easier that train simple deep convolutional network and most importantly the

Easier that train simple deep convolutional network

This preview shows page 1 - 3 out of 4 pages.

easier that train simple deep convolutional network and most importantly the lack of accuracy had been solved; that is the concept of ResNet, there is three typed of ResNet (RN);RN50 , RN101 and RN152. Each one of them provides layers. For my project I used RN50 since I don’t have that amount of classes Figure[1] : evolution of Depth In addition I found a video on the internet that explains and provides more information about the ResNet[2]. As you know without adjustments, deep networks suffer from vanishing gradients its gets smaller and smaller which it causes the learning intractable; and that’s where ResNet comes it provides a feature we can call it “skip connection’ this feature allows the network to learn the function and pass the input through the block without passing through the other layers, from this we can stack layers and build deep networks. DataSet The data that I am going to use in this project is images, it was collected over a period of 3 months, they were obtained from Google Image search based on the 15 most populated countries in the world, after they were collected the dataset was compiled by Moses Olafenwa; each image that contains a professional is contained in a separate folder with the folder name corresponding to the image label (e.g doctor), each image has a size of 224*224 pixel JPEG. The dataset contains 11000 pictures and it had been split into 9000 to the training module and 2000 picture for the testing module,as shown in the figure below: Figure[2]: dataset of IdentProf XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Image of page 1
Train Module In order to train the ImageAI module I used the code in the appendix, using the Custom Prediction Model Training allowed to train the module on my set of images (dataset) after its done using the Google-colab it generates a JSON file that maps the persons in my image dataset. Once in defined the model trainer I can set the type of the network (ResNet) and sets the path to the image dataset, for a high accuracy achievements I used the “enhance-Data” parameter which allows the network to produce a set of copies of the training images, by using this parameter and training the module a set amount of trails(61) I achieved 79% accuracy, without that parameter I achieved 71% after 41 training experiments, every experiment that the network trained (epoch) takes 121 seconds, after every trial using the KERAS library I imported Model- checkpoint so that I can save the trial with the highes accuracy as shown in the figure;
Image of page 2
Image of page 3

You've reached the end of your free preview.

Want to read all 4 pages?

  • Left Quote Icon

    Student Picture

  • Left Quote Icon

    Student Picture

  • Left Quote Icon

    Student Picture

Stuck? We have tutors online 24/7 who can help you get unstuck.
A+ icon
Ask Expert Tutors You can ask You can ask You can ask (will expire )
Answers in as fast as 15 minutes