In the TV show Silicon Valley, there's a famous scene where the developer Jian-Yang demonstrates an app that can identify any kind of food in an image. Of course this being Silicon Valley, there's a catch: the app can only identify hotdogs and classifies everything else as 'not hotdog':
You can watch the full video here: https://www.youtube.com/watch?v=pqTntG1RXSY
In this assignment you are going to build this same app, which must be able to identify hotdogs in any image.
The easiest way to do this is to build a convolutional neural network and train it on a dataset of hotdog and not-hotdog images. The Kaggle Hotdog dataset has exactly what we need.
You can download the dataset here. Unzip the archive and make sure the hotdog and nothotdog folders are created in the project folder you're going to create below.
Here's what the dataset looks like:
These are 499 pictures of hotdogs. We also have a second set with 499 images of fast food that isn't a hotdog.
You will need to train a neural network on these image sets and get the hotdog detection accuracy as high as possible.
Let’s get started. You need to build a new application from scratch by opening a terminal and creating a new NET Core console project:
https://gist.github.com/6722a3ab73182195818f66050fda13e2
Also make sure to copy the dataset folders hotdog and nothotdog into this folder because the code you're going to type next will expect it here.
Now install the following packages
https://gist.github.com/7ec4bb932033e52e4da3ea32245f64cd
The CNTK.GPU library is Microsoft's Cognitive Toolkit that can train and run deep neural networks. And Xplot.Plotly is an awesome plotting library based on Plotly. The library is designed for F# so we also need to pull in the Fsharp.Core library.
The CNTK.GPU package will train and run deep neural networks using your GPU. You'll need an NVidia GPU and Cuda graphics drivers for this to work.
If you don't have an NVidia GPU or suitable drivers, the library will fall back and use the CPU instead. This will work but training neural networks will take significantly longer.
CNTK is a low-level tensor library for building, training, and running deep neural networks. The code to build deep neural network can get a bit verbose, so I've developed a little wrapper called CNTKUtil that will help you write code faster.
Please download the CNTKUtil files in a new CNTKUtil folder at the same level as your project folder.
Then make sure you're in the console project folder and crearte a project reference like this:
https://gist.github.com/87ed6549221443572dfd40aede0b6b41
Now you are ready to start writing code. Edit the Program.cs file with Visual Studio Code and add the following code:
https://gist.github.com/206fbb0d77c18d055ebbb10c607370cf
The first thing you'll need to do is add a method to build the mapping files. These are text files that map each image in the dataset to a corresponding label. We will encode a hotdog with a '1' and a not-hotdog with a '0' value. So the mapping file should look like this:
You can see that each image has been paired with a label indicating if the image contains a hotdog or not.
You're now going to add a method that will automatically create the mapping files.
Add the following code:
https://gist.github.com/ad364db86f1639dcf12180f7dc655a93
This method uses Directory.GetFiles to collect all image files in the dataset and then uses nested loops to write the file names to train_map.txt and test_map.txt files. These files contain all image file names for training and testing the neural network.
You can experiment with the sizes of the training and testing partitions by changing the values of the trainingSetSize and testingSetSize constants. Just make sure they both add up to 499.
Now it's time to start writing the main program method:
https://gist.github.com/d0988639f748e16eab4d6e5bb9391518
This code calls CreateMappingFiles to set up the training and testing mapping files. Then it calls GetImageReader twice to set up two image readers, one for the training images and one for the testing images.
Note that the images in the training set are randomized. We do this to prevent the neural network from learning patterns associated with the specific sorting of the images in the dataset.
Note the imageWidth, imageHeight, and numChannels constants. We are going to rescale every image to 150x150 pixels and feed all 3 color channels into the neural network. This means we will be training directly on color images without transforming them to grayscale first.
Now we need to tell CNTK what shape the input data has that we'll train the neural network on, and what shape the output data of the neural network will have:
https://gist.github.com/15e631a525ba1fdb158057787e3d21ad
Note the first Var method which tells CNTK that our neural network will use a 3-dimensional tensor of 150 by 150 pixels with 3 color channels each. This matches the shape of the images returned by the trainingReader and testingReader.
The second Var method tells CNTK that we want our neural network to output a 1-dimensional tensor of 2 float values. The first float will indicate the probability that the image does not contain a hotdog, and the second float indicates the probability that the image does contain a hotdog.
Our next step is to design the neural network.
We will use a deep convolutional neural network with a mix of convolution and pooling layers, a dropout layer to stabilize training, and two dense layers for classification. We'll use the ReLU activation function for the convolution layers and the classifier, and Softmax activation for the final dense layer.
The network looks like this:
The network has the following layers:
- A 3x3 convolution layer with 32 filters and ReLU
- A 2x2 max pooling layer with stride 2
- A 3x3 convolution layer with 64 filters and ReLU
- A 2x2 max pooling layer with stride 2
- A 3x3 convolution layer with 128 filters and ReLU
- A 2x2 max pooling layer with stride 2
- A 3x3 convolution layer with 128 filters and ReLU
- A 2x2 max pooling layer with stride 2
- A dropout layer with a 50% dropout rate
- A 512-node hidden layer with ReLU
- A 2-node output layer with softmax
Remember: the sofmax function creates a mutually exclusive list of output classes where only a single class can be the correct answer. If we had used sigmoid, the neural network might predict that an images represents both a hotdog and not a hotdog. We don't want that here.
