AIB 딥러닝 섹션 회고
한 달 동안 딥러닝을 배웠지만, 너무 방대한 양이므로 정리하기도 힘들다. 내가 얻은 교훈은…마음을 잘 정리해야 한다. 몇 년 걸려서 배울 내용을 한 달 안에 억지로 욱여 넣으면 당연히 안 되는데 그걸 진행하고 있는 부트캠프들은 대단하다. 왜 하나의 딥러닝 세부분야를 골라서 면밀히 가르치지 않고 “이런 것이 있습니다”라면서 매일 같이 새로운 것을 던져 주는지 이해할 수 없다. 마음을 비우고 시간 날 때 하나씩 여러 예제를 찾아보는 것이 좋다. 노트도 어차피 TensorFlow 튜토리얼 그대로 가져오는데, 최신 버전도 아니고 예전 버전이다. 계속 에러 난다. “노답” 이라는 생각을 저 구석에 치워 놓고 스스로 배워야 한다. 별로 좋지 않은 교육자료로 스스로 좋은 자료를 찾아보도록 자기주도학습을 유도한 것인가 싶다.
마음 한 쪽 답답했던 문제가 하나 있었는데 오늘 아침에 해결했다. 그 문제는 N332의 U-Net에서 왜 몇 개의 layer만 골라서 모델로 구성하는지였다. 튜토리얼
base_model = tf.keras.applications.MobileNetV2(input_shape=[128, 128, 3], include_top=False)
# Use the activations of these layers
layer_names = [
'block_1_expand_relu', # 64x64
'block_3_expand_relu', # 32x32
'block_6_expand_relu', # 16x16
'block_13_expand_relu', # 8x8
'block_16_project', # 4x4
]
base_model_outputs = [base_model.get_layer(name).output for name in layer_names]
# Create the feature extraction model
down_stack = tf.keras.Model(inputs=base_model.input, outputs=base_model_outputs)
down_stack.trainable = False
딱 이 부분인데, layer_names로 다섯 개의 layer만 구성한 인코더로 보인다. 사전학습 모델에서 중간 중간 사용할 layer만 선택하는 것으로 보이는 문제. 그런데 down_stack.summary()를 보면 마지막 layer까지 다 들어 있다.
block_3_pad (ZeroPadding2D) (None, 33, 33, 144) 0 ['block_3_expand_relu[0][0]']
block_3_depthwise (DepthwiseCo (None, 16, 16, 144) 1296 ['block_3_pad[0][0]']
nv2D)
block_3_depthwise_BN (BatchNor (None, 16, 16, 144) 576 ['block_3_depthwise[0][0]']
malization)
block_3_depthwise_relu (ReLU) (None, 16, 16, 144) 0 ['block_3_depthwise_BN[0][0]']
block_3_project (Conv2D) (None, 16, 16, 32) 4608 ['block_3_depthwise_relu[0][0]']
block_3_project_BN (BatchNorma (None, 16, 16, 32) 128 ['block_3_project[0][0]']
lization)
block_4_expand (Conv2D) (None, 16, 16, 192) 6144 ['block_3_project_BN[0][0]']
block_4_expand_BN (BatchNormal (None, 16, 16, 192) 768 ['block_4_expand[0][0]']
ization)
block_4_expand_relu (ReLU) (None, 16, 16, 192) 0 ['block_4_expand_BN[0][0]']
block_4_depthwise (DepthwiseCo (None, 16, 16, 192) 1728 ['block_4_expand_relu[0][0]']
nv2D)
block_4_depthwise_BN (BatchNor (None, 16, 16, 192) 768 ['block_4_depthwise[0][0]']
malization)
block_4_depthwise_relu (ReLU) (None, 16, 16, 192) 0 ['block_4_depthwise_BN[0][0]']
block_4_project (Conv2D) (None, 16, 16, 32) 6144 ['block_4_depthwise_relu[0][0]']
block_4_project_BN (BatchNorma (None, 16, 16, 32) 128 ['block_4_project[0][0]']
lization)
block_4_add (Add) (None, 16, 16, 32) 0 ['block_3_project_BN[0][0]',
'block_4_project_BN[0][0]']
block_5_expand (Conv2D) (None, 16, 16, 192) 6144 ['block_4_add[0][0]']
