[AAIS] E04. Transfer Learning - 1
2021-01-27 # AAIS # AI # AAIS # keras

Transfer Learning


About VGG16

https://analysisbugs.tistory.com/65?category=839091

https://bskyvision.com/504

https://medium.com/@msmapark2/vgg16-%EB%85%BC%EB%AC%B8-%EB%A6%AC%EB%B7%B0-very-deep-convolutional-networks-for-large-scale-image-recognition-6f748235242a

이전의 모델들은 대부분 7x7 filter, 11x11 filter를 사용하였으나 이부분부터 3x3 filter를 사용하였습니다.

예시로 10x10을 7x7 filter를 통과시키면 feature map이 4x4가 만들어지는데 필요한 파라미터 수는 7x7 = 49이고
10x10모델을 3x3 filter를 한번통과시키면 8x8 map > 두번통과시키면 6x6 map > 세번통과시키면 4x4 featur map이 만들어지는데 필요한 파라미터수는 3x3x3 = 27로 크게 줄어듭니다.



Q.이렇게 할 경우 왜 좋은가? : 3x3일경우 패딩이 1이라서 경계선의 추론이 더 쉬워진다? + relu를 세번통과하기 때문에 비 선형성이 증가하게 된다.



transfer learning

  1. 모델을 가져와서 ‘include_top=False’를 통해 분류층(classification)을 삭제시킨다.
  2. ‘model.trainable = False’로 설정시켜서 동결(freezing)을 시킨다.
  3. global_average_pooling과 dense layer를 쌓은 후 컴파일한다.
  4. ImageDataGenerator를 통해 적은량의 데이터를 data augmentation을 진행시킨 후 학습시킨다.
  5. ‘model.trainable = True’로 설정시켜서 동결을 해제한다.
  6. learning_rate을 매우 낮게 설정하여 fine-tuning하기

1
!nvidia-smi
Mon Jan 25 13:30:28 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.104      Driver Version: 410.104      CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla V100-DGXS...  On   | 00000000:07:00.0 Off |                    0 |
| N/A   41C    P0    49W / 300W |    947MiB / 32478MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Tesla V100-DGXS...  On   | 00000000:08:00.0 Off |                    0 |
| N/A   39C    P0    40W / 300W |     12MiB / 32478MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   2  Tesla V100-DGXS...  On   | 00000000:0E:00.0 Off |                    0 |
| N/A   40C    P0    52W / 300W |   5134MiB / 32478MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   3  Tesla V100-DGXS...  On   | 00000000:0F:00.0 Off |                    0 |
| N/A   39C    P0    38W / 300W |     38MiB / 32478MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

1. classification층을 삭제


1
keras.application.VGG16(weights,input_shape,include_top)

include_top은 최상단 Layer인 fully connected layer를 포함시키는지 여부를 전달시키는 것으로 pre-trained된 것을 새롭게 prediction하기위해 이 layer는 제거시켜줍니다.

1
2
3
4
5
6
from tensorflow import keras

base_model = keras.applications.VGG16(
    weights='imagenet',  # Load weights pre-trained on ImageNet.
    input_shape=(224, 224, 3),
    include_top=False)
1
base_model.summary()
Model: "vgg16"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
input_1 (InputLayer)         [(None, 224, 224, 3)]     0
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 224, 224, 64)      1792
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 224, 224, 64)      36928
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 112, 112, 64)      0
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 112, 112, 128)     73856
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 112, 112, 128)     147584
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 56, 56, 128)       0
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 56, 56, 256)       295168
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 56, 56, 256)       590080
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 56, 56, 256)       590080
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 28, 28, 256)       0
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 28, 28, 512)       1180160
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 28, 28, 512)       2359808
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 28, 28, 512)       2359808
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 14, 14, 512)       0
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 14, 14, 512)       2359808
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 14, 14, 512)       2359808
_________________________________________________________________
block5_conv3 (Conv2D)        (None, 14, 14, 512)       2359808
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 7, 7, 512)         0
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0
_________________________________________________________________

이전과 다르게 classification layer이 삭제되었다. (Flatten ~ Dense)

1
base_model.trainable = False

2. 기존의 네트워크 동결시키기

trainable옵션을 False로 하여 새로운 학습을 할 때에 기존의 학습된 네트워크가 같이 학습되지 않도록 합니다.