Here's the code to build the neural network:
https://gist.github.com/9680fdb4b3d5736f0fb99924f759fa18
Each Convolution2D call adds a convolution layer, Pooling adds a pooling layer, Dropout adds a dropout layer, and Dense adds a dense feedforward layer to the network. We're using ReLU activation almost everywhere, with Softmax only in the final dense layer.
Then we use the ToSummary method to output a description of the architecture of the neural network to the console.
Now we need to decide which loss function to use to train the neural network, and how we are going to track the prediction error of the network during each training epoch.
For this assignment we'll use CrossEntropyWithSoftmax as the loss function because it's the standard metric for measuring multiclass classification loss with softmax.
We'll track the error with the ClassificationError metric. This is the number of times (expressed as a percentage) that the model predictions are wrong. An error of 0 means the predictions are correct all the time, and an error of 1 means the predictions are wrong all the time.
https://gist.github.com/979f3a83d4eea7948ea4f4bf98b98ad2
Next we need to decide which algorithm to use to train the neural network. There are many possible algorithms derived from Gradient Descent that we can use here.
For this assignment we're going to use the AdamLearner. You can learn more about the Adam algorithm here: https://machinelearningmastery.com/adam...
https://gist.github.com/8ddbc2aa95af645e3c324e6f194ef877
These configuration values are a good starting point for many machine learning scenarios, but you can tweak them if you like to try and improve the quality of your predictions.
We're almost ready to train. Our final step is to set up a trainer and an evaluator for calculating the loss and the error during each training epoch:
https://gist.github.com/339d111460f3a0b038edee73e69c20f5
The GetTrainer method sets up a trainer which will track the loss and the error for the training partition. And GetEvaluator will set up an evaluator that tracks the error in the test partition.
Now we're finally ready to start training the neural network!
Add the following code:
https://gist.github.com/276d26703b3d0c2353d3083341f124f3
We're training the network for 100 epochs using a batch size of 16. During training we'll track the loss and errors in the loss, trainingError and testingError arrays.
Once training is done, we show the final testing error on the console. This is the percentage of mistakes the network makes when predicting hotdogs.
Note that the error and the accuracy are related: accuracy = 1 - error. So we also report the final accuracy of the neural network.
Here's the code to train the neural network. Put this inside the for loop:
https://gist.github.com/a59a6c8b309f6c2d2265afd11a7e1fec
The while loop keeps training unit we've processed every image in the training set once. Inside the loop we call GetBatch to get a training batch of images and then access the StreamInfo method to get the feature batch (the images) and the label batch (the zeroes and ones indicating hotdogs). Then we call TrainBatch to train the neural network on these two batches of training data.
The TrainBatch method returns the loss and error, but only for training on the 16-image batch. So we simply add up all these values and divide them by the number of batches in the dataset. That gives us the average loss and error for the predictions on the training partition during the current epoch, and we report this to the console.
So now we know the training loss and error for one single training epoch. The next step is to test the network by making predictions about the data in the testing partition and calculate the testing error.
Put this code inside the epoch loop and right below the training code:
https://gist.github.com/3f6c7d3ee064c308a38200bea4cd8875
Again we use a while loop to process each image in the partition, calling GetBatch to get the images and StreamInfo to access the feature and label batches. But note that we're now providing the testingReader to get the images in the test set.
We call TestBatch to test the neural network on the 16-image test batch. The method returns the error for the batch, and we again add up the errors for each batch and divide by the number of batches.
That gives us the average error in the neural network predictions on the test partition for this epoch.
After training completes, the training and testing errors for each epoch will be available in the trainingError and testingError arrays. Let's use XPlot to create a nice plot of the two error curves so we can check for overfitting:
https://gist.github.com/fed483e7885695842141e30761367eef
This code creates a Plot with two Scatter graphs. The first one plots the trainingError values and the second one plots the testingError values. Also note the WithOptions call that forces the y-axis to start at zero.
Finally we use File.WriteAllText to write the plot to disk as a HTML file.
We're now ready to build the app, so this is a good moment to save your work ;)
Go to the CNTKUtil folder and type the following:
https://gist.github.com/cf6d669ec71bb1b9706dc05a318e9f90
This will build the CNKTUtil project. Note how we're specifying the x64 platform because the CNTK library requires a 64-bit build.
Now go to the HotdogNotHotdog folder and type:
https://gist.github.com/7b5252b35bc16b1b23e5c5aed2c10a38
This will build your app. Note how we're again specifying the x64 platform.
Now run the app:
https://gist.github.com/a9c4927dcd4dfeeca2964305daed5d62
The app will create the neural network, load the dataset, train the network on the images, and create a plot of the training and testing errors for each epoch.
The plot is written to disk in a new file called chart.html. Open the file now and take a look at the training and testing curves.
What are your final classification errors on training and testing? What is the final testing accuracy? And what do the curves look like? Is the neural network overfitting?
Do you think this model is good at predicting hotdogs?
Now let's improve the accuracy of the model. You are going to enable Data Augmentation to artificially increase the size of the dataset and give the neural network more to train on.
Change the code that sets up the image readers as follows:
https://gist.github.com/fa54a84fba5cf985e37e9023795d7488
The change is that the augmentData argument of the trainingReader is now set to true. The reader will randomly rotate, translate, and shear the training images to artificially increase the size of the training set. This will help the neural network recognize hotdogs in more orientations and prevent overfitting.
Now compile and run the app again:
https://gist.github.com/8cb637abb51f6b0b12b890201e99eb16
What is your final accuracy on testing now? Has the model improved?
And what changes do you see in the shape of the training and testing curves? Can you explain how data augmentation has resulted in these changes?
Share your results and conclusions in our group.