01 4.53428800e-01 4.19270130e-01 3.87709920e-01
 3.87221520e-01 3.49848450e-01 3.16229370e-01 3.13858700e-01
 2.80085700e-01 2.61263000e-01 2.30250820e-01 1.98457850e-01
 1.53500440e-01 1.36026740e-01 1.08726785e-01 8.95344000e-02
 6.58975300e-02 5.09539300e-02 3.69604570e-02 2.88708470e-02
 1.54857090e-02 1.99594010e-02 2.33170990e-02 3.26487760e-02
 5.89796340e-02 6.33707600e-02 9.20634640e-02 1.24182590e-01
 1.63985540e-01 1.81501220e-01 2.16830660e-01 2.57668080e-01
 2.86428720e-01 3.21613600e-01 3.65077380e-01 3.85984180e-01
 4.33254400e-01 4.36655340e-01 4.60312520e-01 4.99771740e-01
 5.03849100e-01 5.48682200e-01 5.58029060e-01 5.62947630e-01
 5.99234340e-01 5.96399550e-01 6.04513000e-01 6.19093400e-01
 6.38402700e-01 6.64816900e-01 6.56996000e-01 6.66098600e-01
 6.69130500e-01 6.70661600e-01 6.80202360e-01 6.92821600e-01
 6.96852400e-01 7.02015160e-01 7.00507460e-01 7.00250800e-01
 7.11669300e-01 6.96838260e-01 7.08974360e-01 7.19114100e-01
 7.18034700e-01 7.00589540e-01 7.05873250e-01 6.92893500e-01
 7.14911160e-01 6.85250640e-01 6.82136830e-01 6.81311400e-01
 6.78019200e-01 6.67259900e-01 6.78185700e-01 6.52875660e-01
 6.34932940e-01 6.36031400e-01 6.16502100e-01 5.91461400e-01
 5.74673350e-01 5.57246200e-01 5.42328200e-01 5.19422200e-01
 5.03134850e-01 4.81939820e-01 4.45872100e-01 4.17206170e-01
 4.05306970e-01 3.55970500e-01 3.27808440e-01 2.89738060e-01
 2.44837760e-01 2.16494500e-01 1.69062380e-01 1.23260050e-01
 1.03297760e-01 6.77379500e-02 3.89807700e-02 2.52380610e-02
 2.59922780e-02 1.38827340e-02 1.90851740e-02 6.68724550e-02
 7.21453900e-02 1.14929430e-01 1.77696590e-01 2.11153720e-01
 2.55038900e-01 2.95599580e-01 3.21243700e-01 3.74151170e-01
 3.91486200e-01 4.27665320e-01 4.61769700e-01 4.75576340e-01
 4.78055980e-01 5.16503040e-01 5.30934450e-01 5.23902500e-01
 5.51259300e-01 5.65235300e-01 5.67701600e-01 5.70858800e-01
 5.76418300e-01 5.75021860e-01 6.01967200e-01 5.79546750e-01
 5.81157900e-01 5.93576100e-01 5.84070600e-01 6.04394850e-01
 5.87139800e-01 5.93731460e-01 5.91678200e-01 5.63344300e-01
 5.69886800e-01 5.46404960e-01 5.43537260e-01 5.42394760e-01
 5.42762940e-01 5.26218060e-01 4.89471520e-01 4.72608000e-01
 4.59845570e-01 4.39476280e-01 4.24489440e-01 4.16756120e-01
 4.11724180e-01 3.81418440e-01 3.60050380e-01 3.37973600e-01
 3.20442830e-01 2.97488420e-01 2.85940900e-01 2.82688900e-01
 2.78886770e-01 2.60735480e-01 2.64526520e-01 2.38374640e-01
 2.46523140e-01 2.54422630e-01 2.63443470e-01 2.77559100e-01
 3.05792480e-01 3.04755750e-01 3.47745950e-01 3.42683080e-01
 3.59969050e-01 3.90264000e-01 4.21746500e-01 4.33448700e-01
 4.50429620e-01 4.94950400e-01 4.90516330e-01 5.15429500e-01
 5.46570300e-01 5.64438460e-01 5.68110800e-01 5.71491600e-01
 5.81447660e-01 5.84012000e-01 6.05881900e-01 5.98874870e-01
 5.97313170e-01 5.95610300e-01]