block_5_expand_BN (BatchNormal (None, 16, 16, 192) 768 ['block_5_expand[0][0]']
ization)
block_5_expand_relu (ReLU) (None, 16, 16, 192) 0 ['block_5_expand_BN[0][0]']
block_5_depthwise (DepthwiseCo (None, 16, 16, 192) 1728 ['block_5_expand_relu[0][0]']
nv2D)
block_5_depthwise_BN (BatchNor (None, 16, 16, 192) 768 ['block_5_depthwise[0][0]']
malization)
block_5_depthwise_relu (ReLU) (None, 16, 16, 192) 0 ['block_5_depthwise_BN[0][0]']
block_5_project (Conv2D) (None, 16, 16, 32) 6144 ['block_5_depthwise_relu[0][0]']
block_5_project_BN (BatchNorma (None, 16, 16, 32) 128 ['block_5_project[0][0]']
lization)
block_5_add (Add) (None, 16, 16, 32) 0 ['block_4_add[0][0]',
'block_5_project_BN[0][0]']
block_6_expand (Conv2D) (None, 16, 16, 192) 6144 ['block_5_add[0][0]']
block_6_expand_BN (BatchNormal (None, 16, 16, 192) 768 ['block_6_expand[0][0]']
ization)
block_6_expand_relu (ReLU) (None, 16, 16, 192) 0 ['block_6_expand_BN[0][0]']
block_6_pad (ZeroPadding2D) (None, 17, 17, 192) 0 ['block_6_expand_relu[0][0]']
block_6_depthwise (DepthwiseCo (None, 8, 8, 192) 1728 ['block_6_pad[0][0]']
nv2D)
block_6_depthwise_BN (BatchNor (None, 8, 8, 192) 768 ['block_6_depthwise[0][0]']
malization)
block_6_depthwise_relu (ReLU) (None, 8, 8, 192) 0 ['block_6_depthwise_BN[0][0]']
block_6_project (Conv2D) (None, 8, 8, 64) 12288 ['block_6_depthwise_relu[0][0]']
block_6_project_BN (BatchNorma (None, 8, 8, 64) 256 ['block_6_project[0][0]']
lization)
block_7_expand (Conv2D) (None, 8, 8, 384) 24576 ['block_6_project_BN[0][0]']
block_7_expand_BN (BatchNormal (None, 8, 8, 384) 1536 ['block_7_expand[0][0]']
ization)
block_7_expand_relu (ReLU) (None, 8, 8, 384) 0 ['block_7_expand_BN[0][0]']
block_7_depthwise (DepthwiseCo (None, 8, 8, 384) 3456 ['block_7_expand_relu[0][0]']
nv2D)
block_7_depthwise_BN (BatchNor (None, 8, 8, 384) 1536 ['block_7_depthwise[0][0]']
malization)
block_7_depthwise_relu (ReLU) (None, 8, 8, 384) 0 ['block_7_depthwise_BN[0][0]']
block_7_project (Conv2D) (None, 8, 8, 64) 24576 ['block_7_depthwise_relu[0][0]']
block_7_project_BN (BatchNorma (None, 8, 8, 64) 256 ['block_7_project[0][0]']
lization)
block_7_add (Add) (None, 8, 8, 64) 0 ['block_6_project_BN[0][0]',
'block_7_project_BN[0][0]']
block_8_expand (Conv2D) (None, 8, 8, 384) 24576 ['block_7_add[0][0]']
block_8_expand_BN (BatchNormal (None, 8, 8, 384) 1536 ['block_8_expand[0][0]']
ization)
block_8_expand_relu (ReLU) (None, 8, 8, 384) 0 ['block_8_expand_BN[0][0]']
block_8_depthwise (DepthwiseCo (None, 8, 8, 384) 3456 ['block_8_expand_relu[0][0]']
nv2D)
block_8_depthwise_BN (BatchNor (None, 8, 8, 384) 1536 ['block_8_depthwise[0][0]']
