One place for hosting & domains

      Python

      How To Apply Computer Vision to Build an Emotion-Based Dog Filter in Python 3


      The author selected Girls Who Code to receive a donation as part of the Write for DOnations program.

      Introduction

      Computer vision is a subfield of computer science that aims to extract a higher-order understanding from images and videos. This field includes tasks such as object detection, image restoration (matrix completion), and optical flow. Computer vision powers technologies such as self-driving car prototypes, employee-less grocery stores, fun Snapchat filters, and your mobile device’s face authenticator.

      In this tutorial, you will explore computer vision as you use pre-trained models to build a Snapchat-esque dog filter. For those unfamiliar with Snapchat, this filter will detect your face and then superimpose a dog mask on it. You will then train a face-emotion classifier so that the filter can pick dog masks based on emotion, such as a corgi for happy or a pug for sad. Along the way, you will also explore related concepts in both ordinary least squares and computer vision, which will expose you to the fundamentals of machine learning.

      A working dog filter

      As you work through the tutorial, you’ll use OpenCV, a computer-vision library, numpy for linear algebra utilities, and matplotlib for plotting. You’ll also apply the following concepts as you build a computer-vision application:

      • Ordinary least squares as a regression and classification technique.
      • The basics of stochastic gradient neural networks.

      While not necessary to complete this tutorial, you’ll find it easier to understand some of the more detailed explanations if you’re familiar with these mathematical concepts:

      • Fundamental linear algebra concepts: scalars, vectors, and matrices.
      • Fundamental calculus: how to take a derivative.

      You can find the complete code for this tutorial at https://github.com/do-community/emotion-based-dog-filter.

      Let’s get started.

      Prerequisites

      To complete this tutorial, you will need the following:

      Step 1 — Creating The Project and Installing Dependencies

      Let’s create a workspace for this project and install the dependencies we’ll need. We’ll call our workspace DogFilter:

      Navigate to the DogFilter directory:

      Then create a new Python virtual environment for the project:

      • python3 -m venv dogfilter

      Activate your environment.

      • source dogfilter/bin/activate

      The prompt changes, indicating the environment is active. Now install PyTorch, a deep-learning framework for Python that we'll use in this tutorial. The installation process depends on which operating system you're using.

      On macOS, install Pytorch with the following command:

      • python -m pip install torch==0.4.1 torchvision==0.2.1

      On Linux, use the following commands:

      • pip install http://download.pytorch.org/whl/cpu/torch-0.4.1-cp35-cp35m-linux_x86_64.whl
      • pip install torchvision

      And for Windows, install Pytorch with these commands:

      • pip install http://download.pytorch.org/whl/cpu/torch-0.4.1-cp35-cp35m-win_amd64.whl
      • pip install torchvision

      Now install prepackaged binaries for OpenCV and numpy, which are computer vision and linear algebra libraries, respectively. The former offers utilities such as image rotations, and the latter offers linear algebra utilities such as a matrix inversion.

      • python -m pip install opencv-python==3.4.3.18 numpy==1.14.5

      Finally, create a directory for our assets, which will hold the images we'll use in this tutorial:

      With the dependencies installed, let's build the first version of our filter: a face detector.

      Step 2 — Building a Face Detector

      Our first objective is to detect all faces in an image. We'll create a script that accepts a single image and outputs an annotated image with the faces outlined with boxes.

      Fortunately, instead of writing our own face detection logic, we can use pre-trained models. We'll set up a model and then load pre-trained parameters. OpenCV makes this easy by providing both.

      OpenCV provides the model parameters in its source code. but we need the absolute path to our locally-installed OpenCV to use these parameters. Since that absolute path may vary, we'll download our own copy instead and place it in the assets folder:

      • wget -O assets/haarcascade_frontalface_default.xml https://github.com/opencv/opencv/raw/master/data/haarcascades/haarcascade_frontalface_default.xml

      The -O option specifies the destination as assets/haarcascade_frontalface_default.xml. The second argument is the source URL.

      We'll detect all faces in the following image from Pexels (CC0, link to original image).

