5.0 / 5
We will cover the following tasks in 49 minutes:
We will understand the Rhyme interface and our learning environment. You will get a virtual machine, you will need Jupyter Notebook and TensorFlow for this course and both of these are already installed on your virtual machine. Jupyter Notebooks are very popular with Data Science and Machine Learning Engineers as one can write code in cells and use other cells for documentation. Computer vision is a complicated subject but the primary aim is to help machines gain a high-level understanding derived from images and videos. Like the name implies, computer vision is simply helping computer achieve a human-like visual system.
We will import TensorFlow and its implementation of Keras - a high-level API to build and train deep learning models. We will also import NumPy - which is the fundamental package for scientific computing in python. We will import the Sequential class from Keras as well as many of the layers that we will be using to build our convolutional neural network model.
We are going to use the Fashion MNIST dataset for this course. This dataset actually comes with Keras so it’s really simple to use it if you’re using Keras to build your model, which we are. This dataset has 60k training examples and 10k examples for the test set. The labels are numeric values from 0 to 9 for the 10 classes that this dataset has. The Fashion MNIST dataset has grayscale images of different type of clothing items. We will take a look at the images in a minute. The images are grayscale so there is no channel information and all the images are of 28 by 28 pixels with each pixel having a value between 0 to 255.
The labels in our dataset are just integers from 0 to 9. We need to transform the labels in a way that each column corresponds to each class. So, if a label has value 3, this means that the new transformed labels array will have 0s for all the columns except for column at index 3 which will have a value of 1. This type of transformation is called one hot encoding and fortunately for us, Keras provides a simple way for achieving this with the
We will also normalize the training and test data. This is easy to achieve. We already know the maximum value a pixel can have is 255 so all we have to do is divide our examples by 255. We will also expand the dimensions on our training and test examples. This is because our images are grayscale so in the dataset there is no information provided for channels because for a grayscale image we don’t need that information. But the convolutional layers, the ones we are going to use in our model, do expect the channels to be present even if the value for channels is just 1. So, we will use NumPy’s expand dimension method to add channel information.
Create the Model
We will use the Sequential class we loaded earlier. We can simply pass in a list of layers to the initializer. The first two layers are convolutional 2D layer. Then, we will flatten the convolutional layers so that the 4D tensors are converted to 2D tensors. The flattened output goes to a fully connected layer. We will use the Dense class here. And finally, for the output, we will need a Dense layer with the
softmax activation function and number of classes for number of nodes.
Compiling the Model
We will compile this model with a variant of stochastic gradient descent algorithm which is called
adaptive moment estimation or
adam in short. We will use the loss function called
categorical cross entropy and we will use accuracy as our training metric.
Training the Model
We will use the
fit method to train the model that we created in the last chapter. Let’s pass on the expanded version of the training examples. Let’s also use 20% of the training set for validation. Use a batch size of 128. This is the number of samples to be used for each training step in each epoch. And also, we will use total 2 iterations of the entire training set.
Now that the model is trained, we can start making predictions using this trained instance. All we have to do is use the
predict method on the model. We will make predictions on the first 25 examples in our test set and display the results visually.
About the Host (Amit Yadav)
I am a Software Engineer with many years of experience in writing commercial software. My current areas of interest include computer vision and sequence modelling for automated signal processing using deep learning as well as developing chatbots.