{"id":884,"date":"2020-09-25T19:32:26","date_gmt":"2020-09-25T10:32:26","guid":{"rendered":"http:\/\/leenux.kr\/?p=884"},"modified":"2020-10-08T23:49:40","modified_gmt":"2020-10-08T14:49:40","slug":"identify-the-role-concept-and-parameters-of-each-layer-in-tensorflow","status":"publish","type":"post","link":"https:\/\/leenux.kr\/?p=884","title":{"rendered":"Identify the role concept and parameters of each layer in TensorFlow"},"content":{"rendered":"\n
\n\n\n\n

Feature Extraction<\/strong><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\uc774\ubbf8\uc9c0\uc744 \ud2b9\uc9d5, \ud328\ud134\uc73c\ub85c \ub2e4\uc591\ud558\uac8c \ub370\uc774\ud130 \uc14b\uc744 \ud574\ubcf4\uace0 \ubaa8\ub378\uc744 \uad6c\uc131\ud574 \ub370\uc774\ud130 \uc14b \uac12\uc744 \ubaa8\ub378\uc5d0 \ub118\uaca8\ubcf4\ub294 \uc608\uc81c\ub97c \ud574\ubcf4\uacd8\ub2e4. <\/strong><\/p>\n\n\n\n

\n\n\n\n

Convolution<\/strong><\/p>\n\n\n\n

\uc774\ubbf8\uc9c0\uc744 \ud2b9\uc9d5, \ud328\ud134 \ub370\uc774\ud130 SET<\/strong><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\ud150\uc11c\ud50c\ub85c\uc6b0 \ucf00\ub77c\uc2a4\uc5d0\uc11c \uc81c\uacf5\ud574\uc8fc\ub294 Convolution\uc744 \ud1b5\ud574 \uc6d0\ud558\ub294 \uc635\uc158 \uc815\uc758\ud558\uace0 \uadf8 \uc635\uc158 \uc124\uc815\uac12\uc73c\ub85c \uc774\ubbf8\uc9c0\uac00 \ub2e4\uc591\ud558\uac8c \ub098\ud0c0\ub0b4\ub294 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uae30\uae4c\uc9c0 \ud558\uaca0\ub2e4.<\/strong><\/p>\n\n\n\n

<\/p>\n\n\n\n

\ud150\uc11c\ud50c\ub85c\uc6b0, \ub118\ud30c\uc774, \ub9c8\ud50c\ub78f \uac01 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub4e4\uc744 \ubd88\ub7ec\uc628\ub2e4.<\/strong><\/p>\n\n\n\n