      Picture of children

      First, download the image. The following command saves the downloaded image as children.png in the assets folder:

      • wget -O assets/children.png https://xpresservers.com/wp-content/uploads/2019/04/How-To-Apply-Computer-Vision-to-Build-an-Emotion-Based-Dog-Filter-in-Python-3.png

      To check that the detection algorithm works, we will run it on an individual image and save the resulting annotated image to disk. Create an outputs folder for these annotated results.

      Now create a Python script for the face detector. Create the file step_1_face_detect using nano or your favorite text editor:

      • nano step_2_face_detect.py

      Add the following code to the file. This code imports OpenCV, which contains the image utilities and face classifier. The rest of the code is typical Python program boilerplate.

      step_2_face_detect.py

      """Test for face detection"""
      
      import cv2
      
      
      def main():
          pass
      
      if __name__ == '__main__':
          main()
      

      Now replace pass in the main function with this code which initializes a face classifier using the OpenCV parameters you downloaded to your assets folder:

      step_2_face_detect.py

      def main():
          # initialize front face classifier
          cascade = cv2.CascadeClassifier("assets/haarcascade_frontalface_default.xml")
      

      Next, add this line to load the image children.png.

      step_2_face_detect.py

          frame = cv2.imread('assets/children.png')
      

      Then add this code to convert the image to black and white, as the classifier was trained on black-and-white images. To accomplish this, we convert to grayscale and then discretize the histogram:

      step_2_face_detect.py

          # Convert to black-and-white
          gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
          blackwhite = cv2.equalizeHist(gray)
      

      Then use OpenCV's detectMultiScale function to detect all faces in the image.

      step_2_face_detect.py

          rects = cascade.detectMultiScale(
              blackwhite, scaleFactor=1.3, minNeighbors=4, minSize=(30, 30),
              flags=cv2.CASCADE_SCALE_IMAGE)
      
      • scaleFactor specifies how much the image is reduced along each dimension.
      • minNeighbors denotes how many neighboring rectangles a candidate rectangle needs to be retained.
      • minSize is the minimum allowable detected object size. Objects smaller than this are discarded.

      The return type is a list of tuples, where each tuple has four numbers denoting the minimum x, minimum y, width, and height of the rectangle in that order.

      Iterate over all detected objects and draw them on the image in green using cv2.rectangle:

      step_2_face_detect.py

          for x, y, w, h in rects:
              cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
      
      • The second and third arguments are opposing corners of the rectangle.
      • The fourth argument is the color to use. (0, 255, 0) corresponds to green for our RGB color space.
      • The last argument denotes the width of our line.

      Finally, write the image with bounding boxes into a new file at outputs/children_detected.png:

      step_2_face_detect.py

          cv2.imwrite('outputs/children_detected.png', frame)
      

      Your completed script should look like this:

      step_2_face_detect.py

      """Tests face detection for a static image."""  
      
      import cv2  
      
      
      def main():  
      
          # initialize front face classifier  
          cascade = cv2.CascadeClassifier(  
              "assets/haarcascade_frontalface_default.xml")  
      
          frame = cv2.imread('assets/children.png')  
      
          # Convert to black-and-white  
          gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)  
          blackwhite = cv2.equalizeHist(gray)  
      
          rects = cascade.detectMultiScale(  
              blackwhite, scaleFactor=1.3, minNeighbors=4, minSize=(30, 30),  
          flags=cv2.CASCADE_SCALE_IMAGE)  
      
          for x, y, w, h in rects:  
              cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)  
      
          cv2.imwrite('outputs/children_detected.png', frame)  
      
      if __name__ == '__main__':  
          main()
      

      Save the file and exit your editor. Then run the script:

      • python step_2_face_detect.py

      Open outputs/children_detected.png. You'll see the following image that shows the faces outlined with boxes:

      Picture of children with bounding boxes

      At this point, you have a working face detector. It accepts an image as input and draws bounding boxes around all faces in the image, outputting the annotated image. Now let's apply this same detection to a live camera feed.

