If you are trying to create a model for classifying images and have a set of examples for training, you should try a convolutional neural network. Knowing that training a convnet could be somehow struggling, I wrote this blog to help you in your first touch with this technology. Here, I will teach you how to train a customized convolutional neural network for image classification.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”]
First of all, you will need to download our convolutional neural network repository<\/a>, which is based on Tensorflow2.3<\/a> with Keras. In this repository, you will find implementations of well-known architectures like AlexNet and ResNet. Moreover, to facilitate the data manipulation, we will store it as tfrecords (a nice format of Tensorflow) that are then read by a Dataset object. This allows the model to accelerate the reading process during training.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”] To make this example easy, we will use the well-known MNIST dataset during this blog. Please download the training<\/a> y testing<\/a> datasets. It consists of 60000 images for training and 10000 images for testing. Figure 1 shows a sample of the images we can find in the MNIST dataset.<\/p>\n If you want to try with a smaller dataset, you can use MNIST-5000<\/a>, which is a sample of MNIST with 5000 images for training.<\/p>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/mnist.png” title_text=”A sample of images from MNIST dataset” align=”center” _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” header_font=”Roboto|900||on|||||” text_orientation=”center” min_height=”31px”] Figure 1. A sample of images from the MNIST dataset.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”] These parameters define one experiment but we can include different sets of parameters for various experiments. To identify each set of parameters we use a section name. An example of a configuration file (mnist_full.config<\/a>) that we will use for the MNIST dataset is shown below:<\/p>\n [MNIST]<\/p>\n An efficient way to store the data is through tfrecords. This allows the data to efficiently feed the underlying neural network. In our example we create the tfrecords file using the following command:\n\u00a0<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” background_color=”rgba(0,0,0,0.29)” min_height=”72px” custom_margin=”||42px|||” custom_padding=”||0px|||” border_color_all=”#000000″ inline_fonts=”Courier Prime,Abril Fatface”] With the outputs produced by the previous step, we are ready to train the model. We define a simple architecture composed of two convolutional layers at the beginning, and two fully connected layers at the end. The last layer is the classification layers with as many neurons as classes the task has. The implementation of this architecture can be found aqu\u00ed<\/a>.<\/p>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/simple_net.png” title_text=”simple_net” align=”center” _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” header_font=”Roboto|900||on|||||” text_orientation=”center” min_height=”31px”] Figure 2.\u00a0 An implementation of a simple convnet for MNIST.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”56px” custom_padding=”||0px|||”] Now we are ready to train our model. You can do it running the following command:<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” background_color=”rgba(0,0,0,0.29)” min_height=”56px” custom_margin=”||42px|||” custom_padding=”||0px|||” border_color_all=”#000000″ inline_fonts=”Courier Prime”] At the beginning of the execution, you will see a report of the involved parameters. Then, you will see the performance of the training for each iteration showing the loss and accuracy of the model. At the end of each epoch, the same metrics are reported for the testing (validation) dataset.<\/p>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/simple_net_output1.png” title_text=”Number of parameters of each netwrok’s module” align=”center” _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” header_font=”Roboto|900||on|||||” text_orientation=”center” min_height=”31px”] Figure 3.\u00a0 Parameters required for the network.<\/p>\n <\/p>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/simple_net_output_2.png” title_text=”simple_net_output_2″ _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” header_font=”Roboto|900||on|||||” text_orientation=”center” min_height=”29px” custom_margin=”||15px|||”] Figure 4.\u00a0 Output produced during training.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”68px” custom_padding=”||0px|||”] The training process will yield checkpoints after each epoch. In this particular case, you will see 10 files named from 001.h5 to 010.h5<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”] You can also test your model. To this end, you need to set the CKPFILE parameter in the configuration file with the checkpoint you want to try (e.g. 010.h5), and then run the command below, which is very similar to the previous one changing the parameter mode<\/em> for testing.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” background_color=”rgba(0,0,0,0.29)” min_height=”56px” custom_margin=”||42px|||” custom_padding=”||0px|||” border_color_all=”#000000″ inline_fonts=”Courier Prime”] Figure 5.\u00a0 Output produced by testing.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”] Figure 6.\u00a0 Example of a prediction of the MNIST model.