Notebook 07: Cloned model with loaded weights performs poor in Food Vision Milestone project

[reshma93s]

Hi,

I am working on Food Vision Milestone project. I was able to get a model similar to the tutorial while fitting the test data and saved the checkpoints for weights as shown below.

#############################################################################
history_101_food_classes_feature_extract = model.fit(train_data,
                                                     epochs=3,
                                                     steps_per_epoch=(len(train_data)),
                                                     validation_data=test_data,
                                                     validation_steps=int(0.15*len(test_data)),
                                                     callbacks=[create_tensorboard_callback(dir_name="training_logs",
                                                                                            experiment_name="efficientnetb0_101"),
                                                                model_checkpoint])

Saving TensorBoard log files to: training_logs/efficientnetb0_101/20220720-053759
Epoch 1/3
2368/2368 [==============================] - 170s 71ms/step - loss: 1.1328 - accuracy: 0.7048 - val_loss: 1.0842 - val_accuracy: 0.7052
Epoch 2/3
2368/2368 [==============================] - 170s 71ms/step - loss: 1.0383 - accuracy: 0.7292 - val_loss: 1.0708 - val_accuracy: 0.7055
Epoch 3/3
2368/2368 [==============================] - 169s 71ms/step - loss: 0.9670 - accuracy: 0.7449 - val_loss: 1.0768 - val_accuracy: 0.7076
###############################################################################

results_feature_extract_model = model.evaluate(test_data)
results_feature_extract_model

790/790 [==============================] - 50s 63ms/step - loss: 1.0642 - accuracy: 0.7098
[1.0641952753067017, 0.7097821831703186]

###############################################################################

Now, when I clone the model and load the weights, it is giving me very low accuracy. Can you please help me with it?

###############################################################################
cloned_model = tf.keras.models.clone_model(model)
cloned_model.summary()

Model: "model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_2 (InputLayer)        [(None, 224, 224, 3)]     0         
                                                                 
 efficientnetb0 (Functional)  (None, None, None, 1280)  4049571  
                                                                 
 global_average_pooling2d (G  (None, 1280)             0         
 lobalAveragePooling2D)                                          
                                                                 
 dense (Dense)               (None, 101)               129381    
                                                                 
 activation (Activation)     (None, 101)               0         
                                                                 
=================================================================
Total params: 4,178,952
Trainable params: 129,381
Non-trainable params: 4,049,571
_________________________________________________________________

###########################################################################
cloned_model.load_weights(checkpoint_path)

<tensorflow.python.training.tracking.util.CheckpointLoadStatus at 0x7fc84333af50>
###########################################################################

cloned_model.compile(loss="sparse_categorical_crossentropy",
                     optimizer=tf.keras.optimizers.Adam(),
                     metrics=["accuracy"])
results_cloned_model_with_loaded_weights = cloned_model.evaluate(test_data)

790/790 [==============================] - 49s 61ms/step - loss: 1.7362 - accuracy: 0.5509

[santiagorg2401]

Hey,
I just did the Milestone project and I noticed exactly the same:

I found out that the EfficientNetBX family has issues serializing to save the model on TF2.10 >, I am unable to conclude if this issue is related to the loss of accuracy after loading weights, did you ever find out the reason?

[mrdbourke]

Hi @reshma93s and @santiagorg2401 ,

Thank you for your troubleshooting steps and tips.

After a fair bit of troubleshooting, it looks like the tf.keras.clone_model method is not performing reliably.

As I was able to create a new model (using the same modelling code as the original model was made with) and load the weights to achieve the same performance.

Alongside this, I installed tf-nightly (May 2023 version which is 2.14.0) to help fix with tf.keras.applications.efficientnet issues.

See the full update/solution here: #550