      Step 3 — Linking the Camera Feed

      The next objective is to link the computer's camera to the face detector. Instead of detecting faces in a static image, you'll detect all faces from your computer's camera. You will collect camera input, detect and annotate all faces, and then display the annotated image back to the user. You'll continue from the script in Step 2, so start by duplicating that script:

      • cp step_2_face_detect.py step_3_camera_face_detect.py

      Then open the new script in your editor:

      • nano step_3_camera_face_detect.py

      You will update the main function by using some elements from this test script from the official OpenCV documentation. Start by initializing a VideoCapture object that is set to capture live feed from your computer's camera. Place this at the start of the main function, before the other code in the function:

      step_3_camera_face_detect.py

      def main():
          cap = cv2.VideoCapture(0)
          ...
      

      Starting from the line defining frame, indent all of your existing code, placing all of the code in a while loop.

      step_3_camera_face_detect.py

          while True:
              frame = cv2.imread('assets/children.png')
              ...
              for x, y, w, h in rects:  
                  cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)  
      
              cv2.imwrite('outputs/children_detected.png', frame)
      

      Replace the line defining frame at the start of the while loop. Instead of reading from an image on disk, you're now reading from the camera:

      step_3_camera_face_detect.py

          while True:
              # frame = cv2.imread('assets/children.png') # DELETE ME
              # Capture frame-by-frame
              ret, frame = cap.read()
      

      Replace the line cv2.imwrite(...) at the end of the while loop. Instead of writing an image to disk, you'll display the annotated image back to the user's screen:

      step_3_camera_face_detect.py

            cv2.imwrite('outputs/children_detected.png', frame)  # DELETE ME
            # Display the resulting frame
            cv2.imshow('frame', frame)
      

      Also, add some code to watch for keyboard input so you can stop the program. Check if the user hits the q character and, if so, quit the application. Right after cv2.imshow(...) add the following:

      step_3_camera_face_detect.py

      ...
              cv2.imshow('frame', frame)
              if cv2.waitKey(1) & 0xFF == ord('q'):
                  break
      ...
      

      The line cv2.waitkey(1) halts the program for 1 millisecond so that the captured image can be displayed back to the user.

      Finally, release the capture and close all windows. Place this outside of the while loop to end the main function.

      step_3_camera_face_detect.py

      ...
      
          while True:
          ...
      
      
          cap.release()
          cv2.destroyAllWindows()
      

      Your script should look like the following:

      step_3_camera_face_detect.py

      """Test for face detection on video camera.
      
      Move your face around and a green box will identify your face.
      With the test frame in focus, hit `q` to exit.
      Note that typing `q` into your terminal will do nothing.
      """
      
      import cv2
      
      
      def main():
          cap = cv2.VideoCapture(0)
      
          # initialize front face classifier
          cascade = cv2.CascadeClassifier(
              "assets/haarcascade_frontalface_default.xml")
      
          while True:
              # Capture frame-by-frame
              ret, frame = cap.read()
      
              # Convert to black-and-white
              gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
              blackwhite = cv2.equalizeHist(gray)
      
              # Detect faces
              rects = cascade.detectMultiScale(
                  blackwhite, scaleFactor=1.3, minNeighbors=4, minSize=(30, 30),
                  flags=cv2.CASCADE_SCALE_IMAGE)
      
              # Add all bounding boxes to the image
              for x, y, w, h in rects:
                  cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
      
              # Display the resulting frame
              cv2.imshow('frame', frame)
              if cv2.waitKey(1) & 0xFF == ord('q'):
                  break
      
          # When everything done, release the capture
          cap.release()
          cv2.destroyAllWindows()
      
      
      if __name__ == '__main__':
          main()
      

      Save the file and exit your editor.

      Now run the test script.

      • python step_3_camera_face_detect.py

      This activates your camera and opens a window displaying your camera's feed. Your face will be boxed by a green square in real time:

      Working face detector

      Note: If you find that you have to hold very still for things to work, the lighting in the room may not be adequate. Try moving to a brightly lit room where you and your background have high constrast. Also, avoid bright lights near your head. For example, if you have your back to the sun, this process might not work very well.

