Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
AI勉強会_Kerasハンズオン#1_分類
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
taashi
April 17, 2019
Technology
140
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
AI勉強会_Kerasハンズオン#1_分類
社内勉強会資料
Keras
Google Colaboratory
(修正中...)
- precision, recallの説明追加
taashi
April 17, 2019
Transcript
45 slides
AI #1 (classification)
Keras
$# Google% Jupyter notebook$# Google Colaboratory (https://colab.research.google.com/)
12! " GPU, TPU! Google Drive &! … etc
Google Colab PYTHON Notebook
Colab!
(%"# ) : ! :
$ ! Notebook
$" !% GPU% #
GPU
%# & ’shift + enter’ "$ !
$’!’ " Linux !# Linux
ColabDrive&+ Google Drive! drive" Colb)' drive"%* from
google.colab import drive drive.mount('/content/drive') "( google.colabdrive ! Drive.mount$# Google Drive!
' !# URL (& ) Google
Drive% "$ (1)
$ # %! & ' " (’enter’ Google
Drive (2)
E@$#%"! <; *(7- ?.=384 :E@ 96 <;7 <;&'!Slack>
1, PLAYGROUND / 0C D5)+A - $# -2B
,89 Team Drives*(@2)/$#!7 "' 10Team Drives$#!5< (3:-
…) # [001] Install use items &% !.+ Team Drives $#!"' .+46 =; >? # [002] Mount GoogleDrive # [003] Copy data from google drive # [004] Check directory structure
Keras")* $&! '&
% $ )*#( )* $
:'<$ /1 !% '<9 '<7 *)0
(C F ) '<!% '<93" !5 '<9> (,?) '< 8) CF'<9A'< D+('24)(6. 8) FBC;&C'<-= (@+ #E)
Keras PythonA 8<BNNW0!,01 +($"4*TensorFlow, CNTK, Theano ; ID-3)'!- @
CPUGPU6%.2&;@ … etc :>A =?7#&&/ (=?7 ) 59 (JH) E : =?7 GCF
#$ # [005] Define constant value # [006] Make data
file list # [007] Check dog data # [008] Check cat data " ! " ‘_t_XXX.png’ &(Train) ‘_v_XXX.png’ (Validation) ‘_e_XXX.png’ %(Evaluate)
+.0 )C57 ,? %98 /- %98%:=>E04 @) H*(:[1, 0],
G*(:[0, 1] ,? ;F %98G or H OneHotVector )CA("$%&#!),? 6 2 %98'.(B1D ) numpy>E04<3
!#'% KerasOneHotVactor -*' &#$ import keras.utils as ku class_ids
= np.array([1, 0, 1, 1, 2, 0]) one_hots = ku.to_categorical(class_ids) print(one_hots) > [[0, 1, 0], [1, 0, 0], [0, 1, 0], [0, 0, 1], [1, 0, 0]] # [009] Make input and teacher data " (,cv2imread Numpy)+ OK
!$ Keras!$ 1) Sequential " 2) functional API# "
Keras!$ " 2) &( " 2)functional API# "'%
#&) % functional API$ % input_layer = Input((64,
64, 1)) cv_1_layer = Conv2D(32, 3)(input_layer) cv_2_layer = Conv2D(64, 3)(cv_1_layer) max_layer = MaxPooling2D()(cv_2_layer) flat_layer = Flatten()(max_layer) output_layer = Dense(2)(flat_layer) *'(" Keras !
Input&! input_layer = Input((64, 64, 1))
( Input(input_shape) #% input_shape ' $ "
Convolution!Conv2D1) Convolution! cv_1_layer = Conv2D(32, 3)(input_layer) 4 Conv2D(filters, kernel_size,
strides=(1, 1), padding=‘valid’, activation=None) *,- filters ( ) (30/%.-) kernel_size - 5 #"+% strides - 5 #"+% padding’valid’padding3 ’same’ 2('&$ activation
MaxPoolingMaxPooling2D.& MaxPooling max_layer = MaxPooling2D()(cv_layer) 3 MaxPooling2D(pool_size=(2, 2), strides=None)
')* pool_size *4 " (# strides *4 " (# ,%2+$ . /5610-!
Flatten# flat_layer = Flatten()(max_layer) % Flatten()
"! $&
Dense output_layer = Dense(2)(flat_layer)
Dense(units, activation=None) units activation
! Activate act_layer = Activate(activation='relu’)(cv_layer)
" Activate(activation=None) activation
;5F(%'- #! (%'- #!3D CB 14,"* (Input)!G.E7<,"* 9:input_shape==3A8 CB
@?& $!6>/ 0/2 & $!)+%'( @?)A8 3D # [010] Define Network
input_layer = Input(IMAGE_SHAPE) cv1_1 = Conv2D(32, 3, padding='same', input_shape=IMAGE_SHAPE)(input_layer)
cv1_1 = Activation(activation='relu')(cv1_1) cv1_2 = Conv2D(32, 3, padding='same')(cv1_1) cv1_2 = Activation(activation='relu')(cv1_2) cv1_max = MaxPooling2D()(cv1_2) cv2_1 = Conv2D(64, 3, padding='same')(cv1_max) cv2_1 = Activation(activation='relu')(cv2_1) cv2_2 = Conv2D(64, 3, padding='same')(cv2_1) cv2_2 = Activation(activation='relu')(cv2_2) cv2_max = MaxPooling2D()(cv2_2) flat_layer = Flatten()(cv2_max) fc = Dense(2)(flat_layer) output_layer = Activation(activation='softmax')(fc)
" Model# Keras model = Model(inputs=[input_layer],
outputs=[output_layer]) Model ! inputs ( ) outputs ( )( ) % $
()'# ! & '# model.summary() +*$ from keras.utils
import plot_model plot_model(model, to_file=MODEL_PNG_NAME) to_file " dot ()%
!- !- model.compile(optimizer=Adam(), loss='categorical_crossentropy',
metrics=['accuracy']) " %& optimizer loss metrics " [’accuracy’]( $ ,#' +*& crossentropy)
fit history = model.fit(train_inputs, train_teachers, batch_size=20,
epochs=100 , validation_data=(valid_inputs, valid_teachers) , shuffle=True, verbose=1, callbacks=callbacks) train_inputs$*#! train_teachers$*#') batch_size epochs % validation_data&(# (!, ')) shuffle! verbose$*!" (1 ) callbacks %()
+ "& + model.save_weights(save_file_name) $+ & ' )*%#(!