1
2
3
4
5
6
7
8
inputs = keras.Input(shape=(224, 224, 3))
# Separately from setting trainable on the model, we set training to False
x = base_model(inputs, training=False)
x = keras.layers.GlobalAveragePooling2D()(x)
# A Dense classifier with a single unit (binary classification)
outputs = keras.layers.Dense(1)(x)
model = keras.Model(inputs, outputs)
model.summary()
Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
input_2 (InputLayer)         [(None, 224, 224, 3)]     0
_________________________________________________________________
vgg16 (Model)                (None, 7, 7, 512)         14714688
_________________________________________________________________
global_average_pooling2d (Gl (None, 512)               0
_________________________________________________________________
dense (Dense)                (None, 1)                 513
=================================================================
Total params: 14,715,201
Trainable params: 513
Non-trainable params: 14,714,688
_________________________________________________________________

about layer

  1. GlobalAveragePooing2D
  • 기존의 fully convolutional Network는 많은 parameter수와 계산이 오래걸리고, 위치의 정보가 모두 사라지는 단점이 있습니다.
  • 같은 chennel의 feature들의 평균을 내어 채널의 개수만큼 원소를 가지는 벡터를 출력 (이경우에는 입력 층이 (7,7,512)이므로 (512,)를 출력시키는 layer)
  • fully connected layer를 없애기 위한 방법
  • 코드
  1. Dense(1)
  • binary로 클래스를 분류할 것이기 때문에 0과 1로 classification
1
2
# Important to use binary crossentropy and binary accuracy as we now have a binary classification problem
model.compile(loss=keras.losses.BinaryCrossentropy(from_logits=True), metrics=[keras.metrics.BinaryAccuracy()])

before data augmentation

check

위 그림과같이 ImageDataGenerator.flow_from_directory를 사용할 때에 hidden folder가 있는지 확인해야합니다.

1
2
3
4
5
6
7
dog
        test
                bo
                not_bo
        train
                bo
                not_bo

위와 같이 directory별로 class를 구분하기 때문에 .ipynb_checkpoints폴더를 삭제해야합니다.

1
2
3
4
5
6
7
8
9
10
from tensorflow.keras.preprocessing.image import ImageDataGenerator
# create a data generator
datagen = ImageDataGenerator(
        samplewise_center=True,  # set each sample mean to 0
        rotation_range=10,  # randomly rotate images in the range (degrees, 0 to 180)
        zoom_range = 0.1, # Randomly zoom image
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
        horizontal_flip=True,  # randomly flip images
        vertical_flip=False) # we don't expect Bo to be upside-down so we will not flip vertically

data augmentation

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
# load and iterate training dataset

