[Python] Why the loss values of class type and function type of tensorflow models are...

I made two different type models with tensorflow, one is Class type, the other is function type for learning Fashion MNIST. I think the structure of two models is exactly same, but the results are not.

What is the difference of two models?, why the values are different? and how can I make these two models exactly same?

The code is like below:

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

import keras
import tensorflow as tf

from sklearn.model_selection import train_test_split

keras.utils.set_random_seed(42)
tf.random.set_seed(42)

(train_input, train_target), (test_input, test_target) = \
keras.datasets.fashion_mnist.load_data()

train_scaled = train_input / 255.0

train_scaled, val_scaled, train_target, val_target = train_test_split(
train_scaled, train_target, test_size=0.2, random_state=42)

class TFModel(keras.Model):
def __init__(self, a_layer=None):
super(TFModel, self).__init__()
self.flatten = keras.layers.Flatten()
self.dense1 = keras.layers.Dense(100, activation='relu')
self.option = a_layer
self.dense2 = keras.layers.Dense(10, activation='softmax')

def call(self, x):
x = self.flatten(x)
x = self.dense1(x)
if self.option is not None:
x = self.option(x, training=True)
x = self.dense2(x)
return x

def model_fn(a_layer=None):
model = keras.Sequential()
model.add(keras.layers.Input(shape=(28,28)))
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(100, activation='relu'))
if a_layer:
model.add(a_layer)
model.add(keras.layers.Dense(10, activation='softmax'))
return model

dropout = keras.layers.Dropout(0.3)

# model = model_fn(dropout)
# model2 = TFModel(dropout)

model = model_fn()
model2 = TFModel()

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model2.compile(optimizer='adam', loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
history = model.fit(train_scaled, train_target, epochs=20, verbose=0,
validation_data=(val_scaled, val_target))
history2 = model2.fit(train_scaled, train_target, epochs=20, verbose=0,
validation_data=(val_scaled, val_target))

plt.plot(history.history['loss'], label='train')
plt.plot(history.history['val_loss'], label='val')
# plt.xlabel('epoch')
# plt.ylabel('loss')
# plt.legend()
# plt.show()

plt.plot(history2.history['loss'], label='train2')
plt.plot(history2.history['val_loss'], label='val2')
plt.xlabel('epoch')
plt.ylabel('loss')
plt.legend()
plt.show()


The result is like this:

without dropout:



with dropout:



I think both models should show same train loss and validation loss graph.

Continue reading...