Hyper-Parameter Tuning to Optimize DL Models (MobileNet v2)

Photo by Eric Nopanen on Unsplash

After having build multiple machine learning and deep learning models, you will sometimes feel the need to improve upon your existing models. In such a scenario, fine-tuning your model may help you build a better model and get more from your data.

But what exactly is Fine-Tuning and what are some of the results you can expect by using it? Read the blog to find out more.

What is meant by Fine-Tuning a model ?

  • The goal of fine-tuning is to adapt specialized features from a generic dataset to work with the new dataset, rather than overwrite the generic learning.
  • Here I have trained (or “fine-tuned”) the weights of the top layers of the pre-trained model i.e. MobileNet-V2.

Code

Find the Jupyter Notebook with the code here.

Parameters to tune

  • Learning rate: Choosing an appropriate learning_rate helps improve the accuracy. Here we have used learning_rate = 0.0001)
  • No. of epochs: Increasing the no. of epochs upto which the model should train, can increase the accuracy of the model. Here we have trained the model again on 10 epochs (20 in total).
  • Unfreezing the top layers of the model : We try and fine-tune a small number of top layers rather than the whole MobileNet model.
Photo by Jr Korpa on Unsplash

Un-freezing the top layers of the model

  • First step is to unfreeze the base_model and set the bottom layers to be un-trainable.
  • We then recompile the model (necessary for these changes to take effect), and resume training for 10 more epochs.
  • Lastly we compile the model using a much lower learning rate (0.0001).
  • After fine-tuning the model nearly reaches 98% accuracy.
# Setting the top layers as trainable
base_mode.trainable = True

# To check how many layers are in the base model
print('Number of layers in the base model: ',len(base_model.layers))

# Fine-tune from this layer onwards
fine_tune_at = 100

# Freeze all the layers before the `fine_tune_at` layer
for layer in base_model.layers[:fine_tune_at]:
layer.trainable = False

# This step improves the accuracy of the model by a few steps
model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), optimizer = tf.keras.optimizers.RMSprop(lr=base_learning_rate/10),
metrics=['accuracy'])

Fine Tuning the parameter weights

  • To further improve performance, we can repurpose the top-level layers of the pre-trained models to the new dataset via fine-tuning.
  • In this case, we tuned our weights such that our model learned high-level features specific to the dataset.
  • This technique is usually recommended when the training dataset is large and very similar to the original dataset that the pre-trained model was trained on.

Results after Fine-Tuning the model

  • Before fine-tuning the model, the model has an accuracy of 93.25% and loss of 0.14 on the Test set.
  • After fine-tuning the model, the model has an accuracy of 97.89% and loss of 0.058 on the Test set.

Accuracy and Loss on Test set (Before Fine Tuning)

Accuracy and Loss on Test set (After Fine Tuning)

Now go on, try to apply Hyper-parameter tuning on your Deep Learning Projects.

--

--

--

300k+ Views on Medium | 4xTop Writer | Technology, Productivity, Books and Life | Linkedin: linkedin.com/in/afroz-chakure-489780168 | InteractiveGeneration.tech

Love podcasts or audiobooks? Learn on the go with our new app.

Recommended from Medium

Supercharge your Donkey with a NVidia Jetson TX2

Multi-Agent Reinforcement Learning: The Gist

Support Vector Machine (SVM)

Cellular Image Classification for Drug Discovery

Text Recognition in Flutter Using Firebase’s ML Kit

In Search Of New Activation Functions

Face Recognition Using Transfer Learning

No one knows where America’s helipads are, except this neural network

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Afroz Chakure

Afroz Chakure

300k+ Views on Medium | 4xTop Writer | Technology, Productivity, Books and Life | Linkedin: linkedin.com/in/afroz-chakure-489780168 | InteractiveGeneration.tech

More from Medium

Switching between TensorFlow and PyTorch with ONNX

Transfer Learning with MobileNet-v2 💡

What is Data preprocessing, why we need that❓

Data Preprocessing

MLOps: A brief introduction