seed = 1

train_it = datagen.flow_from_directory('../sungjin/dog/train/',
#                                        save_to_dir='../sungjin/dog2/train/', //2700여개의 파일이 생성됨
                                       seed = seed,
                                       target_size=(224, 224),
                                       color_mode='rgb',
                                       class_mode='binary',
                                       batch_size=8)
# load and iterate test dataset
test_it = datagen.flow_from_directory('../sungjin/dog/test/',
#                                       save_to_dir='../sungjin/dog2/test/', // 1600여개의 파일이 생성됨
                                      seed = seed,
                                      target_size=(224, 224),
                                      color_mode='rgb',
                                      class_mode='binary',
                                      batch_size=8)
Found 139 images belonging to 2 classes.
Found 20 images belonging to 2 classes.
1
history = model.fit(train_it, steps_per_epoch=12, validation_data=test_it, validation_steps=4, epochs=20)
Epoch 1/20
12/12 [==============================] - 3s 225ms/step - loss: 1.4831 - binary_accuracy: 0.6562 - val_loss: 3.5233 - val_binary_accuracy: 0.5357
Epoch 2/20
12/12 [==============================] - 2s 178ms/step - loss: 1.1998 - binary_accuracy: 0.7473 - val_loss: 1.9760 - val_binary_accuracy: 0.6667
Epoch 3/20
12/12 [==============================] - 2s 161ms/step - loss: 0.4746 - binary_accuracy: 0.8462 - val_loss: 1.3999 - val_binary_accuracy: 0.7143
Epoch 4/20
12/12 [==============================] - 2s 172ms/step - loss: 0.3132 - binary_accuracy: 0.8646 - val_loss: 1.0363 - val_binary_accuracy: 0.7857
Epoch 5/20
12/12 [==============================] - 2s 175ms/step - loss: 0.2785 - binary_accuracy: 0.9011 - val_loss: 1.0750 - val_binary_accuracy: 0.7917
Epoch 6/20
12/12 [==============================] - 2s 178ms/step - loss: 0.0651 - binary_accuracy: 0.9670 - val_loss: 1.3541 - val_binary_accuracy: 0.7143
Epoch 7/20
12/12 [==============================] - 2s 198ms/step - loss: 0.1798 - binary_accuracy: 0.9167 - val_loss: 0.9755 - val_binary_accuracy: 0.8214
Epoch 8/20
12/12 [==============================] - 2s 166ms/step - loss: 0.0315 - binary_accuracy: 0.9890 - val_loss: 0.5196 - val_binary_accuracy: 0.9167
Epoch 9/20
12/12 [==============================] - 2s 179ms/step - loss: 0.0822 - binary_accuracy: 0.9341 - val_loss: 0.3601 - val_binary_accuracy: 0.8929
Epoch 10/20
12/12 [==============================] - 2s 181ms/step - loss: 0.0169 - binary_accuracy: 0.9896 - val_loss: 0.3563 - val_binary_accuracy: 0.9286
Epoch 11/20
12/12 [==============================] - 2s 191ms/step - loss: 0.0437 - binary_accuracy: 0.9890 - val_loss: 0.4602 - val_binary_accuracy: 0.8333
Epoch 12/20
12/12 [==============================] - 2s 158ms/step - loss: 0.0405 - binary_accuracy: 0.9890 - val_loss: 0.2497 - val_binary_accuracy: 0.8929
Epoch 13/20
12/12 [==============================] - 2s 191ms/step - loss: 0.0083 - binary_accuracy: 1.0000 - val_loss: 0.4642 - val_binary_accuracy: 0.8571
Epoch 14/20
12/12 [==============================] - 2s 163ms/step - loss: 0.0133 - binary_accuracy: 0.9890 - val_loss: 0.2259 - val_binary_accuracy: 0.9167
Epoch 15/20
12/12 [==============================] - 2s 178ms/step - loss: 0.0050 - binary_accuracy: 1.0000 - val_loss: 0.2705 - val_binary_accuracy: 0.8929
Epoch 16/20
12/12 [==============================] - 2s 185ms/step - loss: 0.0027 - binary_accuracy: 1.0000 - val_loss: 0.2617 - val_binary_accuracy: 0.8929
Epoch 17/20
12/12 [==============================] - 2s 179ms/step - loss: 0.0203 - binary_accuracy: 0.9890 - val_loss: 0.1751 - val_binary_accuracy: 0.9583
Epoch 18/20
12/12 [==============================] - 2s 173ms/step - loss: 0.0055 - binary_accuracy: 1.0000 - val_loss: 0.0705 - val_binary_accuracy: 1.0000
Epoch 19/20
12/12 [==============================] - 2s 199ms/step - loss: 0.0015 - binary_accuracy: 1.0000 - val_loss: 0.0275 - val_binary_accuracy: 1.0000
Epoch 20/20
12/12 [==============================] - 2s 189ms/step - loss: 0.0056 - binary_accuracy: 1.0000 - val_loss: 0.1403 - val_binary_accuracy: 0.9583

기존의 네트워크 동결 해제 및 fine-tuning

새롭게 컴파일할 때 기존의 네트워크를 unfreezing하여 fine-tuning(미세조정)한다.

단, 이때 다음과 같은 세가지의 전략이 있다.

  1. 전체모델을 학습시키기
  • 이 방법은 pre-trained model의 구조만 사용하고 새로운 데이터셋에 맞게 전부 새로 학습시키는 방법입니다. 큰 사이즈의 데이터셋이 필요합니다.