malization)
block_8_depthwise_relu (ReLU) (None, 8, 8, 384) 0 ['block_8_depthwise_BN[0][0]']
block_8_project (Conv2D) (None, 8, 8, 64) 24576 ['block_8_depthwise_relu[0][0]']
block_8_project_BN (BatchNorma (None, 8, 8, 64) 256 ['block_8_project[0][0]']
lization)
block_8_add (Add) (None, 8, 8, 64) 0 ['block_7_add[0][0]',
'block_8_project_BN[0][0]']
block_9_expand (Conv2D) (None, 8, 8, 384) 24576 ['block_8_add[0][0]']
block_9_expand_BN (BatchNormal (None, 8, 8, 384) 1536 ['block_9_expand[0][0]']
ization)
block_9_expand_relu (ReLU) (None, 8, 8, 384) 0 ['block_9_expand_BN[0][0]']
block_9_depthwise (DepthwiseCo (None, 8, 8, 384) 3456 ['block_9_expand_relu[0][0]']
nv2D)
block_9_depthwise_BN (BatchNor (None, 8, 8, 384) 1536 ['block_9_depthwise[0][0]']
malization)
block_9_depthwise_relu (ReLU) (None, 8, 8, 384) 0 ['block_9_depthwise_BN[0][0]']
block_9_project (Conv2D) (None, 8, 8, 64) 24576 ['block_9_depthwise_relu[0][0]']
block_9_project_BN (BatchNorma (None, 8, 8, 64) 256 ['block_9_project[0][0]']
lization)
block_9_add (Add) (None, 8, 8, 64) 0 ['block_8_add[0][0]',
'block_9_project_BN[0][0]']
block_10_expand (Conv2D) (None, 8, 8, 384) 24576 ['block_9_add[0][0]']
block_10_expand_BN (BatchNorma (None, 8, 8, 384) 1536 ['block_10_expand[0][0]']
lization)
block_10_expand_relu (ReLU) (None, 8, 8, 384) 0 ['block_10_expand_BN[0][0]']
block_10_depthwise (DepthwiseC (None, 8, 8, 384) 3456 ['block_10_expand_relu[0][0]']
onv2D)
block_10_depthwise_BN (BatchNo (None, 8, 8, 384) 1536 ['block_10_depthwise[0][0]']
rmalization)
block_10_depthwise_relu (ReLU) (None, 8, 8, 384) 0 ['block_10_depthwise_BN[0][0]']
block_10_project (Conv2D) (None, 8, 8, 96) 36864 ['block_10_depthwise_relu[0][0]']
block_10_project_BN (BatchNorm (None, 8, 8, 96) 384 ['block_10_project[0][0]']
alization)
block_11_expand (Conv2D) (None, 8, 8, 576) 55296 ['block_10_project_BN[0][0]']
block_11_expand_BN (BatchNorma (None, 8, 8, 576) 2304 ['block_11_expand[0][0]']
lization)
block_11_expand_relu (ReLU) (None, 8, 8, 576) 0 ['block_11_expand_BN[0][0]']
block_11_depthwise (DepthwiseC (None, 8, 8, 576) 5184 ['block_11_expand_relu[0][0]']
onv2D)
block_11_depthwise_BN (BatchNo (None, 8, 8, 576) 2304 ['block_11_depthwise[0][0]']
rmalization)
block_11_depthwise_relu (ReLU) (None, 8, 8, 576) 0 ['block_11_depthwise_BN[0][0]']
block_11_project (Conv2D) (None, 8, 8, 96) 55296 ['block_11_depthwise_relu[0][0]']
block_11_project_BN (BatchNorm (None, 8, 8, 96) 384 ['block_11_project[0][0]']
alization)
block_11_add (Add) (None, 8, 8, 96) 0 ['block_10_project_BN[0][0]',
'block_11_project_BN[0][0]']
