Hello everyone, in my last blog post, I wanted to discuss a simple application about my favorite topic, CNN. I chose the mnist data set, which is one of the easiest data sets for this topic.
The MNIST database (Modified National Institute of Standards and Technology database) is a large database of handwritten digits that is commonly used for training various image processing systems. The MNIST database contains 60,000 training images and 10,000 testing images. Half of the training set and half of the test set were taken from NIST’s training dataset, while the other half of the training set and the other half of the test set were taken from NIST’s testing dataset. The images contained within have a width of 28 pixels and a height of 28 pixels.
Figure : Sample images from MNIST test dataset
Data set is imported from tensorflow library. Session function has used to running codes. Global_variables_initializer has activated for codes to work. Data should be given piece by piece to train the model so batch_size has taken as 128. A function called “training step” has been created for the realization of the training. The for loop has defined as the loop that will perform the training in the function. MNIST pictures have taken with this code x_batch, y_batch = mnist. train. next_batch(batch_size) so we have feed pictures to our model in the form of batch. Feed_dict_train has defined to assign images and tags in the data set to our place holders. The code has written in one line to simultaneously optimize the model and see the variability of the loss value. The if loop has been used to observe the situation in our training. It is coded for training accuracy and training loss printing every 100 iterations. The test_accuracy function has been defined to see how our model predicts data that it has not encountered before.
2 convolutional layers have used to implement the MNIST data set. As a result of trials, when the number of convolutional layers, training step and filter sizes have increased, it has seen that the accuracy increased.First convolutional layer has 16 filters and they all have 5×5 size filters. Second convolutional layer has 32 filters and they all have 5×5 size filters. Layers have combined by making necessary arrangements with max pooling function. ReLU and SoftMax functions have used as activation function. Adam has been used as an optimization algorithm. A very small value of 0.0005 was taken as the learning rate. Batch size is set to 128 for make the training better. Training accuracy and training loss have printed on the output every 100 iterations to check the accuracy of the model. Test accuracy 0.9922 has obtained because of 10000 iterations when the codes have executed.
Figure : Estimation mistakes made by the model
In the figure above, some examples that our model incorrectly predicted are given. Our model can sometimes make wrong predictions, which may be because the text is faint or unclear. In the first example, we see that our model estimates the number 4 as 2.
Figure : Graph of Loss function
The loss graph gives us a visualized version of the loss values we observed during the training. As shown in the figure, we have a decreasing loss graph over time. Our goal is bringing the loss value closer to zero. Through the loss graph, we can see the appropriateness of the learning rate. When we look at the figure, we can say that our learning rate value is good because there is no slowdown in the decrease in the graph.
In this blog, I made an application on Python using CNN with the Mnist data set. Thank you to everyone who has followed my blogs closely until today, goodbye until we see you again …