<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||” min_height=”13px”] The code in train_simple.py<\/a> is composed of the following sections:<\/p>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”] In this section, the tfrecord file is fed into a dataset object, which allows us to organize the data into batches, apply data augmentation using the function map<\/em>, and shuffle the data after each epoch.<\/p>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/code_section_1.png” title_text=”code_section_1″ _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”] Here, we configure a ModelCheckpoint object to save checkpoints after a certain number of iterations during training (e.g., after each epoch). Check the ModelCheckpoint<\/a> object for more details.<\/p>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/code_section_2_1.png” title_text=”code_section_2_1″ _builder_version=”4.9.4″ _module_preset=”default” min_height=”129px” custom_margin=”-12px|||||”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”] Here, we create an instance of the network you will use. In this example, we create an instance of SimpleModel<\/a>. The last part of this section builds the model defining the input size.<\/p>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/code_section_3.png” title_text=”code_section_3″ _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”]The example dataset<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”100px” custom_padding=”||0px|||”]
Prepare the dataset for training\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”100px” custom_padding=”||0px|||”]We will require two text files, one listing the images that will be used for training and the other listing the images for testing (or validation). These files should be named as train.txt and test.txt, respectively. These should be in a two-column format, where the first column describes the absolute filename for each image, and the second is for the corresponding class (0-indexed). The separator between columns is the tab character. Finally, to organize the files, we will call DATA_DIR the folder where the described text files are located.[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”]
Prepare a configuration file\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”100px” custom_padding=”||0px|||”]Training a convnet requires defining a set of hyperparameters. A good manner of tuning these parameters is through a configuration file, where we define the parameters required for the training and testing stages. In this toolkit, we encourage users to create a configuration file where they can set parameters like:[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”516px” custom_padding=”||0px|||”]
\n
preprocessing operations to be applied to the input images.<\/li>\n\n
Create tfrecords<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”71px” custom_padding=”||0px|||” custom_margin=”||29px|||”]
$python datasets\/create_tfrecords.py -type all -config configs\/mnist_full.config -name MNIST<\/span>
<\/span><\/h4>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|900|||||||” header_font=”Roboto|900||on|||||”]Train your own model\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”100px” custom_padding=”||0px|||”]
$python train_simple.py -mode train\u00a0 -config configs\/mnist_full.config -name MNIST<\/span><\/h4>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”56px” custom_padding=”||0px|||”]
Test your own model\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”100px” custom_padding=”||0px|||”]
$python train_simple.py -mode test\u00a0 -config configs\/mnist_full.config -name MNIST<\/span><\/h4>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/simple_net_test-1.png” title_text=”simple_net_test” _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” header_font=”Roboto|900||on|||||” text_orientation=”center” min_height=”29px” custom_margin=”||15px|||”]
Predict with your trained model\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” background_color=”rgba(0,0,0,0.29)” min_height=”56px” custom_margin=”||42px|||” custom_padding=”||0px|||” border_color_all=”#000000″ inline_fonts=”Courier Prime”]
$python train_simple.py -mode predict\u00a0 -config configs\/mnist_full.config -name MNIST<\/span><\/h4>[\/et_pb_text][et_pb_image src=”https:\/\/impresee.com\/wp-content\/uploads\/simple_net_predict.png” title_text=”simple_net_predict” align=”center” _builder_version=”4.9.4″ _module_preset=”default”][\/et_pb_image][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” header_font=”Roboto|900||on|||||” text_orientation=”center” min_height=”29px” custom_margin=”||15px|||” custom_padding=”||31px|||”]
Understanding the code\n<\/h1>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”36px” custom_margin=”||19px|||” custom_padding=”4px||0px|||” hover_enabled=”0″ sticky_enabled=”0″]
Data loading\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”66px” custom_padding=”||0px|||”]
Define callbacks for saving checkpoints\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”66px” custom_margin=”||35px|||” custom_padding=”||0px|||”]
Create an instance of your network\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”65px” custom_margin=”||26px|||” custom_padding=”||0px|||” hover_enabled=”0″ sticky_enabled=”0″]
Loading pre-trained weights\n<\/h2>[\/et_pb_text][et_pb_text _builder_version=”4.9.4″ text_font=”Roboto|300|||||||” text_text_color=”#000000″ text_font_size=”18px” text_line_height=”1.5em” min_height=”43px” custom_padding=”||0px|||” hover_enabled=”0″ sticky_enabled=”0″]