block_12_expand (Conv2D) (None, 8, 8, 576) 55296 ['block_11_add[0][0]']
block_12_expand_BN (BatchNorma (None, 8, 8, 576) 2304 ['block_12_expand[0][0]']
lization)
block_12_expand_relu (ReLU) (None, 8, 8, 576) 0 ['block_12_expand_BN[0][0]']
block_12_depthwise (DepthwiseC (None, 8, 8, 576) 5184 ['block_12_expand_relu[0][0]']
onv2D)
block_12_depthwise_BN (BatchNo (None, 8, 8, 576) 2304 ['block_12_depthwise[0][0]']
rmalization)
block_12_depthwise_relu (ReLU) (None, 8, 8, 576) 0 ['block_12_depthwise_BN[0][0]']
block_12_project (Conv2D) (None, 8, 8, 96) 55296 ['block_12_depthwise_relu[0][0]']
block_12_project_BN (BatchNorm (None, 8, 8, 96) 384 ['block_12_project[0][0]']
alization)
block_12_add (Add) (None, 8, 8, 96) 0 ['block_11_add[0][0]',
'block_12_project_BN[0][0]']
block_13_expand (Conv2D) (None, 8, 8, 576) 55296 ['block_12_add[0][0]']
block_13_expand_BN (BatchNorma (None, 8, 8, 576) 2304 ['block_13_expand[0][0]']
lization)
block_13_expand_relu (ReLU) (None, 8, 8, 576) 0 ['block_13_expand_BN[0][0]']
block_13_pad (ZeroPadding2D) (None, 9, 9, 576) 0 ['block_13_expand_relu[0][0]']
block_13_depthwise (DepthwiseC (None, 4, 4, 576) 5184 ['block_13_pad[0][0]']
onv2D)
block_13_depthwise_BN (BatchNo (None, 4, 4, 576) 2304 ['block_13_depthwise[0][0]']
rmalization)
block_13_depthwise_relu (ReLU) (None, 4, 4, 576) 0 ['block_13_depthwise_BN[0][0]']
block_13_project (Conv2D) (None, 4, 4, 160) 92160 ['block_13_depthwise_relu[0][0]']
block_13_project_BN (BatchNorm (None, 4, 4, 160) 640 ['block_13_project[0][0]']
alization)
block_14_expand (Conv2D) (None, 4, 4, 960) 153600 ['block_13_project_BN[0][0]']
block_14_expand_BN (BatchNorma (None, 4, 4, 960) 3840 ['block_14_expand[0][0]']
lization)
block_14_expand_relu (ReLU) (None, 4, 4, 960) 0 ['block_14_expand_BN[0][0]']
block_14_depthwise (DepthwiseC (None, 4, 4, 960) 8640 ['block_14_expand_relu[0][0]']
onv2D)
block_14_depthwise_BN (BatchNo (None, 4, 4, 960) 3840 ['block_14_depthwise[0][0]']
rmalization)
block_14_depthwise_relu (ReLU) (None, 4, 4, 960) 0 ['block_14_depthwise_BN[0][0]']
block_14_project (Conv2D) (None, 4, 4, 160) 153600 ['block_14_depthwise_relu[0][0]']
block_14_project_BN (BatchNorm (None, 4, 4, 160) 640 ['block_14_project[0][0]']
alization)
block_14_add (Add) (None, 4, 4, 160) 0 ['block_13_project_BN[0][0]',
'block_14_project_BN[0][0]']
block_15_expand (Conv2D) (None, 4, 4, 960) 153600 ['block_14_add[0][0]']
block_15_expand_BN (BatchNorma (None, 4, 4, 960) 3840 ['block_15_expand[0][0]']
lization)
block_15_expand_relu (ReLU) (None, 4, 4, 960) 0 ['block_15_expand_BN[0][0]']