데이터를 분석하여 모델에 학습할 수 있는 형태로 정리합니다.

train = train_d[:,4:] # 729000, 230 4번째 데이터부터 가져온다.
label = train_d[:,:4]

t_train = test_d[:,4:] # 81000, 230개 4번째 까지 데이터를 가져온다.
t_label = test_d[:,:4]

batch_size = 256

# for train
train_dataset = tf.data.Dataset.from_tensor_slices((train, label))
train_dataset = train_dataset.shuffle(10000).repeat().batch(batch_size=batch_size)
print(train_dataset)

# for test
test_dataset = tf.data.Dataset.from_tensor_slices((t_train, t_label))
test_dataset = test_dataset.batch(batch_size=batch_size)
print(test_dataset)

#<_BatchDataset element_spec=(TensorSpec(shape=(None, 226), dtype=tf.float64, name=None), TensorSpec(shape=(None, 4), dtype=tf.float64, name=None))>
#<_BatchDataset element_spec=(TensorSpec(shape=(None, 226), dtype=tf.float64, name=None), TensorSpec(shape=(None, 4), dtype=tf.float64, name=None))>

모델 구성

현재 가장 간단한 모델로 구성되어 있습니다.
데이터셋 자체를 가장 잘 학습할 수 있는 모델을 구현해 학습을 진행합니다.

8 * 8 * 4

#256

input_tensor = layers.Input(shape=(226,))

x = layers.Dense(8 * 8 * 4, activation='relu')(input_tensor)
x = layers.Reshape(target_shape=(8, 8, 4))(x)


output_tensor = layers.Dense(4)(x)

model = tf.keras.Model(input_tensor, output_tensor)

모델 학습

4개층의 박막의 두께를 예측하는 모델을 학습해봅시다.

# the save point
if use_colab:
    checkpoint_dir ='./drive/My Drive/train_ckpt/semiconductor/exp1'
    if not os.path.isdir(checkpoint_dir):
        os.makedirs(checkpoint_dir)
else:
    checkpoint_dir = 'semiconductor/exp1'

cp_callback = tf.keras.callbacks.ModelCheckpoint(checkpoint_dir,
                                                 save_weights_only=True,
                                                 monitor='val_loss',
                                                 mode='auto',
                                                 save_best_only=True,
                                                 verbose=1)

model.compile(loss='mae', #mse
              optimizer='adam',
              metrics=['mae']) #mse

history = model.fit(train_dataset,
                    steps_per_epoch=len(train) / batch_size, # train data의 길이 // batch 길이
                    epochs=100,
                    validation_data=test_dataset,
                    validation_steps=len(t_train) / batch_size)

