fine_tuning_begin.py

# import necessary layers  
from tensorflow.keras.layers import Input, Conv2D , Dropout, MaxPool2D, Flatten, Dense 
from tensorflow.keras import Model
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.regularizers import l2
import tensorflow as tf
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
import os
import matplotlib.pyplot as plt
import sys
from tensorflow.keras.callbacks import CSVLogger


MODEL_FNAME = "trained_model.h5"
base_dir = "dataset"
tmp_model_name = "tmp.h5"
INPUT_SIZE = 224
BATCH_SIZE = 16


physical_devices = tf.config.list_physical_devices()
print("DEVICES : \n", physical_devices)


print('Using:')
print('\t\u2022 Python version:',sys.version)
print('\t\u2022 TensorFlow version:', tf.__version__)
print('\t\u2022 tf.keras version:', tf.keras.__version__)
print('\t\u2022 Running on GPU' if tf.test.is_gpu_available() else '\t\u2022 GPU device not found. Running on CPU')

count = 0
previous_acc = 0


if not os.path.exists(MODEL_FNAME):
    
    """ Create Shallow Model"""
    
    input = Input(shape =(INPUT_SIZE,INPUT_SIZE,3))
    
    weight_initializer = tf.keras.initializers.RandomNormal(mean=0.0, stddev=0.01, seed=None)
    bias_initializer=tf.keras.initializers.Zeros()
     
    x = Conv2D (filters =32, kernel_size =3, padding ='same', activation='relu',kernel_initializer=weight_initializer,bias_initializer=bias_initializer)(input)
    x = MaxPool2D(pool_size =2, strides =2, padding ='same')(x)
      
    x = Conv2D (filters =64, kernel_size =3, padding ='same', activation='relu',kernel_initializer=weight_initializer,bias_initializer=bias_initializer)(x)
    x = MaxPool2D(pool_size =2, strides =2, padding ='same')(x)
    
    x = Conv2D (filters =128, kernel_size =3, padding ='same', activation='relu',kernel_initializer=weight_initializer,bias_initializer=bias_initializer)(x)
    x = MaxPool2D(pool_size =2, strides =2, padding ='same')(x)
    
    x = Conv2D (filters =128, kernel_size =3, padding ='same', activation='relu',kernel_initializer=weight_initializer,bias_initializer=bias_initializer)(x)
    x = MaxPool2D(pool_size =2, strides =2, padding ='same')(x)
   
    x = Flatten()(x)
    x = Dense(units = 64, activation ='relu', kernel_initializer=weight_initializer,bias_initializer=bias_initializer)(x) 
    output = Dense(units=1,activation='sigmoid')(x)
    
    # creating the model
    
    model = Model (inputs=input, outputs =output)
    
    # to be sure GPU memory is cleaned after last train
    m = model
    m.save(tmp_model_name)
    del m
    tf.keras.backend.clear_session()
    
    
    model.summary()
    
    """ Prepare the Dataset for Training"""
    
    
    train_dir = os.path.join(base_dir, 'train')
    val_dir = os.path.join(base_dir, 'validation')
 
    
    
    train_batches = ImageDataGenerator(rescale = 1 / 255.).flow_from_directory(train_dir,
                                                         target_size=(INPUT_SIZE,INPUT_SIZE),
                                                         shuffle=True,
                                                         seed=42,
                                                         class_mode = "binary",
                                                         batch_size=BATCH_SIZE)

    val_batches = ImageDataGenerator(rescale = 1 / 255.).flow_from_directory(val_dir,
                                                         target_size=(INPUT_SIZE,INPUT_SIZE),
                                                         shuffle=True,
                                                         seed=42,
                                                         class_mode = "binary",
                                                         batch_size=BATCH_SIZE)
    
    """ Train """
    
    class CustomLearningRateScheduler(tf.keras.callbacks.Callback):
        def __init__(self, schedule):
            super(CustomLearningRateScheduler, self).__init__()
            self.schedule = schedule
            self.weights_monitor = open("weights.txt", "w+")
            