block_15_depthwise (DepthwiseC (None, 4, 4, 960) 8640 ['block_15_expand_relu[0][0]']
onv2D)
block_15_depthwise_BN (BatchNo (None, 4, 4, 960) 3840 ['block_15_depthwise[0][0]']
rmalization)
block_15_depthwise_relu (ReLU) (None, 4, 4, 960) 0 ['block_15_depthwise_BN[0][0]']
block_15_project (Conv2D) (None, 4, 4, 160) 153600 ['block_15_depthwise_relu[0][0]']
block_15_project_BN (BatchNorm (None, 4, 4, 160) 640 ['block_15_project[0][0]']
alization)
block_15_add (Add) (None, 4, 4, 160) 0 ['block_14_add[0][0]',
'block_15_project_BN[0][0]']
block_16_expand (Conv2D) (None, 4, 4, 960) 153600 ['block_15_add[0][0]']
block_16_expand_BN (BatchNorma (None, 4, 4, 960) 3840 ['block_16_expand[0][0]']
lization)
block_16_expand_relu (ReLU) (None, 4, 4, 960) 0 ['block_16_expand_BN[0][0]']
block_16_depthwise (DepthwiseC (None, 4, 4, 960) 8640 ['block_16_expand_relu[0][0]']
onv2D)
block_16_depthwise_BN (BatchNo (None, 4, 4, 960) 3840 ['block_16_depthwise[0][0]']
rmalization)
block_16_depthwise_relu (ReLU) (None, 4, 4, 960) 0 ['block_16_depthwise_BN[0][0]']
block_16_project (Conv2D) (None, 4, 4, 320) 307200 ['block_16_depthwise_relu[0][0]']
==================================================================================================
Total params: 1,841,984
Trainable params: 0
Non-trainable params: 1,841,984
__________________________________________________________________________________________________
down_stack으로 모델 저장할 때 outputs=base_model_outputs로 하면 down_stack의 outputs은 다섯 개의 layer의 output이 출력되고, 이것은 down_stack.layers의 각각의 output과는 다르다. 이렇게 5 layer의 outputs을 정의했으니, 쉽게 가져다 사용할 수 있게 된다:
def unet_model(output_channels:int):
inputs = tf.keras.layers.Input(shape=[128, 128, 3])
# Downsampling through the model
skips = down_stack(inputs)
x = skips[-1]
skips = reversed(skips[:-1])
# Upsampling and establishing the skip connections
for up, skip in zip(up_stack, skips):
x = up(x)
concat = tf.keras.layers.Concatenate()
x = concat([x, skip])
# This is the last layer of the model
last = tf.keras.layers.Conv2DTranspose(
filters=output_channels, kernel_size=3, strides=2,
padding='same') #64x64 -> 128x128
x = last(x)
return tf.keras.Model(inputs=inputs, outputs=x)
기타 모델 구성하는 것도 어느 정도 익숙해졌다. 왜 이렇게 구성해야 하는지 모르겠지만, layer 넣으라면 넣을 수 있을 정도. 섹션 3 프로젝트에서 autoencoder 잠깐 시도해봤는데, 모델 중간 레이어를 뽑아서 그걸 loss로 넣어보는 것도 해봤다. 다른 것도 여러 번 해보면 잘 되겠지.
이번 한 달간 몸이 너무 안 좋은데 때려치고 싶은 생각도 많이 들었다. 왜 이제까지 컴퓨터 앞에 앉아도 자세 관리 잘 했었는데 뭐가 달라졌는지 모르겠다. 허리 아픈것부터 시작해서 뒷목이 너무 땡기니 뭘 할 수가 없다. 의자에 기대어 앉으면 내 몸이 스스로 근육에 힘을 주어 앉는 것이 아니라서 일부터 딱딱하고 등도 기댈 수 없는 의자에 앉는데 비싼 돈주고 컴퓨터 의자 사야할 것 같다. 몸이 조금 괜찮았다가 다시 나빠졌다가 하는데 정말 스트레스다.
댓글남기기