Epoch 1/100
2847/2847 [==============================] - 44s 12ms/step - loss: 98.3696 - mae: 98.3696 - val_loss: 55.0984 - val_mae: 55.0984
Epoch 2/100
2847/2847 [==============================] - 33s 12ms/step - loss: 41.6744 - mae: 41.6744 - val_loss: 32.9359 - val_mae: 32.9359
Epoch 3/100
2847/2847 [==============================] - 35s 12ms/step - loss: 25.3576 - mae: 25.3576 - val_loss: 18.7295 - val_mae: 18.7295
Epoch 4/100
2847/2847 [==============================] - 33s 12ms/step - loss: 18.0173 - mae: 18.0173 - val_loss: 14.7011 - val_mae: 14.7011
Epoch 5/100
2847/2847 [==============================] - 33s 12ms/step - loss: 15.6218 - mae: 15.6218 - val_loss: 12.6688 - val_mae: 12.6688
Epoch 6/100
2847/2847 [==============================] - 33s 12ms/step - loss: 14.1532 - mae: 14.1532 - val_loss: 11.6533 - val_mae: 11.6533
Epoch 7/100
2847/2847 [==============================] - 33s 12ms/step - loss: 13.1076 - mae: 13.1076 - val_loss: 10.7831 - val_mae: 10.7831
Epoch 8/100
2847/2847 [==============================] - 34s 12ms/step - loss: 12.3764 - mae: 12.3764 - val_loss: 10.1933 - val_mae: 10.1933
Epoch 9/100
2847/2847 [==============================] - 33s 11ms/step - loss: 11.7947 - mae: 11.7947 - val_loss: 9.5747 - val_mae: 9.5747
Epoch 10/100
2847/2847 [==============================] - 34s 12ms/step - loss: 11.2987 - mae: 11.2987 - val_loss: 9.3836 - val_mae: 9.3836
Epoch 11/100
2847/2847 [==============================] - 33s 12ms/step - loss: 10.8654 - mae: 10.8654 - val_loss: 9.2192 - val_mae: 9.2192
Epoch 12/100
2847/2847 [==============================] - 33s 12ms/step - loss: 10.5522 - mae: 10.5522 - val_loss: 8.2976 - val_mae: 8.2976
Epoch 13/100
2847/2847 [==============================] - 32s 11ms/step - loss: 10.2725 - mae: 10.2725 - val_loss: 8.2708 - val_mae: 8.2708
Epoch 14/100
2847/2847 [==============================] - 34s 12ms/step - loss: 10.0172 - mae: 10.0172 - val_loss: 7.6924 - val_mae: 7.6924
Epoch 15/100
2847/2847 [==============================] - 33s 12ms/step - loss: 9.7965 - mae: 9.7965 - val_loss: 7.5464 - val_mae: 7.5464
Epoch 16/100
2847/2847 [==============================] - 33s 12ms/step - loss: 9.6332 - mae: 9.6332 - val_loss: 7.3972 - val_mae: 7.3972
Epoch 17/100
2847/2847 [==============================] - 33s 11ms/step - loss: 9.4276 - mae: 9.4276 - val_loss: 7.2775 - val_mae: 7.2775
Epoch 18/100
2847/2847 [==============================] - 32s 11ms/step - loss: 9.2707 - mae: 9.2707 - val_loss: 7.1137 - val_mae: 7.1137
Epoch 19/100
2847/2847 [==============================] - 33s 12ms/step - loss: 9.1206 - mae: 9.1206 - val_loss: 6.9142 - val_mae: 6.9142
Epoch 20/100
2847/2847 [==============================] - 34s 12ms/step - loss: 8.9705 - mae: 8.9705 - val_loss: 6.8319 - val_mae: 6.8319
Epoch 21/100
2847/2847 [==============================] - 35s 12ms/step - loss: 8.8938 - mae: 8.8938 - val_loss: 6.7131 - val_mae: 6.7131
Epoch 22/100
2847/2847 [==============================] - 34s 12ms/step - loss: 8.7869 - mae: 8.7869 - val_loss: 6.4702 - val_mae: 6.4702
Epoch 23/100
2847/2847 [==============================] - 36s 13ms/step - loss: 8.6596 - mae: 8.6596 - val_loss: 6.5114 - val_mae: 6.5114
Epoch 24/100
2847/2847 [==============================] - 37s 13ms/step - loss: 8.5861 - mae: 8.5861 - val_loss: 6.3033 - val_mae: 6.3033
Epoch 25/100
2847/2847 [==============================] - 33s 12ms/step - loss: 8.4791 - mae: 8.4791 - val_loss: 6.2558 - val_mae: 6.2558
Epoch 26/100
2847/2847 [==============================] - 35s 12ms/step - loss: 8.3915 - mae: 8.3915 - val_loss: 6.1372 - val_mae: 6.1372
Epoch 27/100
2847/2847 [==============================] - 34s 12ms/step - loss: 8.3137 - mae: 8.3137 - val_loss: 6.0292 - val_mae: 6.0292
Epoch 28/100
2847/2847 [==============================] - 33s 12ms/step - loss: 8.2411 - mae: 8.2411 - val_loss: 6.2980 - val_mae: 6.2980
Epoch 29/100
2847/2847 [==============================] - 32s 11ms/step - loss: 8.1478 - mae: 8.1478 - val_loss: 6.0103 - val_mae: 6.0103
Epoch 30/100
2847/2847 [==============================] - 32s 11ms/step - loss: 8.0947 - mae: 8.0947 - val_loss: 5.9317 - val_mae: 5.9317
Epoch 31/100
2847/2847 [==============================] - 33s 12ms/step - loss: 8.0531 - mae: 8.0531 - val_loss: 5.8171 - val_mae: 5.8171
Epoch 32/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.9609 - mae: 7.9609 - val_loss: 5.9632 - val_mae: 5.9632
Epoch 33/100
2847/2847 [==============================] - 34s 12ms/step - loss: 7.9161 - mae: 7.9161 - val_loss: 5.9289 - val_mae: 5.9289
Epoch 34/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.8580 - mae: 7.8580 - val_loss: 5.7152 - val_mae: 5.7152
Epoch 35/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.8196 - mae: 7.8196 - val_loss: 5.6733 - val_mae: 5.6733
Epoch 36/100
2847/2847 [==============================] - 33s 11ms/step - loss: 7.7775 - mae: 7.7775 - val_loss: 5.9246 - val_mae: 5.9246
Epoch 37/100
2847/2847 [==============================] - 35s 12ms/step - loss: 7.6940 - mae: 7.6940 - val_loss: 5.4900 - val_mae: 5.4900
Epoch 38/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.6567 - mae: 7.6567 - val_loss: 5.6425 - val_mae: 5.6425
Epoch 39/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.6151 - mae: 7.6151 - val_loss: 5.5744 - val_mae: 5.5744
Epoch 40/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.5781 - mae: 7.5781 - val_loss: 5.4983 - val_mae: 5.4983
Epoch 41/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.5228 - mae: 7.5228 - val_loss: 5.4053 - val_mae: 5.4053
Epoch 42/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.4723 - mae: 7.4723 - val_loss: 5.3742 - val_mae: 5.3742
Epoch 43/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.3971 - mae: 7.3971 - val_loss: 5.8738 - val_mae: 5.8738
Epoch 44/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.4046 - mae: 7.4046 - val_loss: 5.3615 - val_mae: 5.3615
Epoch 45/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.3576 - mae: 7.3576 - val_loss: 5.2890 - val_mae: 5.2890
Epoch 46/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.3018 - mae: 7.3018 - val_loss: 5.4128 - val_mae: 5.4128
Epoch 47/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.2977 - mae: 7.2977 - val_loss: 5.2234 - val_mae: 5.2234
Epoch 48/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.2586 - mae: 7.2586 - val_loss: 5.2990 - val_mae: 5.2990
Epoch 49/100
2847/2847 [==============================] - 34s 12ms/step - loss: 7.1975 - mae: 7.1975 - val_loss: 5.1710 - val_mae: 5.1710
Epoch 50/100
2847/2847 [==============================] - 32s 11ms/step - loss: 7.1891 - mae: 7.1891 - val_loss: 5.0942 - val_mae: 5.0942
Epoch 51/100
2847/2847 [==============================] - 34s 12ms/step - loss: 7.1484 - mae: 7.1484 - val_loss: 4.9883 - val_mae: 4.9883
Epoch 52/100
2847/2847 [==============================] - 33s 12ms/step - loss: 7.1019 - mae: 7.1019 - val_loss: 5.4390 - val_mae: 5.4390
