해당 글은 제로베이스데이터스쿨 학습자료를 참고하여 작성되었습니다

CNN_Feature_Maps

사용 데이터 및 모델

MNIST Load

import tensorflow as tf

mnist = tf.keras.datasets.mnist

(X_train, y_train),(X_test, y_test) = mnist.load_data()
X_train, X_test = X_train / 255.0, X_test / 255.0

X_train = X_train.reshape((60000, 28, 28, 1))
X_test = X_test.reshape((10000, 28, 28, 1))

Simple Modeling

from tensorflow.keras import layers, models

model = models.Sequential([
    layers.Conv2D(3, kernel_size=(3, 3), strides=(1,1), padding="same", activation="relu", input_shape=(28,28,1)),
    layers.MaxPooling2D(pool_size=(2,2), strides=(2,2)),
    layers.Dropout(0.25),
    layers.Flatten(),
    layers.Dense(1000, activation="relu"),
    layers.Dense(10, activation="softmax")              
])

model.summary()
--------------------------------------------------------------------
Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d (Conv2D)             (None, 28, 28, 3)         30        
                                                                 
 max_pooling2d (MaxPooling2D  (None, 14, 14, 3)        0         
 )                                                               
                                                                 
 dropout (Dropout)           (None, 14, 14, 3)         0         
                                                                 
 flatten (Flatten)           (None, 588)               0         
                                                                 
 dense (Dense)               (None, 1000)              589000    
                                                                 
 dense_1 (Dense)             (None, 10)                10010     
                                                                 
=================================================================
Total params: 599,040
Trainable params: 599,040
Non-trainable params: 0
_________________________________________________________________

layer 조회

model.layers
--------------------------------------------------------------------
[<keras.layers.convolutional.conv2d.Conv2D at 0x1428ca6d310>,
 <keras.layers.pooling.max_pooling2d.MaxPooling2D at 0x1428d08ead0>,
 <keras.layers.regularization.dropout.Dropout at 0x1428e790150>,
 <keras.layers.reshaping.flatten.Flatten at 0x1428e7b2a90>,
 <keras.layers.core.dense.Dense at 0x1428e7b3690>,
 <keras.layers.core.dense.Dense at 0x1428e793a10>]

아직 학습하지 않은 Conv 레이어의 weigths

conv = model.layers[0]
conv_weights = conv.weights[0].numpy()
conv_weights.mean(), conv_weights.std()
------------------------------------------
(0.012864082, 0.23187771)

weights 분포도

import matplotlib.pyplot as plt

plt.hist(conv_weights.reshape(-1, 1))
plt.xlabel("weights")
plt.ylabel("count")
plt.show()

학습 전 conv filter

fig, ax = plt.subplots(1, 3, figsize=(15, 5))
for i in range(3):
    ax[i].imshow(conv_weights[:,:,0,i], vmin=-0.5, vmax=0.5)
    ax[i].axis("off")
    
plt.show()

모델 학습

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
hist = model.fit(X_train, y_train, epochs=5, verbose=1, validation_data = (X_test, y_test))

학습 후 conv filter

fig, ax = plt.subplots(1, 3, figsize=(15, 5))
for i in range(3):
    ax[i].imshow(conv_weights[:,:,0,i], vmin=-0.5, vmax=0.5)
    ax[i].axis("off")
    
plt.show()

FeatureMaps

FeatureMaps에 사용할 데이터

plt.imshow(X_train[0], cmap="gray")

FeatureMaps 생성

inputs = X_train[0].reshape(-1, 28, 28, 1)
conv_layer_output = tf.keras.Model(model.input, model.layers[0].output)

feature_maps = conv_layer_output.predict(inputs)

feature_maps

def draw_feature_maps(n):
    inputs = X_train[n].reshape(-1, 28, 28, 1)
    feature_maps = conv_layer_output.predict(inputs)
    
    fig, ax = plt.subplots(1,4, figsize=(15,5))
    ax[0].imshow(feature_maps[0,:,:,0], cmap="gray")
    
    for i in range(1, 4):
        ax[i].imshow(feature_maps[0,:,:,i-1])
        ax[i].axis("Off")
    plt.show()
    
draw_feature_maps(1)

draw_feature_maps(4)

채널 증가

• 입력 수 3 -> 8

from tensorflow.keras import layers, models

model1 = models.Sequential([
    layers.Conv2D(8, kernel_size=(3,3), strides=(1,1), padding="same", activation="relu", input_shape=(28,28,1)),
    layers.MaxPooling2D(pool_size=(2,2), strides=(2,2)),
    layers.Dropout(0.25),
    layers.Flatten(),
    layers.Dense(1000, activation="relu"),
    layers.Dense(10, activation="softmax")              
])

model1.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
hist = model1.fit(X_train, y_train, epochs=5, verbose=1, validation_data = (X_test, y_test))
---------------------------------------------------------------------------------------
Epoch 1/5
1875/1875 [==============================] - 46s 24ms/step - loss: 0.1633 - accuracy: 0.9499 - val_loss: 0.0579 - val_accuracy: 0.9810
Epoch 2/5
1875/1875 [==============================] - 45s 24ms/step - loss: 0.0680 - accuracy: 0.9784 - val_loss: 0.0442 - val_accuracy: 0.9856
Epoch 3/5
1875/1875 [==============================] - 47s 25ms/step - loss: 0.0460 - accuracy: 0.9854 - val_loss: 0.0403 - val_accuracy: 0.9876
Epoch 4/5
1875/1875 [==============================] - 50s 26ms/step - loss: 0.0350 - accuracy: 0.9884 - val_loss: 0.0353 - val_accuracy: 0.9881
Epoch 5/5
1875/1875 [==============================] - 52s 28ms/step - loss: 0.0270 - accuracy: 0.9909 - val_loss: 0.0422 - val_accuracy: 0.9870

feature_maps

conv_layer_output = tf.keras.Model(model1.input, model1.layers[0].output)

def draw_feature_maps(n):
    inputs = X_train[n].reshape(-1, 28, 28, 1)
    feature_maps = conv_layer_output.predict(inputs)
    
    fig, ax = plt.subplots(1,9, figsize=(15,5))
    ax[0].imshow(feature_maps[0,:,:,0], cmap="gray")
    
    for i in range(1, 9):
        ax[i].imshow(feature_maps[0,:,:,i-1])
        ax[i].axis("Off")
    plt.show()
    
draw_feature_maps(1)

draw_feature_maps(9)