        
        # learning rate scheduler is called at the end of each epoch. learning rate decreases if needed
        def on_epoch_end(self, epoch, logs=None):
            if not hasattr(self.model.optimizer, "lr"):
                raise ValueError('Optimizer must have a "lr" attribute.')
            # Get the current learning rate from model's optimizer.
            lr = float(tf.keras.backend.get_value(self.model.optimizer.learning_rate))
            # Call schedule function to get the scheduled learning rate.
            # keys = list(logs.keys())
            # print("keys",keys)
            val_acc = logs.get("val_binary_accuracy")
            scheduled_lr = self.schedule(lr, val_acc)
            # Set the value back to the optimizer before this epoch starts
            tf.keras.backend.set_value(self.model.optimizer.lr, scheduled_lr)
         
        #first and last convolutional layers' and dense layer's weights are saved at the beginning of each epoch
        def on_epoch_begin(self, epoch, logs=None):
            epoch_str = "beginning of epoch : "+ str(epoch)
            self.weights_monitor.write(epoch_str)
            self.weights_monitor.write(str(model.get_layer("conv2d").weights))
            self.weights_monitor.write(str(model.get_layer("conv2d_1").weights))
            self.weights_monitor.write(str(model.get_layer("dense_1").weights))
          
          
    def learning_rate_scheduler(lr, val_acc):
        global count
        global previous_acc
      
        if val_acc <= previous_acc:
          #  print("acc ", val_acc, "previous acc ", previous_acc)
            count += 1
        else:
            previous_acc = val_acc
            count = 0
        
        if count >= 10:
            print("acc doesnt improve for 10 epoch, learnin rate decreased by /10")
            count = 0
            lr /= 10
            print("new learning rate:", lr)     
           
        return lr
    
    #compile the model by determining loss function Binary Cross Entropy, optimizer as SGD
    model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.00001),
                  loss=tf.keras.losses.BinaryCrossentropy(),
                  metrics=[tf.keras.metrics.BinaryAccuracy()],
                  sample_weight_mode=[None])
    
    #if validation accuracy doesnt improve for 15 epoch, stop training
    early_stopping = EarlyStopping(monitor='val_binary_accuracy', patience=15)
        
    #save the model if a better validation accuracy then previous better accuracy is obtained  
    checkpointer = ModelCheckpoint(filepath=MODEL_FNAME, verbose=1, save_best_only=True)
    
    # write accuracy and loss history to the log.csv
    csv_logger = CSVLogger('log.csv', append=True, separator=' ')
    
    history=model.fit(train_batches,
        validation_data = val_batches,
        epochs = 100,
        verbose = 1,
        shuffle = True,
        callbacks = [checkpointer,early_stopping,CustomLearningRateScheduler(learning_rate_scheduler),
                     csv_logger])
    
    """ Plot the train and validation Loss """
    
    plt.plot(history.history['loss'])
    plt.plot(history.history['val_loss'])
    plt.title('model loss')
    plt.ylabel('loss')
    plt.xlabel('epoch')
    plt.legend(['train', 'validation'], loc='upper left')
    plt.show()
    
    """ Plot the train and validation Accuracy """
  
    plt.plot(history.history['binary_accuracy'])
    plt.plot(history.history['val_binary_accuracy'])
    plt.title('model accuracy')
    plt.ylabel('accuracy')
    plt.xlabel('epoch')
    plt.legend(['train', 'validation'], loc='upper left')
    plt.show()
  
    print("End of Training")
    
else: 
    
    """ Test """
    
    test_dir = os.path.join(base_dir, 'test')
    
    test_batches = ImageDataGenerator(rescale = 1 / 255.).flow_from_directory(test_dir,
                                                       target_size=(INPUT_SIZE,INPUT_SIZE),
                                                       class_mode='binary',
                                                       shuffle=False,
                                                       seed=42,
                                                       batch_size=1)
    model = tf.keras.models.load_model(MODEL_FNAME)
    model.summary()
    
    # Evaluate on test data
    scores = model.evaluate(test_batches)
    print("metric names",model.metrics_names) 
  
    print(model.metrics_names[0], scores[0])
    print(model.metrics_names[1], scores[1])
        

tf.keras.backend.clear_session()