Epoch 53/100
2847/2847 [==============================] - 35s 12ms/step - loss: 7.0700 - mae: 7.0700 - val_loss: 5.0882 - val_mae: 5.0882
Epoch 54/100
2847/2847 [==============================] - 34s 12ms/step - loss: 7.0510 - mae: 7.0510 - val_loss: 5.0578 - val_mae: 5.0578
Epoch 55/100
2847/2847 [==============================] - 35s 12ms/step - loss: 7.0202 - mae: 7.0202 - val_loss: 5.0349 - val_mae: 5.0349
Epoch 56/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.9821 - mae: 6.9821 - val_loss: 5.2390 - val_mae: 5.2390
Epoch 57/100
2847/2847 [==============================] - 34s 12ms/step - loss: 6.9828 - mae: 6.9828 - val_loss: 5.0327 - val_mae: 5.0327
Epoch 58/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.9384 - mae: 6.9384 - val_loss: 5.0759 - val_mae: 5.0759
Epoch 59/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.8780 - mae: 6.8780 - val_loss: 4.9260 - val_mae: 4.9260
Epoch 60/100
2847/2847 [==============================] - 37s 13ms/step - loss: 6.8704 - mae: 6.8704 - val_loss: 4.8593 - val_mae: 4.8593
Epoch 61/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.8608 - mae: 6.8608 - val_loss: 4.8831 - val_mae: 4.8831
Epoch 62/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.8275 - mae: 6.8275 - val_loss: 5.1727 - val_mae: 5.1727
Epoch 63/100
2847/2847 [==============================] - 34s 12ms/step - loss: 6.7859 - mae: 6.7859 - val_loss: 4.8546 - val_mae: 4.8546
Epoch 64/100
2847/2847 [==============================] - 33s 11ms/step - loss: 6.7657 - mae: 6.7657 - val_loss: 4.8685 - val_mae: 4.8685
Epoch 65/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.7401 - mae: 6.7401 - val_loss: 4.8768 - val_mae: 4.8768
Epoch 66/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.7265 - mae: 6.7265 - val_loss: 4.6834 - val_mae: 4.6834
Epoch 67/100
2847/2847 [==============================] - 33s 11ms/step - loss: 6.6693 - mae: 6.6693 - val_loss: 4.8329 - val_mae: 4.8329
Epoch 68/100
2847/2847 [==============================] - 34s 12ms/step - loss: 6.6614 - mae: 6.6614 - val_loss: 4.7936 - val_mae: 4.7936
Epoch 69/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.6552 - mae: 6.6552 - val_loss: 4.8079 - val_mae: 4.8079
Epoch 70/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.6337 - mae: 6.6337 - val_loss: 4.7778 - val_mae: 4.7778
Epoch 71/100
2847/2847 [==============================] - 33s 11ms/step - loss: 6.6061 - mae: 6.6061 - val_loss: 4.7467 - val_mae: 4.7467
Epoch 72/100
2847/2847 [==============================] - 34s 12ms/step - loss: 6.5800 - mae: 6.5800 - val_loss: 4.8474 - val_mae: 4.8474
Epoch 73/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.5879 - mae: 6.5879 - val_loss: 4.6442 - val_mae: 4.6442
Epoch 74/100
2847/2847 [==============================] - 33s 11ms/step - loss: 6.5656 - mae: 6.5656 - val_loss: 4.7916 - val_mae: 4.7916
Epoch 75/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.5286 - mae: 6.5286 - val_loss: 4.5983 - val_mae: 4.5983
Epoch 76/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.4862 - mae: 6.4862 - val_loss: 4.6008 - val_mae: 4.6008
Epoch 77/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.4822 - mae: 6.4822 - val_loss: 4.5863 - val_mae: 4.5863
Epoch 78/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.4512 - mae: 6.4512 - val_loss: 4.6937 - val_mae: 4.6937
Epoch 79/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.4517 - mae: 6.4517 - val_loss: 4.7971 - val_mae: 4.7971