      Our next objective is to take the detected faces and superimpose dog masks on each one.

      Step 4 — Building the Dog Filter

      Before we build the filter itself, let's explore how images are represented numerically. This will give you the background needed to modify images and ultimately apply a dog filter.

      Let's look at an example. We can construct a black-and-white image using numbers, where 0 corresponds to black and 1 corresponds to white.

      Focus on the dividing line between 1s and 0s. What shape do you see?

      0 0 0 0 0 0 0 0 0
      0 0 0 0 1 0 0 0 0
      0 0 0 1 1 1 0 0 0
      0 0 1 1 1 1 1 0 0
      0 0 0 1 1 1 0 0 0
      0 0 0 0 1 0 0 0 0
      0 0 0 0 0 0 0 0 0
      

      The image is a diamond. If save this matrix of values as an image. This gives us the following picture:

      Diamond as picture

      We can use any value between 0 and 1, such as 0.1, 0.26, or 0.74391. Numbers closer to 0 are darker and numbers closer to 1 are lighter. This allows us to represent white, black, and any shade of gray. This is great news for us because we can now construct any grayscale image using 0, 1, and any value in between. Consider the following, for example. Can you tell what it is? Again, each number corresponds to the color of a pixel.

      1  1  1  1  1  1  1  1  1  1  1  1
      1  1  1  1  0  0  0  0  1  1  1  1
      1  1  0  0 .4 .4 .4 .4  0  0  1  1
      1  0 .4 .4 .5 .4 .4 .4 .4 .4  0  1
      1  0 .4 .5 .5 .5 .4 .4 .4 .4  0  1
      0 .4 .4 .4 .5 .4 .4 .4 .4 .4 .4  0
      0 .4 .4 .4 .4  0  0 .4 .4 .4 .4  0
      0  0 .4 .4  0  1 .7  0 .4 .4  0  0
      0  1  0  0  0 .7 .7  0  0  0  1  0
      1  0  1  1  1  0  0 .7 .7 .4  0  1
      1  0 .7  1  1  1 .7 .7 .7 .7  0  1
      1  1  0  0 .7 .7 .7 .7  0  0  1  1
      1  1  1  1  0  0  0  0  1  1  1  1
      1  1  1  1  1  1  1  1  1  1  1  1
      

      Re-rendered as an image, you can now tell that this is, in fact, a Poké Ball:

      Pokeball as picture

      You've now seen how black-and-white and grayscale images are represented numerically. To introduce color, we need a way to encode more information. An image has its height and width expressed as h x w.

      Image

      In the current grayscale representation, each pixel is one value between 0 and 1. We can equivalently say our image has dimensions h x w x 1. In other words, every (x, y) position in our image has just one value.

      Grayscale image

      For a color representation, we represent the color of each pixel using three values between 0 and 1. One number corresponds to the "degree of red," one to the "degree of green," and the last to the "degree of blue." We call this the RGB color space. This means that for every (x, y) position in our image, we have three values (r, g, b). As a result, our image is now h x w x 3:

      Color image

      Here, each number ranges from 0 to 255 instead of 0 to 1, but the idea is the same. Different combinations of numbers correspond to different colors, such as dark purple (102, 0, 204) or bright orange (255, 153, 51). The takeaways are as follows:

      1. Each image will be represented as a box of numbers that has three dimensions: height, width, and color channels. Manipulating this box of numbers directly is equivalent to manipulating the image.
      2. We can also flatten this box to become just a list of numbers. In this way, our image becomes a vector. Later on, we will refer to images as vectors.

      Now that you understand how images are represented numerically, you are well-equipped to begin applying dog masks to faces. To apply a dog mask, you will replace values in the child image with non-white dog mask pixels. To start, you will work with a single image. Download this crop of a face from the image you used in Step 2.

      • wget -O assets/child.png https://xpresservers.com/wp-content/uploads/2019/04/1554419826_451_How-To-Apply-Computer-Vision-to-Build-an-Emotion-Based-Dog-Filter-in-Python-3.png

      Cropped face

      Additionally, download the following dog mask. The dog masks used in this tutorial are my own drawings, now released to the public domain under a CC0 License.

