如题，在调试各种超参数后，改变隐藏层结构，准确度还是无法提升至7成以上。
将资料丢入迁移式学习，准确度可达到0.95，确认过这个资料集是没问题的。
资料集：https://www.kaggle.com/datasets/biancaferreira/african-wildlife
资料集分为，60% Train 20% Validation 20% Test (901張train 300張validation 300張test)
希望大家能帮忙解答下，这到底是参数问题还是模型问题。
还有个问题就是当train set被归一化或标准化后，val loss没法收敛，会从个位数升到十位或百位数，val acc也升不上去。
from keras.preprocessing.image import ImageDataGenerator
train_datagen=ImageDataGenerator( #資料預處理
horizontal_flip=True,
width_shift_range=0.1,
height_shift_range=0.1,
zoom_range=0.2
)
os.makedirs("/content/train/")
train_generator=train_datagen.flow(tr,tr_la,
batch_size=16
)
model = models.Sequential()
model.add(layers.Conv2D(64, (3,3),strides=2,input_shape=(256, 256, 3)))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3),strides=2))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(256, (3, 3),strides=2))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten())
model.add(layers.Dense(64, kernel_regularizer=regularizers.l2(0.001), activation='relu'))
model.add(layers.Dropout(0.25))
model.add(layers.Dense(16, kernel_regularizer=regularizers.l2(0.001), activation='relu'))
model.add(layers.Dropout(0.25))
model.add(layers.Dense(4, activation='softmax'))
model.summary()
Earlystopping=tf.keras.callbacks.EarlyStopping(
monitor='val_loss', min_delta=0,
patience=10, verbose=1, mode='max',
baseline=None, restore_best_weights=False
)
model.compile(optimizer='Adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False),
metrics=['accuracy'])
initial_learning_rate = 0.001
def lr_step_decay(epoch, lr):
drop_rate = 0.1
epochs_drop =5.0
return initial_learning_rate * math.pow(drop_rate, math.floor(epoch/epochs_drop))
history = model.fit(train_generator, epochs=15,
callbacks=[LearningRateScheduler(lr_step_decay, verbose=1),Earlystopping],
validation_data=(va,va_la))
用以上參數跑過
Epoch 1: LearningRateScheduler setting learning rate to 0.001.
Epoch 1/15
57/57 [==============================] - 15s 254ms/step - loss: 1.6336 - accuracy: 0.3263 - val_loss: 1.8376 - val_accuracy: 0.2900 - lr: 0.0010
Epoch 2: LearningRateScheduler setting learning rate to 0.001.
Epoch 2/15
57/57 [==============================] - 14s 252ms/step - loss: 1.3830 - accuracy: 0.4340 - val_loss: 2.3658 - val_accuracy: 0.3533 - lr: 0.0010
Epoch 3: LearningRateScheduler setting learning rate to 0.001.
Epoch 3/15
57/57 [==============================] - 14s 250ms/step - loss: 1.3083 - accuracy: 0.4528 - val_loss: 1.2480 - val_accuracy: 0.5167 - lr: 0.0010
Epoch 4: LearningRateScheduler setting learning rate to 0.001.
Epoch 4/15
57/57 [==============================] - 14s 250ms/step - loss: 1.2538 - accuracy: 0.4861 - val_loss: 1.1332 - val_accuracy: 0.5700 - lr: 0.0010
Epoch 5: LearningRateScheduler setting learning rate to 0.001.
Epoch 5/15
57/57 [==============================] - 14s 250ms/step - loss: 1.1707 - accuracy: 0.5572 - val_loss: 1.2636 - val_accuracy: 0.5467 - lr: 0.0010
Epoch 6: LearningRateScheduler setting learning rate to 0.0001.
Epoch 6/15
57/57 [==============================] - 14s 252ms/step - loss: 1.1375 - accuracy: 0.5339 - val_loss: 1.0421 - val_accuracy: 0.6033 - lr: 1.0000e-04
Epoch 7: LearningRateScheduler setting learning rate to 0.0001.
Epoch 7/15
57/57 [==============================] - 14s 251ms/step - loss: 1.0945 - accuracy: 0.5583 - val_loss: 0.9792 - val_accuracy: 0.6400 - lr: 1.0000e-04
Epoch 8: LearningRateScheduler setting learning rate to 0.0001.
Epoch 8/15
57/57 [==============================] - 14s 250ms/step - loss: 1.1035 - accuracy: 0.5405 - val_loss: 0.9621 - val_accuracy: 0.6433 - lr: 1.0000e-04
Epoch 9: LearningRateScheduler setting learning rate to 0.0001.
Epoch 9/15
57/57 [==============================] - 14s 251ms/step - loss: 1.1143 - accuracy: 0.5505 - val_loss: 0.9472 - val_accuracy: 0.6367 - lr: 1.0000e-04
Epoch 10: LearningRateScheduler setting learning rate to 0.0001.
Epoch 10/15
57/57 [==============================] - 14s 249ms/step - loss: 1.0645 - accuracy: 0.5605 - val_loss: 0.9538 - val_accuracy: 0.6433 - lr: 1.0000e-04
Epoch 11: LearningRateScheduler setting learning rate to 1.0000000000000003e-05.
Epoch 11/15
57/57 [==============================] - 14s 249ms/step - loss: 1.0316 - accuracy: 0.5594 - val_loss: 0.9564 - val_accuracy: 0.6500 - lr: 1.0000e-05
Epoch 12: LearningRateScheduler setting learning rate to 1.0000000000000003e-05.
Epoch 12/15
57/57 [==============================] - 16s 276ms/step - loss: 1.0271 - accuracy: 0.6049 - val_loss: 0.9626 - val_accuracy: 0.6500 - lr: 1.0000e-05
Epoch 12: early stopping
用了迁移学习？还是咋的？