我的模型的验证精度卡住了！ 我如何解决它？ [张量流/Keras]

Question

I am trying to train a Regression model for a series of images on my Local PC running on an Nvidia RTX 2070 Super.

Input image Shape: (96,160,3), 16072 Samples (12857 for training, 3215 for validation)

However, the training is not taking place. I tried using a custom model, as mentioned in this reference .

model = Sequential()
model.add(Conv2D(input_shape=(96, 320, 3), filters=32, kernel_size=3, padding="valid"))
model.add(MaxPooling2D(pool_size=(3,3)))
model.add(Activation('relu'))

model.add(Conv2D(filters=256, kernel_size=3, padding="valid"))
model.add(MaxPooling2D(pool_size=(3,3)))
model.add(Activation('relu'))

model.add(Conv2D(filters=512, kernel_size=3, padding="valid"))
model.add(MaxPooling2D(pool_size=(3,3)))
model.add(Activation('relu'))

model.add(Flatten())
model.add(Dense(2048))
model.add(Activation('relu'))
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dense(128))
model.add(Activation('relu'))
model.add(Dropout(0.25))
model.add(Dense(1))

checkpoint = ModelCheckpoint(filepath="./ckpts/model.ckpt", monitor='val_loss', save_best_only=True)
stopper = EarlyStopping(monitor='val_acc', min_delta=0.0003, patience = 10)

lr_schedule = ExponentialDecay(initial_learning_rate=0.1, decay_steps=10000, decay_rate=0.9)
optimizer = Adam(learning_rate=lr_schedule)
loss = Huber(delta=0.5, reduction="auto", name="huber_loss")
model.compile(loss = loss, optimizer = optimizer, metrics=['accuracy'])
model.fit(X_train, y_train, validation_split = 0.2, shuffle = True, epochs = 100, 
          callbacks=[checkpoint, stopper])

The training went on for 10 epochs before automatically terminating after exhausting patience of Early Stopping Algorithm. Essentially, the model did not train. The validation accuracy was stuck at 0.5353.

Then, suspecting something may be wrong with my model itself, I tried Transfer Learning, importing pre-trained Inception V3 with Imagenet weights, and tried training with frozen weights (only training the last two layers with the pre-trained weights frozen).

from tensorflow.keras import Sequential
from tensorflow.keras.layers import Input, Lambda, Dense, Flatten, Conv2D, Activation, MaxPooling2D, 
      Dropout, GlobalAveragePooling2D
from tensorflow.keras.applications.inception_v3 import InceptionV3, preprocess_input
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
from tensorflow.keras.models import Model
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.losses import Huber
from tensorflow.keras.optimizers.schedules import ExponentialDecay

X_train = preprocess_input(X_train)
inception = InceptionV3(weights='imagenet', include_top=False, input_shape=(299,299,3))
inception.trainable = False

driving_input = Input(shape=(96,320,3))
resized_input = Lambda(lambda image: tf.image.resize(image,(299,299)))(driving_input)
inp = inception(resized_input)

x = GlobalAveragePooling2D()(inp)
x = Dense(512, activation = 'relu')(x)
result = Dense(1, activation = 'relu')(x)


model = Model(inputs = driving_input, outputs = result)
model.compile(optimizer='Adam', loss='mse', metrics = ['accuracy'])

checkpoint = ModelCheckpoint(filepath="./ckpts/model.ckpt", monitor='val_loss', save_best_only=True)
stopper = EarlyStopping(monitor='val_acc', min_delta=0.0003, patience = 10)

batch_size = 32
epochs = 100

model.fit(x=X_train, y=y_train, shuffle=True, validation_split=0.2, epochs=epochs, 
          batch_size=batch_size, verbose=1, callbacks=[checkpoint, stopper])

The result I got did not solve my issue.

I finally also tried unfreezing the weights and retraining the entire InceptionV3 Network, initialized with the Imagenet pre-trained weights.

inception.trainable = True

This also did not produce any change in my model's performance. The validation accuracy is still very much stuck at the same place.

I don't understand why my model is not training. Please help me solve this issue. Do note that I do not expect help with improving my model in general, I can do that myself, but what I can't do is getting the model unstuck. Unless that happens, I can't proceed with any fine tuning or optimization. Any help would be very appreciated.

Answer 1

Thanks to Dr. Snoopy for pointing out the fundamental flaw that accuracy is a classification metric. I switched to Validation Loss using the following modified code:

checkpoint = ModelCheckpoint(filepath="./ckpts/model.h5", monitor='val_loss', 
                             save_best_only=True)
stopper = EarlyStopping(monitor='val_loss', min_delta=0.0003, patience = 10)

lr_schedule = ExponentialDecay(initial_learning_rate=0.1, decay_steps=10000, 
                                decay_rate=0.9)
optimizer = Adam(learning_rate=lr_schedule)
loss = Huber(delta=0.5, reduction="auto", name="huber_loss")
model.compile(loss = loss, optimizer = optimizer)
model.fit(X_train, y_train, validation_split = 0.2, shuffle = True, epochs = 100, 
          callbacks=[checkpoint, stopper])

model.load_weights('./ckpts/model.h5')
model.save('model.h5')

The result was an unstuck model: