在以下用於分類的 ResNet50 模型中的 0 驗證准確率背后的推理

Question

我編寫了這段代碼來使用 Resnet50 對貓和狗進行分類。 實際上，在學習過程中，我得出的結論是，遷移學習為深度學習模型提供了非常好的准確性，但最終得到的結果要差得多，而且我不明白其原因。 任何帶有推理的描述都會非常有幫助。 該數據集包含 2000 張貓狗圖像作為訓練，1000 張圖像作為驗證集。

下面總結了我的模型

from tensorflow.keras.applications import ResNet50
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, InputLayer, Flatten, GlobalAveragePooling2D
num_classes = 2
IMG_SIZE = 224
IMG_SHAPE = (IMG_SIZE, IMG_SIZE, 3)
my_new_model=tf.keras.applications.ResNet50(include_top=False, weights='imagenet', input_shape=IMG_SHAPE, pooling='avg', classes=2)
my_new_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])


from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.applications.resnet50 import preprocess_input
train_datagen = ImageDataGenerator(
 preprocessing_function=preprocess_input,
 rotation_range=40,
 width_shift_range=0.2,
 height_shift_range=0.2,
 shear_range=0.2,
 zoom_range=0.2,
 horizontal_flip=True,)

# Note that the validation data should not be augmented!
test_datagen = ImageDataGenerator(preprocessing_function=preprocess_input)

train_generator = train_datagen.flow_from_directory(
     train_dir,  # This is the source directory for training images
     target_size=(224,224),  # All images will be resized to 224x224
     batch_size=20,
     class_mode='binary')

validation_generator = test_datagen.flow_from_directory(
     validation_dir,
     target_size=(224, 224),
     class_mode='binary')

my_new_model.fit_generator(
     train_generator,
     epochs = 8,
     steps_per_epoch=100,
     validation_data=validation_generator)

為此，我將訓練日志作為，

Train for 100 steps, validate for 32 steps
Epoch 1/8
100/100 - 49s - loss: 7889.4051 - accuracy: 0.0000e+00 - val_loss: 7834.5318 - val_accuracy: 0.0000e+00
Epoch 2/8
100/100 - 35s - loss: 7809.7583 - accuracy: 0.0000e+00 - val_loss: 7775.1556 - val_accuracy: 0.0000e+00
Epoch 3/8
100/100 - 35s - loss: 7808.4858 - accuracy: 0.0000e+00 - val_loss: 7765.3964 - val_accuracy: 0.0000e+00
Epoch 4/8
100/100 - 35s - loss: 7808.0520 - accuracy: 0.0000e+00 - val_loss: 7764.0735 - val_accuracy: 0.0000e+00
Epoch 5/8
100/100 - 35s - loss: 7807.7891 - accuracy: 0.0000e+00 - val_loss: 7762.4891 - val_accuracy: 0.0000e+00
Epoch 6/8
100/100 - 35s - loss: 7807.6872 - accuracy: 0.0000e+00 - val_loss: 7762.1766 - val_accuracy: 0.0000e+00
Epoch 7/8
100/100 - 35s - loss: 7807.6633 - accuracy: 0.0000e+00 - val_loss: 7761.9766 - val_accuracy: 0.0000e+00
Epoch 8/8
100/100 - 35s - loss: 7807.6514 - accuracy: 0.0000e+00 - val_loss: 7761.9346 - val_accuracy: 0.0000e+00
<tensorflow.python.keras.callbacks.History at 0x7f5adff722b0>

Answer 1

當我復制您的模型時，在 resnet（即 Base_net）和密集層之間添加了平坦層，如下所示

Base_net = ResNet50(include_top = False, weights = 'imagenet', input_shape = input_shape)

model = Sequential()
model.add(Base_net)
model.add(Flatten())
model.add(Dense(units = num_classes, activation = 'softmax'))
model.summary()

model.compile(loss = 'categorical_crossentropy',
            optimizer = 'Adam',
            metrics = ['accuracy'])

輸出：

Layer (type)                 Output Shape              Param #   
=================================================================
resnet50 (Model)             (None, 7, 7, 2048)        23587712  
_________________________________________________________________
flatten_5 (Flatten)          (None, 100352)            0         
_________________________________________________________________
dense_9 (Dense)              (None, 1)                 100353    
=================================================================
Total params: 23,688,065
Trainable params: 23,634,945
Non-trainable params: 53,120
_________________________________________________________________
Epoch 1/3
100/100 [==============================] - 84s 837ms/step - loss: 0.0000e+00 - accuracy: 0.5085 - val_loss: 0.0000e+00 - val_accuracy: 0.5000
Epoch 2/3
100/100 [==============================] - 81s 806ms/step - loss: 0.0000e+00 - accuracy: 0.4953 - val_loss: 0.0000e+00 - val_accuracy: 0.5050
Epoch 3/3
100/100 [==============================] - 80s 801ms/step - loss: 0.0000e+00 - accuracy: 0.4965 - val_loss: 0.0000e+00 - val_accuracy: 0.5080