      Dog mask

      Download this with wget:

      • wget -O assets/dog.png https://xpresservers.com/wp-content/uploads/2019/04/1554419826_685_How-To-Apply-Computer-Vision-to-Build-an-Emotion-Based-Dog-Filter-in-Python-3.png

      Create a new file called step_4_dog_mask_simple.py which will hold the code for the script that applies the dog mask to faces:

      • nano step_4_dog_mask_simple.py

      Add the following boilerplate for the Python script and import the OpenCV and numpy libraries:

      step_4_dog_mask_simple.py

      """Test for adding dog mask"""
      
      import cv2
      import numpy as np
      
      
      def main():
          pass
      
      if __name__ == '__main__':
          main()
      

      Replace pass in the main function with these two lines which load the original image and the dog mask into memory.

      step_4_dog_mask_simple.py

      ...
      def main():
          face = cv2.imread('assets/child.png')
          mask = cv2.imread('assets/dog.png')
      

      Next, fit the dog mask to the child. The logic is more complicated than what we've done previously, so we will create a new function called apply_mask to modularize our code. Directly after the two lines that load the images, add this line which invokes the apply_mask function:

      step_4_dog_mask_simple.py

      ...
          face_with_mask = apply_mask(face, mask)
      

      Create a new function called apply_mask and place it above the main function:

      step_4_dog_mask_simple.py

      ...
      def apply_mask(face: np.array, mask: np.array) -> np.array:
          """Add the mask to the provided face, and return the face with mask."""
          pass
      
      def main():
      ...
      

      At this point, your file should look like this:

      step_4_dog_mask_simple.py

      """Test for adding dog mask"""
      
      import cv2
      import numpy as np
      
      
      def apply_mask(face: np.array, mask: np.array) -> np.array:
          """Add the mask to the provided face, and return the face with mask."""
          pass
      
      
      def main():
          face = cv2.imread('assets/child.png')
          mask = cv2.imread('assets/dog.png')
          face_with_mask = apply_mask(face, mask)
      
      if __name__ == '__main__':
          main()
      

      Let's build out the apply_mask function. Our goal is to apply the mask to the child's face. However, we need to maintain the aspect ratio for our dog mask. To do so, we need to explicitly compute our dog mask's final dimensions. Inside the apply_mask function, replace pass with these two lines which extract the height and width of both images:

      step_4_dog_mask_simple.py

      ...
          mask_h, mask_w, _ = mask.shape
          face_h, face_w, _ = face.shape
      

      Next, determine which dimension needs to be "shrunk more." To be precise, we need the tighter of the two constraints. Add this line to the apply_mask function:

      step_4_dog_mask_simple.py

      ...
      
          # Resize the mask to fit on face
          factor = min(face_h / mask_h, face_w / mask_w)
      

      Then compute the new shape by adding this code to the function:

      step_4_dog_mask_simple.py

      ...
          new_mask_w = int(factor * mask_w)
          new_mask_h = int(factor * mask_h)
          new_mask_shape = (new_mask_w, new_mask_h)
      

      Here we cast the numbers to integers, as the resize function needs integral dimensions.

      Now add this code to resize the dog mask to the new shape:

      step_4_dog_mask_simple.py

      ...
      
          # Add mask to face - ensure mask is centered
          resized_mask = cv2.resize(mask, new_mask_shape)
      

      Finally, write the image to disk so you can double-check that your resized dog mask is correct after you run the script:

      step_4_dog_mask_simple.py

          cv2.imwrite('outputs/resized_dog.png', resized_mask)
      

      The completed script should look like this:

      step_4_dog_mask_simple.py

      """Test for adding dog mask"""
      import cv2
      import numpy as np
      
      def apply_mask(face: np.array, mask: np.array) -> np.array:
          """Add the mask to the provided face, and return the face with mask."""
          mask_h, mask_w, _ = mask.shape
          face_h, face_w, _ = face.shape
      