  1. 일부의 층만 학습시키고 나머지는 동결시키기
  • 초기 레이어는 추상적인(일반적인)특징을 추출하고 후기 레이어는 구체적이고 특유한 특징을 추출하는데 이런 특성을 이용해서 선택적으로 재학습시킵니다.
  • 데이터셋이 작고 모델의 파라미터가 많다면 오버피팅이 될 수 있으므로 적은양의 계층을 학습시킵니다.
  • 데이터셋이 크고 그에 비해 모델의 파라미터가 적다면 오버피팅의 가능성이 적으므로 더 많은 양의 계층을 학습시킵니다.

  1. 전체 동결시키기 (classification만 학습시키기)
  • 기존의 네트워크는 건들지 않고 그대로 두면서 특징 추출 메커니즘으로써 활용하고, classifier만 재학습시키는 방법을 씁니다.
  • 데이터셋이 너무 작을때나 새로운 문제가 이전 모델이 이미 학습한 데이터셋과 매우 비슷할 때 사용합니다.

1
2
3
4
5
6
7
8
9
10
# Unfreeze the base model
base_model.trainable = True

# It's important to recompile your model after you make any changes
# to the `trainable` attribute of any inner layer, so that your changes
# are taken into account
model.compile(optimizer = keras.optimizers.Adam(lr = .00001),
#               optimizer=keras.optimizers.RMSprop(learning_rate = .00001),  # Very low learning rate
              loss=keras.losses.BinaryCrossentropy(from_logits=True),
              metrics=[keras.metrics.BinaryAccuracy()])