要解決您的問題，您必須將損失函數更改為binary_crossentropy並在最后一個密集層中用sigmoid替換激活函數，從而解決您的問題。 為了獲得更好的精度值，請實施超參數調整。

注意：我們通常觀察到batch size為2的冪，這是因為優化矩陣運算庫的有效工作

請參考下面的工作代碼，其中准確性

%tensorflow_version 2.x

import tensorflow as tf
from tensorflow.keras.applications import ResNet50
import os
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.layers import Dense, Dropout, Flatten
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.models import Sequential

from google.colab import drive
drive.mount('/content/drive')

train_dir  = '/content/drive/My Drive/Dogs_Vs_Cats/train'
test_dir = '/content/drive/My Drive/Dogs_Vs_Cats/test'

img_width, img_height = 224, 224
input_shape = (img_width, img_height, 3)

#train_samples = 2000
#test_samples = 1000
epochs = 8
batch_size = 32

train_datagen = ImageDataGenerator(
    rescale = 1. /255,
    rotation_range=40,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range = 0.2,
    zoom_range = 0.2,
    horizontal_flip = True)

test_datagen = ImageDataGenerator(
    rescale = 1. /255)

train_data = train_datagen.flow_from_directory(
    train_dir,
    target_size = (img_width, img_height),
    batch_size = batch_size,
    class_mode = 'binary')

test_data = test_datagen.flow_from_directory(
    test_dir,
    target_size = (img_width, img_height),
    #batch_size = batch_size,
    class_mode = 'binary')



# The Convolutional Base of the Pre-Trained Model will be added as a Layer in this Model

Conv_Base = ResNet50(include_top = False, weights = 'imagenet', input_shape = input_shape)

model = Sequential()
model.add(Conv_Base)
model.add(Flatten())
model.add(Dense(units = 256, activation = 'relu'))
model.add(Dropout(0.5))
model.add(Dense(units = 1, activation = 'sigmoid'))
model.summary()

model.compile(loss = 'binary_crossentropy',
            optimizer = 'Adam',
            metrics = ['accuracy'])

model.fit_generator(
        train_data,
        steps_per_epoch = 100,
        epochs = 8,
        validation_data = test_data,
        verbose = 1,
        validation_steps = 32)

輸出：

Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.

Model: "sequential_4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
resnet50 (Model)             (None, 7, 7, 2048)        23587712  
_________________________________________________________________
flatten_4 (Flatten)          (None, 100352)            0         
_________________________________________________________________
dense_8 (Dense)              (None, 256)               25690368  
_________________________________________________________________
dropout_4 (Dropout)          (None, 256)               0         
_________________________________________________________________
dense_9 (Dense)              (None, 1)                 257       
=================================================================
Total params: 49,278,337
Trainable params: 49,225,217
Non-trainable params: 53,120
_________________________________________________________________
Epoch 1/8
100/100 [==============================] - 1828s 18s/step - loss: 1.9865 - accuracy: 0.6401 - val_loss: 0.7398 - val_accuracy: 0.5000
Epoch 2/8
100/100 [==============================] - 65s 646ms/step - loss: 0.6642 - accuracy: 0.6951 - val_loss: 0.7094 - val_accuracy: 0.5000
Epoch 3/8
100/100 [==============================] - 65s 652ms/step - loss: 0.5662 - accuracy: 0.7673 - val_loss: 0.7088 - val_accuracy: 0.4950
Epoch 4/8
100/100 [==============================] - 65s 647ms/step - loss: 0.4362 - accuracy: 0.8330 - val_loss: 0.7026 - val_accuracy: 0.4950
Epoch 5/8
100/100 [==============================] - 65s 649ms/step - loss: 0.3998 - accuracy: 0.8405 - val_loss: 0.8281 - val_accuracy: 0.5070
Epoch 6/8
100/100 [==============================] - 65s 647ms/step - loss: 0.5733 - accuracy: 0.7986 - val_loss: 0.8699 - val_accuracy: 0.4970
Epoch 7/8
100/100 [==============================] - 65s 647ms/step - loss: 0.6359 - accuracy: 0.7860 - val_loss: 0.7071 - val_accuracy: 0.5020
Epoch 8/8
100/100 [==============================] - 65s 653ms/step - loss: 0.5514 - accuracy: 0.7673 - val_loss: 0.6590 - val_accuracy: 0.5910