import tensorflow as tf
from tensorflow import keras

# dataset utilizzato https://keras.io/api/datasets/fashion_mnist/#load_data-function

mnist = keras.datasets.fashion_mnist.load_data()
#vettori per trainare l'ia e per testarla
(x_train,y_train),(x_test,y_test) = mnist

import numpy as np

#min = 0 e max = 255
np.min(x_train), np.min(x_train)

#categorie vestiti
class_names = ['top', 'trouser', 'pullover', 'dress', ' coat', 'sandal', 'shirt', 'sneaker', 'bag', 'ankle boot']

#x_train e' una matrice e dobbiamo normalizzarlo (dividendolo per il max value)

X_train = x_train/255.0
X_test = x_test/255.0

from keras import Sequential as Sequ
from tensorflow.keras.layers import Flatten, Dense

model = Sequ()
#aggiungo il layer Flatten con la forma 28 x 28 pixel
model.add(Flatten(input_shape=(28,28)))
#aggiungo 128 neuroni
model.add(Dense(128, activation='relu'))
model.add(Dense(256, activation='relu'))
#qui aggiungo 10 neuroni perche' abbiamo 10 categorie
#con softmax activation -> perche' facciamo un multiclass classification (piu' di 2 classi)
model.add(Dense(10, activation='softmax'))

print(model.summary())

#model compilation
# loss function, optimazer, metrics

#utilizziamo sparse_categorical_crossentropy perche'
#il nostro output e' 0,1,2,3,4,5 -> sparse_categorical_crossentropy
#il nostro output e' one hot encoded -> categorical_entropy
# 0,1 -> binary_cross_entropy
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

#Passiamo i dati per il training e li dividiamo in 10 'epoche' passando 32 dati ogni epoca
model_checkpoint_best_callback = keras.callbacks.ModelCheckpoint(
    filepath = '../my_models/fashion_model.keras',
    monitor = 'val_loss',
    verbose = 1,
    save_best_only = False,
    save_weights_only = False
)
model.fit(X_train, y_train, epochs=10, batch_size=32, callbacks=[model_checkpoint_best_callback])