Today’s article will be exploring how to use Convolutional Neural Networks to classify clothing images through TensorFlow to train a Deep Learning model for an apparel business.
A kind of computer learning called deep learning uses multi-layered neural networks to identify patterns in data. This project’s main goal is to demonstrate how to use Deep Learning principles to solve a picture categorization issue. To do this, we’ll use the Python TensorFlow library to train a Convolutional Neural Network (CNN) to categorize a dataset of apparel.
Convolutional Neural Networks have been performing inhumanly well on various challenging visual tasks during the past few years. They power a variety of things, including automatic video classification systems, self-driving automobiles, and image search services.
The Dataset Investigating For Classifying Images With Tensorflow
70,000 greyscale photos in 10 categories can be found in the Fashion MNIST dataset, which is what we’ll be used. Imagine the resolution of the image, which depicts individual apparel, is 28 x 28. Each image has one class with a value between [0, 9].
A few samples from our train dataset can be shown alongside the appropriate class. There is at least one instance of each class visible when displaying the first 24 samples.
Starting The Data Preprocessing To Classify The Images Through TensorFlow
Normalizing
The value of each pixel in the image should be an integer for the effective function of the model and must be normalized. A function that divides each value by 255.0 can be written. This function will produce normalized values in the [0,1] range for our dataset.
Picture Reshaping
The number of samples (60,000), the size of the pixels (28 x 28), and the color channel are the four input dimensions that the convolutional neural network expects.
One-Hot EnCoding
The labels and the accompanying class do not, however, have an ordinal relationship. In this situation, the model can presume a natural ordering of the categories by employing integer encoding, which could lead to subpar performance or surprising findings from the Deep Learning model. We can use one-hot encoding to fix this issue, which generates a new binary variable for each distinct integer value.
Creating A Model
To classify the images through TensorFlow we will utilize the conventional CNN architecture.
Layering The Model
We will add a convolutional and a pooling layer first, followed by another convolutional and pooling layer, just as we did in the image. Then, we’ll add a flattened layer to turn our 2-dimensional array into a 1-dimensional array and several dense layers. To lessen overfitting, some dropout layers can be added. The probability distribution for each class is created in the final layer by adding a dense layer with the number of classes from our problem (10) and a SoftMax activation.
Compiling The Model
The model must now be assembled. Here, we pass the loss function, which calculates the difference between true and predicted values, the optimizer—which modifies the weights to minimize the loss—and the metrics, which compute the model’s performance.
Training The Model
Our model must be trained as the final phase. The input data, the target data, and the number of epochs—which determines how many complete rounds of the training dataset there will be—must be passed at this point.
Loss Evaluation
The history object that the fit method returns contain the outcomes for each epoch. The loss and accuracy for the validation and training datasets can be plotted on a graph. This graphic allows one to see how, across the epochs, the loss decreases and the accuracy increases.
Speculating And Assessing The Outcomes
Assessing The Test Dataset’s Accuracy
Compared to the test and validation datasets, the accuracy on the training dataset is lower. With 91,55% accuracy, this is still an excellent outcome.
Predictions
For each sample in our test database, we may use our model to predict a class.
the true and predicted labels for a few photos from our test dataset are plotted. The text will be displayed in blue when the model makes a correct prediction and in red if the forecast is incorrect. Additionally, the anticipated class’s estimated probability will be shown.
Crosstab
When we examine our crosstab, we can see that the product for which we had the best accuracy was also the one with the lowest accuracy. The crosstab is a fantastic way to see the quantities that our model predicts for each class.
Classification Report
Now, by using the classification report from the scikit-learn library, let’s plot a summary showing the precision, recall, and f1-score for each class. Here, it is clear that there was no class in which there was a significant difference between precision and recall.
Conclusions
You have learned how to use TensorFlow to build a convolutional neural network to categorize photos in this post. With this model, you may achieve a respectable overall accuracy in your test dataset of 91,55%. However, by utilizing some data augmentation approaches, we might attempt to increase the accuracy of this class if the results aren’t that good. Additionally, if you want a model with greater accuracy, you might want to experiment with altering some hyperparameters or utilizing a different network architecture.