optimizer


reference

1
history = model.fit(train_it, steps_per_epoch=12, validation_data=test_it, validation_steps=4, epochs=20)
Epoch 1/20
12/12 [==============================] - 3s 219ms/step - loss: 0.0849 - binary_accuracy: 0.9670 - val_loss: 0.0158 - val_binary_accuracy: 1.0000
Epoch 2/20
12/12 [==============================] - 2s 204ms/step - loss: 0.0053 - binary_accuracy: 1.0000 - val_loss: 0.0118 - val_binary_accuracy: 1.0000
Epoch 3/20
12/12 [==============================] - 2s 176ms/step - loss: 0.0020 - binary_accuracy: 1.0000 - val_loss: 0.0015 - val_binary_accuracy: 1.0000
Epoch 4/20
12/12 [==============================] - 2s 196ms/step - loss: 1.5012e-04 - binary_accuracy: 1.0000 - val_loss: 0.0154 - val_binary_accuracy: 1.0000
Epoch 5/20
12/12 [==============================] - 3s 211ms/step - loss: 1.7610e-04 - binary_accuracy: 1.0000 - val_loss: 0.0034 - val_binary_accuracy: 1.0000
Epoch 6/20
12/12 [==============================] - 2s 198ms/step - loss: 1.0886e-04 - binary_accuracy: 1.0000 - val_loss: 0.0275 - val_binary_accuracy: 1.0000
Epoch 7/20
12/12 [==============================] - 2s 178ms/step - loss: 2.9665e-05 - binary_accuracy: 1.0000 - val_loss: 0.0093 - val_binary_accuracy: 1.0000
Epoch 8/20
12/12 [==============================] - 2s 205ms/step - loss: 1.7676e-05 - binary_accuracy: 1.0000 - val_loss: 0.0108 - val_binary_accuracy: 1.0000
Epoch 9/20
12/12 [==============================] - 2s 182ms/step - loss: 3.1658e-05 - binary_accuracy: 1.0000 - val_loss: 8.7812e-04 - val_binary_accuracy: 1.0000
Epoch 10/20
12/12 [==============================] - 2s 195ms/step - loss: 3.5808e-05 - binary_accuracy: 1.0000 - val_loss: 0.0020 - val_binary_accuracy: 1.0000
Epoch 11/20
12/12 [==============================] - 2s 191ms/step - loss: 3.2332e-05 - binary_accuracy: 1.0000 - val_loss: 0.0474 - val_binary_accuracy: 0.9643
Epoch 12/20
12/12 [==============================] - 2s 202ms/step - loss: 1.3189e-04 - binary_accuracy: 1.0000 - val_loss: 0.0044 - val_binary_accuracy: 1.0000
Epoch 13/20
12/12 [==============================] - 2s 202ms/step - loss: 1.7362e-04 - binary_accuracy: 1.0000 - val_loss: 0.0073 - val_binary_accuracy: 1.0000
Epoch 14/20
12/12 [==============================] - 2s 194ms/step - loss: 2.4271e-05 - binary_accuracy: 1.0000 - val_loss: 0.0235 - val_binary_accuracy: 1.0000
Epoch 15/20
12/12 [==============================] - 3s 219ms/step - loss: 1.0316e-05 - binary_accuracy: 1.0000 - val_loss: 0.0388 - val_binary_accuracy: 1.0000
Epoch 16/20
12/12 [==============================] - 2s 187ms/step - loss: 2.3622e-05 - binary_accuracy: 1.0000 - val_loss: 0.0031 - val_binary_accuracy: 1.0000
Epoch 17/20
12/12 [==============================] - 2s 201ms/step - loss: 2.0185e-05 - binary_accuracy: 1.0000 - val_loss: 0.0413 - val_binary_accuracy: 0.9643
Epoch 18/20
12/12 [==============================] - 2s 175ms/step - loss: 3.8968e-05 - binary_accuracy: 1.0000 - val_loss: 0.0197 - val_binary_accuracy: 1.0000
Epoch 19/20
12/12 [==============================] - 2s 202ms/step - loss: 9.7374e-06 - binary_accuracy: 1.0000 - val_loss: 0.0902 - val_binary_accuracy: 0.9643
Epoch 20/20
12/12 [==============================] - 2s 206ms/step - loss: 1.0378e-04 - binary_accuracy: 1.0000 - val_loss: 0.0148 - val_binary_accuracy: 1.0000
1
2
3
4
5
6
7
8
9
inputs = keras.Input(shape=(224, 224, 3))
# Separately from setting trainable on the model, we set training to False
x = base_model(inputs, training=False)
x = keras.layers.Flatten(input_shape=(7,7,512))(x)
# x = keras.layers.Dense(256, activation='relu',input_dim=(7*7*512))(x)
# x = keras.layers.Dropout(0.2)(x)
outputs = keras.layers.Dense(1,activation = 'sigmoid')(x)
model = keras.Model(inputs, outputs)
model.summary()
Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
input_2 (InputLayer)         [(None, 224, 224, 3)]     0
_________________________________________________________________
vgg16 (Model)                (None, 7, 7, 512)         14714688
_________________________________________________________________
flatten (Flatten)            (None, 25088)             0
_________________________________________________________________
dense (Dense)                (None, 1)                 25089
=================================================================
Total params: 14,739,777
Trainable params: 25,089
Non-trainable params: 14,714,688
_________________________________________________________________
1
2
3
4
5
# Important to use binary crossentropy and binary accuracy as we now have a binary classification problem
model.compile(
#     optimizer=keras.optimizers.Adam(),
              loss=keras.losses.BinaryCrossentropy(from_logits=True),
              metrics=[keras.metrics.BinaryAccuracy()])
1
2
3
4
5
6
7
8
9
10
from tensorflow.keras.preprocessing.image import ImageDataGenerator
# create a data generator
datagen = ImageDataGenerator(
        samplewise_center=True,  # set each sample mean to 0
        rotation_range=10,  # randomly rotate images in the range (degrees, 0 to 180)
        zoom_range = 0.1, # Randomly zoom image
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
        horizontal_flip=True,  # randomly flip images
        vertical_flip=False) # we don't expect Bo to be upside-down so we will not flip vertically
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
# load and iterate training dataset

