[Python] My tensor flow model cannot get further accuracy

i have been testing this library recently, i'm not a good programer now i fugure it out not even with the support of AI i could be able to improve my tensor flow model, no matter what i try it just doesn't work.

ok i'll start with the base structure of my code and what it supuse to do, first i feed the model with a small data set of less of 3k of images of graphs (2d images), the number of classes are two, and the out put is binary.

def scheduler(epoch, lr):
if epoch % 10 == 0 and epoch != 0:
return lr * 0.5
return lr

lr_callback = LearningRateScheduler(scheduler)

early_stop = EarlyStopping(
monitor='val_loss',
patience=50,
restore_best_weights=True,
verbose=1
)

nombre = f"modelo_AD_{datetime.now().strftime('%Y%m%d_%H%M%S')}.h5"

checkpoint = ModelCheckpoint(
filepath=f'/content/drive/MyDrive/modelos_guardados/{nombre}',
monitor='val_accuracy',
save_best_only=True,
verbose=1
)


Also used this one with the help of grok

from tensorflow.keras.callbacks import ReduceLROnPlateau

reduce_lr = ReduceLROnPlateau(
monitor='val_loss',
factor=0.2, # Reduce lr a 20% del valor actual
patience=5, # Espera 5 épocas sin mejora
min_lr=1e-6, # Límite inferior para lr
verbose=1
)

early_stop = EarlyStopping(
monitor='val_loss',
patience=10, # Reducir a 10 épocas
restore_best_weights=True,
verbose=1
)

import os
from tensorflow.keras.callbacks import ModelCheckpoint
from datetime import datetime

nombre = f"modelo_AD_{datetime.now().strftime('%Y%m%d_%H%M%S')}.h5"
os.makedirs('/content/drive/MyDrive/modelos_guardados', exist_ok=True) # Crear directorio si no existe
checkpoint = ModelCheckpoint(
filepath=f'/content/drive/MyDrive/modelos_guardados/{nombre}',
monitor='val_loss', # Cambiar a val_loss
save_best_only=True,
verbose=1
)

X_entrenamiento_AD = X [:2364]
X_validacion_AD = X [2364:]
y_entrenamiento_AD = y [:2364]
y_validacion_AD = y [2364:]


I used this model

modeloCNN2 = tf.keras.models.Sequential([
tf.keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=input_shape),
tf.keras.layers.MaxPooling2D(2, 2),
BatchNormalization(),
tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
tf.keras.layers.MaxPooling2D(2, 2),
BatchNormalization(),
tf.keras.layers.Conv2D(128, (3,3), activation='relu'),
tf.keras.layers.MaxPooling2D(2, 2),
BatchNormalization(),
tf.keras.layers.Dropout(0.5),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(250, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.001)),
BatchNormalization(),
tf.keras.layers.Dense(1, activation='sigmoid')
])
modeloCNN2.compile(
optimizer="Adam",
loss='binary_crossentropy',
metrics=['accuracy']
)


and also this one

#EXPERIMENTAL

modeloCNN2 = tf.keras.models.Sequential([

tf.keras.layers.Conv2D(256, (3,3), activation='relu', input_shape=input_shape, padding="same"),
tf.keras.layers.MaxPooling2D(2, 2),
BatchNormalization(),

tf.keras.layers.Conv2D(128, (3,3), activation='relu', padding="same"),
tf.keras.layers.MaxPooling2D(2, 2),
BatchNormalization(),

Conv2D(16, (3,3), activation='relu', padding="same"),
AveragePooling2D(2,2),
BatchNormalization(),


tf.keras.layers.Dropout(0.3),
tf.keras.layers.Flatten(),

tf.keras.layers.Dense(128, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.001)),
BatchNormalization(),

tf.keras.layers.Dense(64, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.001)),
BatchNormalization(),
tf.keras.layers.Dense(1, activation='sigmoid')
])

modeloCNN2.compile(
optimizer="Adam",
loss='binary_crossentropy',
metrics=['accuracy']
)


And i started the training with this code in both cases, changing the callbacks depending of what callback block i preferred, in order to test i used both

tensorboardCNN2 = TensorBoard(log_dir='logs/cnn2_AD')
modeloCNN2.fit(data_gen_entrenamiento,
batch_size=32,
validation_data=(X_validacion_AD, y_validacion_AD),
steps_per_epoch=int(np.ceil(len(X_entrenamiento_AD) / float(32))),
validation_steps=int(np.ceil(len(X_validacion_AD) / float(32))),
epochs=100,
callbacks=[tensorboardCNN2, checkpoint, reduce_lr, early_stop])


So the problem is that the result is always the same, High training value accuracy, low loss in training, mid (between 50% of val accuracy) and mid (between 50% of val loss), in both cases, i tested other optimizer instead of adam, but they just get it worst. At first i started with a data set of 20k images but the results were worst, so i started over with a smaller a newer data set that was validated to avoid redundant data.

Also i decreased and increased the layers but it just make the training faster or slower, the best val accuracy i got was 60% with a val loss of 50%, no matter what i move i just can have better results.

Continue reading...