model.save(save_file_name) )* $+ & , &
+ # [012] Train
!*Adam )'$categorical_crossentropy & % (# +",
+./, # [013] Check loss and acc '
&#(3loss41acc41$ )% loss((3loss) % (3 )% val loss(*-loss) 5% "! 0(32
0, 0,&"#$ +( "#$ from keras.models import load_model model
= load_model(model_file) !%-. '/ $ model = Model(inputs=[input_layer], outputs=[output_layer]) model.load_weights(model_file) )'/ $ "#-.* $
predict model.predict(predict_inputs) predict_inputs (
) # [014] Load trained model # [015] Predict (Train data) # [016] Predict (Valid data) # [017] Predict (Evaluate data)
'43 51$ (-". 0 '43 , #!///+
)*//&2 1% //&2
,+#%3417 #%3417 0.68 !'" 5/ 2 3+-4+-Convolution&$* Dropout0.55/ .6
,+0.6 ()
Dropout drop_layer =
Dropout()(cv_layer) Dropout(rate) rate
;JIL0'$)8H "&*<F'$)G 8H"&*BK Keras Callback(!:= 64A;J-.
A;J1 lossE N28H</ Callback;J+C%#M(O@3 %#M) ) D97, ;J(fit)5?callbacls>?@3
# " ModelCheckpoint " import keras.callbacks as kc kc.ModelCheckpoint(
filepath=MODEL_FILE_NAME, save_weights_only=True , save_best_only=True, period=1) filepath ( ) save_weights_only save_best_only period ! val_loss!
AutoEncoder etc
…
Keras
More Decks by taashi
See All by taashi
What is Haskell?
taashi
0
73
数学の世界~フラクタル~(社内勉強会1002)
taashi
0
190
論文LT会_論文紹介(CU-Net)
taashi
1
650
始めようElmでフロント開発_その02_ローカルWebアプリ
taashi
0
150
始めようElmでフロント開発_その01_Elmの基礎
taashi
0
150
Other Decks in Technology
See All in Technology
お風呂場をサーバルームにしよう
yusuke427
0
360
小さなエージェントから始めてみよう!実践コンテキストエンジニアリング
iwamot
PRO
0
130
AIネイティブ組織の作り方。1人のPOとAIから始める、自社での実験
ama_ch
2
1.2k
_マネジメントする人される人ナイト_2026__良いチームを作っていきたい_お気持ち_.pdf
xkxaxkx
0
810
パスワードのない鍵で、 データを手元に置く
ginjihayashi
0
370
目を覚ませ!僕らの認証情報が何者かに侵略されてるぞ
tkito
2
470
「任されビリティ」で引き受けて「よしなりてぃ」に頼らず任せる / Accept tasks with "makasarebility" and hand them over without relying on "yoshinarily"
pauli
1
800
市販IoT機器や自作IoT機器とHomeAssistantを駆使して家の電気代をできる限り見える化する方法の考案
harutiro
0
390
あの日、見つかっていなかったらどうなっていたか
kazzpapa3
2
260
マネジメントされる側の仕事は「文句」を言うこと
riyaamemiya
7
5.8k
OpenTelemetryで変わった 新時代のCloudWatchメトリクス監視
tsukuboshi
0
450
Terraform 0 to 100:100個以上のリポジトリを常に最新にし続けるための自動化
sue445
2
130
Featured
See All Featured
The Art of Programming - Codeland 2020
erikaheidi
57
15k
What the history of the web can teach us about the future of AI
inesmontani
PRO
3
820
Leading Effective Engineering Teams in the AI Era
addyosmani
9
2.9k
The Hidden Cost of Media on the Web [PixelPalooza 2025]
tammyeverts
2
580
So, you think you're a good person
axbom
PRO
2
2.3k
AI: The stuff that nobody shows you
jnunemaker
PRO
12
1.3k
The AI Revolution Will Not Be Monopolized: How open-source beats economies of scale, even for LLMs
inesmontani
PRO
3
3.8k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
16
2.2k
Winning Ecommerce Organic Search in an AI Era - #searchnstuff2025
aleyda
2
2.2k
How STYLIGHT went responsive
nonsquared
100
6.4k
Rebuilding a faster, lazier Slack
samanthasiow
85
9.8k
We Have a Design System, Now What?
morganepeng
55
8.4k