seed = 1

train_it = datagen.flow_from_directory('../sungjin/dog/train/',
#                                        save_to_dir='../sungjin/dog2/train/', //2700여개의 파일이 생성됨
                                       seed = seed,
                                       target_size=(224, 224),
                                       color_mode='rgb',
                                       class_mode='binary',
                                       batch_size=8)
# load and iterate test dataset
test_it = datagen.flow_from_directory('../sungjin/dog/test/',
#                                       save_to_dir='../sungjin/dog2/test/', // 1600여개의 파일이 생성됨
                                      seed = seed,
                                      target_size=(224, 224),
                                      color_mode='rgb',
                                      class_mode='binary',
                                      batch_size=8)
Found 139 images belonging to 2 classes.
Found 20 images belonging to 2 classes.
1
history = model.fit(train_it, steps_per_epoch=12, validation_data=test_it, validation_steps=4, epochs=20)
Epoch 1/20
12/12 [==============================] - 3s 243ms/step - loss: 0.4516 - binary_accuracy: 0.8333 - val_loss: 0.8132 - val_binary_accuracy: 0.5000
Epoch 2/20
12/12 [==============================] - 2s 195ms/step - loss: 0.4823 - binary_accuracy: 0.8352 - val_loss: 0.8549 - val_binary_accuracy: 0.4583
Epoch 3/20
12/12 [==============================] - 2s 179ms/step - loss: 0.4232 - binary_accuracy: 0.8901 - val_loss: 0.8483 - val_binary_accuracy: 0.4643
Epoch 4/20
12/12 [==============================] - 2s 202ms/step - loss: 0.4383 - binary_accuracy: 0.8750 - val_loss: 0.8490 - val_binary_accuracy: 0.4643
Epoch 5/20
12/12 [==============================] - 2s 195ms/step - loss: 0.4781 - binary_accuracy: 0.8352 - val_loss: 0.8133 - val_binary_accuracy: 0.5000
Epoch 6/20
12/12 [==============================] - 2s 199ms/step - loss: 0.4341 - binary_accuracy: 0.8791 - val_loss: 0.7775 - val_binary_accuracy: 0.5357
Epoch 7/20
12/12 [==============================] - 2s 204ms/step - loss: 0.4526 - binary_accuracy: 0.8646 - val_loss: 0.7197 - val_binary_accuracy: 0.6071
Epoch 8/20
12/12 [==============================] - 2s 170ms/step - loss: 0.4012 - binary_accuracy: 0.9121 - val_loss: 0.8549 - val_binary_accuracy: 0.4583
Epoch 9/20
12/12 [==============================] - 2s 197ms/step - loss: 0.4823 - binary_accuracy: 0.8352 - val_loss: 0.8490 - val_binary_accuracy: 0.4643
Epoch 10/20
12/12 [==============================] - 2s 201ms/step - loss: 0.4070 - binary_accuracy: 0.9062 - val_loss: 0.8133 - val_binary_accuracy: 0.5000
Epoch 11/20
12/12 [==============================] - 2s 205ms/step - loss: 0.5111 - binary_accuracy: 0.8022 - val_loss: 0.7716 - val_binary_accuracy: 0.5417
Epoch 12/20
12/12 [==============================] - 2s 181ms/step - loss: 0.4341 - binary_accuracy: 0.8791 - val_loss: 0.8490 - val_binary_accuracy: 0.4643
Epoch 13/20
12/12 [==============================] - 3s 216ms/step - loss: 0.4383 - binary_accuracy: 0.8750 - val_loss: 0.8133 - val_binary_accuracy: 0.5000
Epoch 14/20
12/12 [==============================] - 2s 190ms/step - loss: 0.4451 - binary_accuracy: 0.8681 - val_loss: 0.8133 - val_binary_accuracy: 0.5000
Epoch 15/20
12/12 [==============================] - 2s 191ms/step - loss: 0.4603 - binary_accuracy: 0.8571 - val_loss: 0.8847 - val_binary_accuracy: 0.4286
Epoch 16/20
12/12 [==============================] - 2s 202ms/step - loss: 0.4278 - binary_accuracy: 0.8854 - val_loss: 0.8133 - val_binary_accuracy: 0.5000
Epoch 17/20
12/12 [==============================] - 2s 205ms/step - loss: 0.4451 - binary_accuracy: 0.8681 - val_loss: 0.8549 - val_binary_accuracy: 0.4583
Epoch 18/20
12/12 [==============================] - 2s 179ms/step - loss: 0.4781 - binary_accuracy: 0.8352 - val_loss: 0.8490 - val_binary_accuracy: 0.4643
Epoch 19/20
12/12 [==============================] - 2s 207ms/step - loss: 0.4695 - binary_accuracy: 0.8438 - val_loss: 0.7775 - val_binary_accuracy: 0.5357
Epoch 20/20
12/12 [==============================] - 2s 185ms/step - loss: 0.4232 - binary_accuracy: 0.8901 - val_loss: 0.8549 - val_binary_accuracy: 0.4583

Global Average Pooing vs Fully Convolutional Network

PREV
2021-01-27 # AAIS # AI # AAIS # keras
NEXT