import os\nimport tensorflow as tf\nimport numpy as np\nfrom PIL import Image\nimport matplotlib.pyplot as plt\n%matplotlib inline<\/code><\/pre>\n\n\n\n
path = 'source\/dog.jpg'\nfile = Image.open(path)\nimage = np.array(file)\n# \uc774\ubbf8\uc9c0 \uc55e\uc5d0 \ucc28\uc6d0\uc218 \ub298\ub9ac\uae30\nimage[tf.newaxis, ...]\nimage.shape # \uc250\uc774\ud504 \ud655\uc778<\/code><\/pre>\n\n\n\n
tf.keras.layers.Conv2D(filters, kernel_size, strides, padding, activation) <\/strong><\/p>\n\n\n\n
 tf.keras.layers.Conv2D<\/strong> \uc635\uc158 \uc124\uba85<\/strong><\/p>\n\n\n\n
filters: layer\uc5d0\uc11c \ub098\uac08 \ub54c \uba87 \uac1c\uc758 filter\ub97c \ub9cc\ub4e4 \uac83\uc778\uc9c0 (a.k.a weights, filters, channels)
 kernel_size: filter(Weight)\uc758 \uc0ac\uc774\uc988
 strides: \uba87 \uac1c\uc758 pixel\uc744 skip \ud558\uba74\uc11c \ud6d1\uc5b4\uc9c0\ub098\uac08 \uac83\uc778\uc9c0 (\uc0ac\uc774\uc988\uc5d0\ub3c4 \uc601\ud5a5\uc744 \uc90c)
 padding: zero padding\uc744 \ub9cc\ub4e4 \uac83\uc778\uc9c0. VALID\ub294 Padding\uc774 \uc5c6\uace0, SAME\uc740 Padding\uc774 \uc788\uc74c (\uc0ac\uc774\uc988\uc5d0\ub3c4 \uc601\ud5a5\uc744 \uc90c)
 activation: Activation Function\uc744 \ub9cc\ub4e4\uac83\uc778\uc9c0. \ub2f9\uc7a5 \uc124\uc815 \uc548\ud574\ub3c4 Layer\uce35\uc744 \ub530\ub85c \ub9cc\ub4e4 \uc218 \uc788\uc74c<\/strong><\/p>\n\n\n\n
layer = tf.keras.layers.Conv2D(5,3,1,padding='SAME')\n\noutput = layer(image) # image \ubcc0\uc218\ub294 Image.open(path)\uc744 \ud1b5\ud574 \uac00\uc838\uc628 \uc774\ubbf8\uc9c0 \noutput.shape # \uc250\uc774\ud504 \ud655\uc778 (1, 734, 1100, fillter \uac12\uc5d0 \ub530\ub978 \ubcc0\ud654)<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\uc774\ubbf8\uc9c0 \ud655\uc778 \ubc0f \ube44\uad50<\/strong><\/p>\n\n\n\n
plt.subplot(121)\nplt.imshow(image, 'gray')\nplt.subplot(122)\nplt.imshow(output, 'gray')\nplt.colorbar()\nplt.show()<\/code><\/pre>\n\n\n\n
\"\"
\uc804                                                                                                                \ud6c4<\/strong><\/figcaption><\/figure>\n\n\n\n
Activation Function<\/strong><\/p>\n\n\n\n
Convolution\uc5d0\uc11c \uc0ac\uc6a9\ub418\ub294 \uc635\uc158\uc744 \ubbf8\ub9ac \uc815\uc758\ud574\ub454 \ub2e4\uc74c\uc5d0 \uc0ac\uc6a9\uc774 \uac00\ub2a5\ud558\ub2e4.  <\/strong><\/p>\n\n\n\n
\"\"
ReLU \uc635\uc158 \uac12<\/strong><\/figcaption><\/figure>\n\n\n\n
tf.keras.layers.ReLU<\/strong><\/p>\n\n\n\n
\ucd5c\uc18c\uac12, \ucd5c\ub300\uac12 \uad6c\ud558\uae30<\/strong><\/p>\n\n\n\n
\"\"
\uacb0\uacfc \uc804<\/strong><\/figcaption><\/figure>\n\n\n\n
act_layer = tf.keras.layers.ReLU() # 0 \ubbf8\ub9cc\uc758 \uac12\uc744 \uc804\ubd80 0\uc73c\ub85c \uac12\uc744 \ubc14\uafb8\ub294 \uc5d1\ud2f0\ubca0\uc774\uc158 \ud568\uc218\nact_output = act_layer(output)\nact_output.shape<\/code><\/pre>\n\n\n\n
\"\"
\uacb0\uacfc \ud6c4<\/strong><\/figcaption><\/figure>\n\n\n\n
tf.keras.layers.ReLU() \ub294  0 \ubbf8\ub9cc \uc74c\uc218\uac12\uc744 \uc804\ubd80 0\uc73c\ub85c \uac12\uc744 \ubc14\uafb8\ub294 \uc5d1\ud2f0\ubca0\uc774\uc158 \ud568\uc218\uc774\uae30 \ub584\ubb38\uc5d0 \uacb0\uacfc \ud6c4 \ucd5c\uc18c\uac12\uc744 \ubcf4\uba74 0\uc73c\ub85c \ubc14\ub010 \uc0c1\ud0dc\uc744 \ubcfc \uc218 \uc788\ub2e4.<\/strong><\/p>\n\n\n\n
