I am getting worst accuracy on the same dataste and 3 different models

[nazzal5448]

I have coded these 3 models and all of them are performing worst (less than 1% accuracy mostly) on the 101_food_classes_10_percent dataset. I dont know whats the issue. i have fine tuned two of them but the results did not improve.

train_data_10_percent=tf.keras.preprocessing.image_dataset_from_directory(train_dir_10_percent,
                                                                          label_mode='categorical',
                                                                          image_size=(224,224),
                                                                          batch_size=32)
test_data=tf.keras.preprocessing.image_dataset_from_directory(test_dir,
                                                              label_mode='categorical',
                                                              image_size=(224,224),
                                                              batch_size=32,
                                                              shuffle=False)

aug_model=tf.keras.Sequential([
    preprocessing.RandomFlip('horizontal'),
    preprocessing.RandomRotation(0.2),
    preprocessing.RandomHeight(0.2),
    preprocessing.RandomWidth(0.2),
    preprocessing.RandomZoom(0.2)

],name='augmentation_model')

base_model=tf.keras.applications.MobileNet(include_top=False)
inputs=tf.keras.layers.Input(shape=(224,224,3),name='input_layer')
base_model.trainable=False
x=aug_model(inputs)
x=base_model(x,training=False)
x=tf.keras.layers.GlobalAveragePooling2D(name='global_average_pooling_2d')(x)
outputs=tf.keras.layers.Dense(101, activation='softmax',name='output_layer')(x)
model_1=tf.keras.Model(inputs,outputs)


model_1.compile(loss='categorical_crossentropy',
                optimizer=tf.keras.optimizers.Adam(),
                metrics=['accuracy'])

model_checkpoint=tf.keras.callbacks.ModelCheckpoint('checkpoints/model1.ckpt', monitor='val_accuracy',save_best_only=True,
                                                    save_weights_only=True, verbose=1)
initial_epochs=10
history_1=model_1.fit(train_data_10_percent,
                      epochs=initial_epochs,
                      steps_per_epoch=len(train_data_10_percent),
                      validation_data=test_data,
                      validation_steps=int(0.15*len(test_data)),
                      callbacks=[model_checkpoint,
                                 create_tensorboard_callback('logs/','model_1')])

model_1.evaluate(test_data)

## **Another model with 5 epochs + 5 fine-tuning epochs**

base_model_1=tf.keras.applications.MobileNet(include_top=False)
inputs=tf.keras.layers.Input(shape=(224,224,3),name='input_layer')
base_model_1.trainable=False
x=aug_model(inputs)
x=base_model_1(x,training=False)
x=tf.keras.layers.GlobalAveragePooling2D(name='global_average_pooling_2d')(x)
outputs=tf.keras.layers.Dense(101, activation='softmax',name='output_layer')(x)
model_2=tf.keras.Model(inputs,outputs)

model_2.compile(loss='categorical_crossentropy',
                optimizer=tf.keras.optimizers.Adam(),
                metrics=['accuracy'])

model_checkpoint=tf.keras.callbacks.ModelCheckpoint('checkpoints/model2.ckpt', monitor='val_accuracy',save_best_only=True,
                                                    save_weights_only=True, verbose=1)
initial_epochs=5
history_2=model_2.fit(train_data_10_percent,
                      epochs=initial_epochs,
                      steps_per_epoch=len(train_data_10_percent),
                      validation_data=test_data,
                      validation_steps=int(0.15*len(test_data)),
                      callbacks=[model_checkpoint,
                                 create_tensorboard_callback('logs/','model_2')])

base_model_1.trainable=True
for layer in base_model_1.layers[:-5]:
  layer.trainable=False

model_2.compile(loss='categorical_crossentropy',
                optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),
                metrics=['accuracy'])

model_checkpoint=tf.keras.callbacks.ModelCheckpoint('checkpoints/model3.ckpt', monitor='val_accuracy',save_best_only=True,
                                                    save_weights_only=True, verbose=1)
initial_epochs=5
history_3=model_2.fit(train_data_10_percent,
                      epochs=initial_epochs+5,
                      steps_per_epoch=len(train_data_10_percent),
                      validation_data=test_data,
                      validation_steps=int(0.15*len(test_data)),
                      initial_epoch=history_2.epoch[-1],
                      callbacks=[model_checkpoint,
                                 create_tensorboard_callback('logs/','model_3')])

from keras.api._v2.keras import mixed_precision
base_model_3=tf.keras.applications.ResNet50(include_top=False)
inputs=tf.keras.layers.Input(shape=(224,224,3),name='input_layer')
base_model_3.trainable=False
base_model_3.mixed_precision=True
x=aug_model(inputs)
x=base_model_3(x,training=False)
x=tf.keras.layers.GlobalAveragePooling2D(name='global_average_pooling_2d')(x)
outputs=tf.keras.layers.Dense(101, activation='softmax',name='output_layer')(x)
model_4=tf.keras.Model(inputs,outputs)

model_4.compile(loss='categorical_crossentropy',
                optimizer=tf.keras.optimizers.Adam(),
                metrics=['accuracy'])
model_checkpoint=tf.keras.callbacks.ModelCheckpoint('checkpoints/model4.ckpt', monitor='val_accuracy',save_best_only=True,
                                                    save_weights_only=True, verbose=1)
initial_epochs=5
history_4=model_4.fit(train_data_10_percent,
                      epochs=initial_epochs,
                      steps_per_epoch=len(train_data_10_percent),
                      validation_data=test_data,
                      validation_steps=int(0.15*len(test_data)),
                      callbacks=[model_checkpoint,
                                 create_tensorboard_callback('logs/','model_4')])

I have the idea about them overfitting but they dont't even perform on the training data.