Epoch 80/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.4411 - mae: 6.4411 - val_loss: 4.6173 - val_mae: 4.6173
Epoch 81/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.4007 - mae: 6.4007 - val_loss: 4.6712 - val_mae: 4.6712
Epoch 82/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.3860 - mae: 6.3860 - val_loss: 4.5905 - val_mae: 4.5905
Epoch 83/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.3750 - mae: 6.3750 - val_loss: 4.5082 - val_mae: 4.5082
Epoch 84/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.3653 - mae: 6.3653 - val_loss: 4.6231 - val_mae: 4.6231
Epoch 85/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.3490 - mae: 6.3490 - val_loss: 4.5700 - val_mae: 4.5700
Epoch 86/100
2847/2847 [==============================] - 35s 12ms/step - loss: 6.3441 - mae: 6.3441 - val_loss: 4.5706 - val_mae: 4.5706
Epoch 87/100
2847/2847 [==============================] - 32s 11ms/step - loss: 6.3160 - mae: 6.3160 - val_loss: 4.4801 - val_mae: 4.4801
Epoch 88/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.2917 - mae: 6.2917 - val_loss: 4.4492 - val_mae: 4.4492
Epoch 89/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.2673 - mae: 6.2673 - val_loss: 4.4364 - val_mae: 4.4364
Epoch 90/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.2582 - mae: 6.2582 - val_loss: 4.5396 - val_mae: 4.5396
Epoch 91/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.2437 - mae: 6.2437 - val_loss: 4.4340 - val_mae: 4.4340
Epoch 92/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.2344 - mae: 6.2344 - val_loss: 4.5570 - val_mae: 4.5570
Epoch 93/100
2847/2847 [==============================] - 34s 12ms/step - loss: 6.2096 - mae: 6.2096 - val_loss: 4.5309 - val_mae: 4.5309
Epoch 94/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.2364 - mae: 6.2364 - val_loss: 4.4046 - val_mae: 4.4046
Epoch 95/100
2847/2847 [==============================] - 34s 12ms/step - loss: 6.1718 - mae: 6.1718 - val_loss: 4.4525 - val_mae: 4.4525
Epoch 96/100
2847/2847 [==============================] - 33s 11ms/step - loss: 6.1778 - mae: 6.1778 - val_loss: 4.4234 - val_mae: 4.4234
Epoch 97/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.1731 - mae: 6.1731 - val_loss: 4.4228 - val_mae: 4.4228
Epoch 98/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.1397 - mae: 6.1397 - val_loss: 4.3904 - val_mae: 4.3904
Epoch 99/100
2847/2847 [==============================] - 33s 12ms/step - loss: 6.1304 - mae: 6.1304 - val_loss: 4.4282 - val_mae: 4.4282
Epoch 100/100
2847/2847 [==============================] - 34s 12ms/step - loss: 6.1343 - mae: 6.1343 - val_loss: 4.4304 - val_mae: 4.4304

# np.save("file/path.npy", history.numpy())
# history = np.load("histoy.npy", allow_pickle=True)

loss=history.history['mae'] # mse
val_loss=history.history['val_mae'] # val_mse

epochs_range = range(len(loss))

plt.figure(figsize=(8, 4))
plt.plot(epochs_range, loss, label='Training MAE') # MSE
plt.plot(epochs_range, val_loss, label='Validation MAE') # MSE
plt.legend(loc='upper right')
plt.title('Training and Validation MAE') # MSE
plt.show()

모델 평가

Mean absolute error 를 이용해 모델이 정확히 예측하는지를 확인

# model.load_weights(checkpoint_dir)
results = model.evaluate(test_dataset)
print("MAE :", results[0]) # MAE

#317/317 [==============================] - 1s 4ms/step - loss: 4.4304 - mae: 4.4304
#MAE : 4.430441379547119

Jacob Kim

AI, Information and Communication, Electronics, Computer Science, Bio, Algorithms