AutoML / mnist / keras

인터넷에 찾아보니 예제들이 있어 그대로 하니 잘 안되어서

 

1. 설치 버전

 1) Python : 3.7.6

 2) Tensorflow : 2.1.0

 3) Keras : 2.3.1

 

2. 소스

from keras.datasets import mnist
from autokeras import ImageClassifier

(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.reshape(x_train.shape + (1,)) # (1,) denotes the channles which is 1 in this case
x_test = x_test.reshape(x_test.shape + (1,)) # (1,) denotes the channles which is 1 in this case

clf = ImageClassifier()
clf.fit(x_train, y_train)
(x_train, y_train), (x_test, y_test) = mnist.load_data()

 

3. 결과

Train for 1500 steps, validate for 375 steps
Epoch 1/1000
1500/1500 [==============================] - 52s 35ms/step - loss: 0.1715 - accuracy: 0.9475 - val_loss: 0.0710 - val_accuracy: 0.9792
Epoch 2/1000
................................

1500/1500 [==============================] - 51s 34ms/step - loss: 0.0204 - accuracy: 0.9927 - val_loss: 0.0414 - val_accuracy: 0.9900
Epoch 20/1000
1500/1500 [==============================] - 51s 34ms/step - loss: 0.0217 - accuracy: 0.9927 - val_loss: 0.0417 - val_accuracy: 0.9893

 

 

[Trial complete]
[Trial summary]
 |-Trial ID: e9fea079a980756cf5cb43345c2e712e
 |-Score: 0.038674127771691325
 |-Best step: 0
 > Hyperparameters:
 |-classification_head_1/dropout_rate: 0.5
 |-classification_head_1/spatial_reduction_1/reduction_type: flatten
 |-dense_block_1/dropout_rate: 0
 |-dense_block_1/num_layers: 1
 |-dense_block_1/units_0: 128
 |-dense_block_1/use_batchnorm: False
 |-image_block_1/augment: False
 |-image_block_1/block_type: vanilla
 |-image_block_1/conv_block_1/dropout_rate: 0.25
 |-image_block_1/conv_block_1/filters_0_0: 32
 |-image_block_1/conv_block_1/filters_0_1: 64
 |-image_block_1/conv_block_1/kernel_size: 3
 |-image_block_1/conv_block_1/max_pooling: True
 |-image_block_1/conv_block_1/num_blocks: 1
 |-image_block_1/conv_block_1/num_layers: 2
 |-image_block_1/conv_block_1/separable: False
 |-image_block_1/normalize: True
 |-optimizer: adam
Train for 1500 steps, validate for 375 steps

Epoch 1/1000
1500/1500 [==============================] - 1521s 1s/step - loss: 0.2227 - accuracy: 0.9381 - val_loss: 0.1432 - val_accuracy: 0.9625
Epoch 2/1000
1500/1500 [==============================] - 1633s 1s/step - loss: 0.1563 - accuracy: 0.9605 - val_loss: 0.0721 - val_accuracy: 0.9795
................................................

1500/1500 [==============================] - 1509s 1s/step - loss: 0.0060 - accuracy: 0.9982 - val_loss: 0.0629 - val_accuracy: 0.9877
Epoch 36/1000
1500/1500 [==============================] - 1518s 1s/step - loss: 0.0046 - accuracy: 0.9989 - val_loss: 0.0470 - val_accuracy: 0.9912
[Trial complete]
[Trial summary]
 |-Trial ID: 7fd78984640e35c2dbe0c037d38a266b
 |-Score: 0.039056800348457425
 |-Best step: 0
 > Hyperparameters:
 |-classification_head_1/dropout_rate: 0
 |-dense_block_1/dropout_rate: 0
 |-dense_block_1/num_layers: 2
 |-dense_block_1/units_0: 32
 |-dense_block_1/units_1: 32
 |-dense_block_1/use_batchnorm: False
 |-image_block_1/augment: True
 |-image_block_1/block_type: resnet
 |-image_block_1/normalize: True
 |-image_block_1/res_net_block_1/conv3_depth: 4
 |-image_block_1/res_net_block_1/conv4_depth: 6
 |-image_block_1/res_net_block_1/pooling: avg
 |-image_block_1/res_net_block_1/version: v2
 |-optimizer: adam
Train for 1500 steps, validate for 375 steps
Epoch 1/1000
1500/1500 [==============================] - 53s 35ms/step - loss: 0.1769 - accuracy: 0.9466 - val_loss: 0.0677 - val_accuracy: 0.9808
Epoch 2/1000
1500/1500 [==============================] - 52s 35ms/step - loss: 0.0826 - accuracy: 0.9742 - val_loss: 0.0527 - val_accuracy: 0.9848

................................................

1500/1500 [==============================] - 52s 35ms/step - loss: 0.0199 - accuracy: 0.9934 - val_loss: 0.0477 - val_accuracy: 0.9903
Epoch 27/1000
1500/1500 [==============================] - 51s 34ms/step - loss: 0.0163 - accuracy: 0.9944 - val_loss: 0.0447 - val_accuracy: 0.9910
[Trial complete]
[Trial summary]
 |-Trial ID: 50634a5647ede3546f77215692e42491
 |-Score: 0.039515967257394854
 |-Best step: 0
 > Hyperparameters:
 |-classification_head_1/dropout_rate: 0.5
 |-classification_head_1/spatial_reduction_1/reduction_type: flatten
 |-dense_block_1/dropout_rate: 0
 |-dense_block_1/num_layers: 1
 |-dense_block_1/units_0: 128
 |-dense_block_1/use_batchnorm: False
 |-image_block_1/augment: False
 |-image_block_1/block_type: vanilla
 |-image_block_1/conv_block_1/dropout_rate: 0.25
 |-image_block_1/conv_block_1/filters_0_0: 32
 |-image_block_1/conv_block_1/filters_0_1: 64
 |-image_block_1/conv_block_1/kernel_size: 3
 |-image_block_1/conv_block_1/max_pooling: True
 |-image_block_1/conv_block_1/num_blocks: 1
 |-image_block_1/conv_block_1/num_layers: 2
 |-image_block_1/conv_block_1/separable: False
 |-image_block_1/normalize: True
 |-optimizer: adam
INFO:tensorflow:Oracle triggered exit
Train for 1875 steps, validate for 375 steps
Epoch 1/1000
1875/1875 [==============================] - 65s 35ms/step - loss: 0.1580 - accuracy: 0.9514 - val_loss: 0.0500 - val_accuracy: 0.9843
Epoch 2/1000
1875/1875 [==============================] - 65s 35ms/step - loss: 0.0742 - accuracy: 0.9773 - val_loss: 0.0258 - val_accuracy: 0.9927

................................................

Epoch 41/1000
1875/1875 [==============================] - 64s 34ms/step - loss: 0.0151 - accuracy: 0.9955 - val_loss: 0.0019 - val_accuracy: 0.9996
Epoch 42/1000
1875/1875 [==============================] - 64s 34ms/step - loss: 0.0153 - accuracy: 0.9952 - val_loss: 0.0018 - val_accuracy: 0.9993