[Python] MobileNetV3-Large TFLite Model Works in Notebook But Always Predicts the Same Class...

I trained a 46-class image classification model using MobileNetV3-Large with TensorFlow/Keras and saved it as a .keras model. I converted it to TFLite using tf.lite.TFLiteConverter.from_keras_model()

When testing the TFLite model on my notebook, it produces correct predictions. However, when I build my Flutter APK and run the model there, it always outputs the same class regardless of the input image. My previous model using MobileNetV2 worked correctly in Flutter. Here’s the training code and the conversion code I used. I suspect this might be related to differences in preprocessing or model conversion between MobileNetV2 and MobileNetV3-Large. Has anyone encountered this issue or have suggestions on what might be causing the constant output in the Flutter environment?

Below is the model training code

from tensorflow.keras.regularizers import l2
from tensorflow.keras.layers import Dropout
from tensorflow.keras.applications import MobileNetV3Large

l2_strength = 0.01

base_model = MobileNetV3Large(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
base_model.trainable = False # Freeze the base model

x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='swish', kernel_regularizer=l2(l2_strength))(x)
x = Dropout(0.5)(x) # Add dropout
predictions = Dense(46, activation='softmax', kernel_regularizer=l2(l2_strength))(x)

model = Model(inputs=base_model.input, outputs=predictions)

model.compile(optimizer=Adam(learning_rate=0.0005), loss='categorical_crossentropy', metrics=['accuracy'])

from tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping
from tensorflow.keras.regularizers import l2
from tensorflow.keras.layers import Dense, Dropout

reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=3, min_lr=1e-7)
early_stopping = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)

history = model.fit(
train_generator,
steps_per_epoch=len(train_generator),
validation_data=validation_generator,
validation_steps=len(validation_generator),
epochs=50,
callbacks=[reduce_lr, early_stopping]
)
# Accuracy
test_loss, test_accuracy = model.evaluate(test_generator)
print(f'Test Accuracy: {test_accuracy * 100:.2f}%')
# Model saving
model.save('classifier_V3.keras')


Below is how I convert my keras model into tflite

import tensorflow as tf
from tensorflow.keras.models import load_model
model_path = "classifier_V3.keras"
tflite_model_path = "classifier_V3.tflite"

model = load_model(model_path)

converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()

with open(tflite_model_path, 'wb') as f:
f.write(tflite_model)
print(f"TensorFlow Lite model saved to {tflite_model_path}")


If you have any clue please feel free to share

I am using tensorflow 2.18 and keras 3.9

Continue reading...