\uc774\ubbf8\uc9c0 \ud655\uc778 \ubc0f \ube44\uad50<\/strong><\/p>\n\n\n\n
plt.figure(figsize=(15,5))\nplt.subplot(121)\nplt.hist(act_output.numpy().ravel(), range=[-2,2])\nplt.ylim(0,10000)\n\nplt.subplot(122)\nplt.imshow(act_output[0,:,:,0])\nplt.colorbar()\nplt.show()<\/code><\/pre>\n\n\n\n
\"\"
\uc774\ubbf8\uc9c0 \ub370\uc774\ud130 \uac12                                                                                     \uc774\ubbf8\uc9c0<\/figcaption><\/figure>\n\n\n\n
<\/p>\n\n\n\n
Pooling<\/strong><\/p>\n\n\n\n
Convolution\uc5d0\uc11c \ud544\ud130\uc744 \ud1b5\ud574 \ub098\ub208 \uc774\ubbf8\uc9c0\ub97c \uc6d0\ud558\ub294 \uc635\uc158 \uac12\uc73c\ub85c \uc555\ucd95\ud558\uace0 \uc81c\uc77c \ud070 \ub370\uc774\ud130 \uac12\ub4e4\ub9cc \ubf51\ub294 \uac83<\/strong><\/p>\n\n\n\n
MaxPool2D : <\/strong>https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/MaxPool2D<\/a><\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
pool_layer = tf.keras.layers.MaxPool2D(2, 2, padding='SAME')\npool_output = pool_layer(act_output)\npool_output.shape<\/code><\/pre>\n\n\n\n
\uc704 \ucf54\ub4dc 2\ubc88\uc9f8 \ub9e4\uac1c\ubcc0\uc218\uc5d0 \ub530\ub77c\uc11c shape\uac12\uc774 \ub2ec\ub77c\uc9c4\ub2e4.                                                                                                                          \uc608) 2\uc77c \uacbd\uc6b0 \u00bd  3\uc77c \uacbd\uc6b0 \u2153,  4\uc77c \uacbd\uc6b0 <\/strong> \u00bc<\/strong><\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\uc774\ubbf8\uc9c0 \ud655\uc778 \ubc0f \ube44\uad50<\/strong><\/p>\n\n\n\n
plt.figure(figsize=(15,5))\nplt.subplot(131)\nplt.hist(pool_output.numpy().ravel(), range=[-2,2])\nplt.ylim(0, 10000)\nplt.subplot(132)\nplt.title(pool_output.shape)\nplt.imshow(pool_output[0, :, :, 0], 'gray')\nplt.subplot(133)\nplt.title(act_output.shape)\nplt.imshow(act_output[0, :, :, 0], 'gray')\nplt.colorbar()\nplt.show()<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\n\n\n\n
Fully Connected<\/h1>\n\n\n\n
\ubaa8\ub378 \uad6c\uc131\ud558\uae30<\/strong><\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
Flatten<\/strong><\/p>\n\n\n\n
\ub370\uc774\ud130 \uc0ac\uc774\uc988\uc744 \uc720\uc9c0\ud558\uba74\uc11c \ucc28\uc6d0 \uc218\uc744 \uc904\uc774\ub294 \uac83 <\/strong><\/p>\n\n\n\n
flatten_layer = tf.keras.layers.Flatten()\nflatten = flatten_layer(output)<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\uc774\ubbf8\uc9c0 \ud655\uc778 \ubc0f \ube44\uad50<\/strong><\/p>\n\n\n\n
plt.figure(figsize=(10, 5))\nplt.subplot(211)\nplt.hist(flatten.numpy().ravel())\nplt.subplot(212)\nplt.imshow(flatten[:,:100])\nplt.show()<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
Dense<\/strong><\/p>\n\n\n\n
\uc6d0\ud558\ub294 \ub370\uc774\ud130 \uac2f\uc218\uc744 \ubc1b\ub294 \uac83<\/strong><\/p>\n\n\n\n
dense_layer = tf.keras.layers.Dense(32, activation='relu')\ndense_output = dense_layer(flatten)\ndense_output.shape<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
dense_layer2 = tf.keras.layers.Dense(64, activation='relu') # 64\uac1c\uc758 relu \uc5d1\ud2f0\ubca0\uc774\uc158 \ud568\uc218\uc5d0 \ub370\uc774\ud130\uac12  \ndense_output2 = dense_layer2(flatten)\ndense_output2.shape<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
