KERAS TUNER : Neural Networks Made Easy

2021-03-12T09:58:00.000-08:00

Hey, I recently came across this library known as keras tuner from Krish Naik's youtube channel he has done a detailed tutorial on its implementation. It can be simply installed using the pip command.

!pip install keras-tuner

So basically what this keras tuner does is optimization of a neural network by tweaking it's hyperparameters for better accuracy. It basically reduces the time that we spend on defining the hyperparameters value like number of filters, kernel size and etc throughout the neural network.

def build_model(hp):

  model=keras.models.Sequential()
  model.add(Conv2D( filters=hp.Int('conv_filter1', min_value=64, max_value=128, step=16),
                   kernel_size=hp.Choice('conv_kernel1', values=[3,5]), 
                   activation='relu',
                   input_shape=(28,28,1) ))
  
  model.add(Conv2D( filters=hp.Int('conv_filter2', min_value=32, max_value=64, step=16),
                   kernel_size=hp.Choice('conv_kernel2', values=[3,5]), 
                   activation='relu'))
  model.add(Flatten() )
  model.add(Dense(units=hp.Int('dense_units',min_value=32, max_value=128, step=16)))
  model.add(Dense(units=10,activation='softmax'))

  model.compile(optimizer=keras.optimizers.Adam(hp.Choice('learning_rate',values=[1e-2,1e-3])), 
                loss='sparse_categorical_crossentropy',
                metrics=['accuracy'])
  
  return model

This is the implementation part of the code (GitHub link). It is just a simple model building except we don't have to work hard or refer to any source for attaining better accuracy the keras tuner library does it for you. In my github repository I have tried implementing this model with certain variation which actually had some impact on the models outcome not just accuracy.

For example you could try implementing the same model removing the add() i.e. bringing everything into sequential() which will cause a reduction in the total number of parameters used improving the efficiency of the model. Tried a few other variations but didn't have a significant effect on the outcome.

Anybody who has experience building neural networks would find this implementation efficient and easy instead of defining everything we define a function and number of layers it takes care of the rest. This library and application could be ran in any IDE, I used google colab found it simple and easy to practice. There are other libraries available for the same purpose but this was the one I found better. Maybe in the future if I come across something more similar or better I'll try to post it (comment if you have any doubts). Thank you.

My First Experience with Flask and Streamlit

2021-03-11T12:45:00.007-08:00

So to begin with I obviously started with flask as soon as I wanted to learn deployment, developed a simple application for image recognition using resnet found it really interesting and easy. But sure didn't look easy or interesting when I began. The reason was I didn't have any idea with respect to html (which later on turned out to be the easy part, lol!). It was really an easy implementation with some simple steps.

We use an application file in which we implement flask and the model to be deployed could be any machine learning or deep learning model already completed. According to the application we are implementing we design the respective templates and at last we put them all in the same directory and run it. A proper tutorial and some solid time to learn that's all you need. The pre requisites are just python and nothing else, you'll learn the basic html needed to design a web page in the process.(I have attached a github link for the code I used to practice)

Coming on to streamlit it overcomes those little difficulties with respect to html, you are not needed to write any code except python and the application is relatively easy. A glance at the syntaxes in the library will give you a basic picture of what can be done using it. It eliminates the html files required to design the template, which in turn are done using simple syntax from streamlit library. Unlike flask we don't need to run the code repeatedly in the terminal to view changes made in the code we just have to reload the respective web page making deployment more efficient.

Both the libraries where great to work on but relatively felt streamlit was a easier deployment as they bring in everything under one python application but as beginner learning flask would give you a bigger picture of what exactly is done in deployment of a model including the design of web page. So its better to start with Flask get an idea and then try out streamlit.

I have just begun working on deployment problems so there is a long way ahead, don't want to jump into conclusion yet on which to prefer both has its pros and cons someone else who just started may say Flask is better. So trying to stay focused on the main idea "learning". Thank you!.

GitHub: https://github.com/Pradeep23-01/Model-Deployment-

Flask API: https://flask.palletsprojects.com/en/1.1.x/api/

Streamlit API: https://streamlit.io/

Case_study

KERAS TUNER : Neural Networks Made Easy

My First Experience with Flask and Streamlit