StudentAlone.py

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from tensorflow.keras.utils import plot_model
from tensorflow.keras.models import Sequential, load_model, Model

from tensorflow.keras.layers import Conv2D,GlobalAveragePooling2D,Dense,Softmax,Flatten,MaxPooling2D,Dropout,Activation, Lambda, concatenate
from tensorflow.keras.callbacks import ModelCheckpoint
from tensorflow.keras.losses import kullback_leibler_divergence as KLD_Loss, categorical_crossentropy as logloss
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.metrics import categorical_accuracy
from utils.constants import EPOCHS



# Create the student
def create_StudentAlone(x_train, y_train, x_test, y_test, input_shape, nb_classes, output_directory, filters, filters2, layers=3, separable_conv=False):
    recupereStudentAloneLossAccurayTest1=[]

    inputs = tf.keras.layers.Input(input_shape)
    if separable_conv:
        conv1 = tf.keras.layers.SeparableConv1D(filters=filters, kernel_size=8, padding="same")(inputs)
    else:
        conv1 = tf.keras.layers.Conv1D(filters=filters, kernel_size=8, padding="same")(inputs)
    conv1 = tf.keras.layers.BatchNormalization()(conv1)
    conv1 = tf.keras.layers.ReLU()(conv1)

    if layers>1:
        if separable_conv:
            conv2 = tf.keras.layers.SeparableConv1D(filters=filters2, kernel_size=5, padding="same")(conv1)
        else:
            conv2 = tf.keras.layers.Conv1D(filters=filters2, kernel_size=5, padding="same")(conv1)
        conv2 = tf.keras.layers.BatchNormalization()(conv2)
        conv2 = tf.keras.layers.ReLU()(conv2)
        if layers>2:
            if separable_conv:
                conv3 = tf.keras.layers.SeparableConv1D(filters=filters, kernel_size=3, padding="same")(conv2)
            else:
                conv3 = tf.keras.layers.Conv1D(filters=filters, kernel_size=3, padding="same")(conv2)
            conv3 = tf.keras.layers.BatchNormalization()(conv3)
            conv3 = tf.keras.layers.ReLU()(conv3)
            gap = tf.keras.layers.GlobalAveragePooling1D()(conv3)

        else:
            gap = tf.keras.layers.GlobalAveragePooling1D()(conv2)

    else:
        gap = tf.keras.layers.GlobalAveragePooling1D()(conv1)

    outputs = tf.keras.layers.Dense(nb_classes)(gap)

    studentAlone = keras.Model(inputs=inputs, outputs=outputs, name="studentAlone")

    studentAlone.summary()

    input()
    
    callbacks = [
    keras.callbacks.ModelCheckpoint(
        f"{output_directory}/best_model_student.h5", save_best_only=True, monitor="loss"
    ),
    keras.callbacks.ReduceLROnPlateau(
        monitor="loss", factor=0.5, patience=50, min_lr=0.0001
    ),
      ]
    # Train student as done usually
    studentAlone.compile(
    optimizer=keras.optimizers.Adam(),
    loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True),
    metrics = [tf.keras.metrics.CategoricalAccuracy()]
    )
        
    # Train and evaluate student trained from scratch.
    batch_size = 16    
    mini_batch_size = int(min(x_train.shape[0]/10, batch_size))

    
    history1= studentAlone.fit(x_train, y_train,batch_size=mini_batch_size, epochs=EPOCHS,validation_data=(x_test,y_test),
    callbacks=callbacks,)
    
    resultat = studentAlone.evaluate(x_test, y_test)
    recupereStudentAloneLossAccurayTest1.append(resultat)
    histo_dfstudentAlone = pd.DataFrame(history1.history)
    hist_csv_file = output_directory + '/historystudentAlone'+'.csv'
    with open(hist_csv_file,mode='w' ) as f:
        histo_dfstudentAlone.to_csv(f)
    
    loss1 =history1.history['loss'] 
    val_loss =history1.history['val_loss']
    
    metric1 = "categorical_accuracy"
    
    tech1=recupereStudentAloneLossAccurayTest1
    np.savetxt(output_directory +'tech.out',tech1,delimiter='\t')