DropOut<\/strong><\/p>\n\n\n\n
\ub9e4\ubc88 \uc790\ub3d9\uc73c\ub85c \ud559\uc2b5\ub420\ub54c\ub9c8\ub2e4 \ud559\uc2b5\uc774 \uc548\ub418\ub294 \ubd80\ubd84\uc744 \uc798 \ub418\uac8c \ud574\uc8fc\uc5b4 \uace8\uad6c\ub85c \ud559\uc2b5 \ud558\ub294\uac83<\/strong> <\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
drop_layer = tf.keras.layers.Dropout(0.7)\ndrop_output = drop_layer(dense_output)\ndrop_output.shape<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\n\n\n\n
Build Model<\/strong><\/p>\n\n\n\n
\ubaa8\ub378 \uad6c\uc131 \ud750\ub984 \ud655\uc778\ud558\uae30<\/strong><\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
from tensorflow.keras import layers\n\ninput_shape = (28, 28, 1)\nnum_classes = 10\n\ninputs = layers.Input(input_shape)\nnet = layers.Conv2D(32, (3, 3), padding='SAME')(inputs)\nnet = layers.Activation('relu')(net)\nnet = layers.Conv2D(32, (3, 3), padding='SAME')(net)\nnet = layers.Activation('relu')(net)\nnet = layers.MaxPooling2D(pool_size=(2, 2))(net)\nnet = layers.Dropout(0.25)(net)\n\nnet = layers.Conv2D(64, (3, 3), padding='SAME')(net)\nnet = layers.Activation('relu')(net)\nnet = layers.Conv2D(64, (3, 3), padding='SAME')(net)\nnet = layers.Activation('relu')(net)\nnet = layers.MaxPooling2D(pool_size=(2, 2))(net)\nnet = layers.Dropout(0.25)(net)\n\nnet = layers.Flatten()(net)\nnet = layers.Dense(512)(net)\nnet = layers.Activation('relu')(net)\nnet = layers.Dropout(0.5)(net)\nnet = layers.Dense(num_classes)(net)\nnet = layers.Activation('softmax')(net)\n\nmodel = tf.keras.Model(inputs=inputs, outputs=net, name='Basic_CNN')\n\nmodel.summary()<\/code><\/pre>\n\n\n\n
\"\"<\/figure>\n\n\n\n
<\/p>\n\n\n\n
<\/p>\n","protected":false},"excerpt":{"rendered":"
Feature Extraction \uc774\ubbf8\uc9c0\uc744 \ud2b9\uc9d5, \ud328\ud134\uc73c\ub85c \ub2e4\uc591\ud558\uac8c \ub370\uc774\ud130 \uc14b\uc744 \ud574\ubcf4\uace0 \ubaa8\ub378\uc744 \uad6c\uc131\ud574 \ub370\uc774\ud130 […]<\/p>\n","protected":false},"author":1,"featured_media":883,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[24],"tags":[],"_links":{"self":[{"href":"https:\/\/leenux.kr\/index.php?rest_route=\/wp\/v2\/posts\/884"}],"collection":[{"href":"https:\/\/leenux.kr\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/leenux.kr\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/leenux.kr\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/leenux.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=884"}],"version-history":[{"count":19,"href":"https:\/\/leenux.kr\/index.php?rest_route=\/wp\/v2\/posts\/884\/revisions"}],"predecessor-version":[{"id":1028,"href":"https:\/\/leenux.kr\/index.php?rest_route=\/wp\/v2\/posts\/884\/revisions\/1028"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/leenux.kr\/index.php?rest_route=\/wp\/v2\/media\/883"}],"wp:attachment":[{"href":"https:\/\/leenux.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=884"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/leenux.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=884"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/leenux.kr\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=884"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}