          # Resize the mask to fit on face
          factor = min(face_h / mask_h, face_w / mask_w)
          new_mask_w = int(factor * mask_w)
          new_mask_h = int(factor * mask_h)
          new_mask_shape = (new_mask_w, new_mask_h)
      
          # Add mask to face - ensure mask is centered
          resized_mask = cv2.resize(mask, new_mask_shape)
          cv2.imwrite('outputs/resized_dog.png', resized_mask)
      
      
      def main():
          face = cv2.imread('assets/child.png')
          mask = cv2.imread('assets/dog.png')
          face_with_mask = apply_mask(face, mask)
      
      if __name__ == '__main__':
          main()
      
      

      Save the file and exit your editor. Run the new script:

      • python step_4_dog_mask_simple.py

      Open the image at outputs/resized_dog.png to double-check the mask was resized correctly. It will match the dog mask shown earlier in this section.

      Now add the dog mask to the child. Open the step_4_dog_mask_simple.py file again and return to the apply_mask function:

      • nano step_4_dog_mask_simple.py

      First, remove the line of code that writes the resized mask from the apply_mask function since you no longer need it:

          cv2.imwrite('outputs/resized_dog.png', resized_mask)  # delete this line
          ...
      

      In its place, apply your knowledge of image representation from the start of this section to modify the image. Start by making a copy of the child image. Add this line to the apply_mask function:

      step_4_dog_mask_simple.py

      ...
          face_with_mask = face.copy()
      

      Next, find all positions where the dog mask is not white or near white. To do this, check if the pixel value is less than 250 across all color channels, as we'd expect a near-white pixel to be near [255, 255, 255]. Add this code:

      step_4_dog_mask_simple.py

      ...
          non_white_pixels = (resized_mask < 250).all(axis=2)
      

      At this point, the dog image is, at most, as large as the child image. We want to center the dog image on the face, so compute the offset needed to center the dog image by adding this code to apply_mask:

      step_4_dog_mask_simple.py

      ...
          off_h = int((face_h - new_mask_h) / 2)  
          off_w = int((face_w - new_mask_w) / 2)
      

      Copy all non-white pixels from the dog image into the child image. Since the child image may be larger than the dog image, we need to take a subset of the child image:

      step_4_dog_mask_simple.py

          face_with_mask[off_h: off_h+new_mask_h, off_w: off_w+new_mask_w][non_white_pixels] = 
                  resized_mask[non_white_pixels]
      

      Then return the result:

      step_4_dog_mask_simple.py

          return face_with_mask
      

      In the main function, add this code to write the result of the apply_mask function to an output image so you can manually double-check the result:

      step_4_dog_mask_simple.py

      ...
          face_with_mask = apply_mask(face, mask)
          cv2.imwrite('outputs/child_with_dog_mask.png', face_with_mask)
      

      Your completed script will look like the following:

      step_4_dog_mask_simple.py

      """Test for adding dog mask"""
      
      import cv2
      import numpy as np
      
      
      def apply_mask(face: np.array, mask: np.array) -> np.array:
          """Add the mask to the provided face, and return the face with mask."""
          mask_h, mask_w, _ = mask.shape
          face_h, face_w, _ = face.shape
      
          # Resize the mask to fit on face
          factor = min(face_h / mask_h, face_w / mask_w)
          new_mask_w = int(factor * mask_w)
          new_mask_h = int(factor * mask_h)
          new_mask_shape = (new_mask_w, new_mask_h)
          resized_mask = cv2.resize(mask, new_mask_shape)
      
          # Add mask to face - ensure mask is centered
          face_with_mask = face.copy()
          non_white_pixels = (resized_mask < 250).all(axis=2)
          off_h = int((face_h - new_mask_h) / 2)  
          off_w = int((face_w - new_mask_w) / 2)
          face_with_mask[off_h: off_h+new_mask_h, off_w: off_w+new_mask_w][non_white_pixels] = 
               resized_mask[non_white_pixels]
      
          return face_with_mask
      
      def main():
          face = cv2.imread('assets/child.png')
          mask = cv2.imread('assets/dog.png')
          face_with_mask = apply_mask(face, mask)
          cv2.imwrite('outputs/child_with_dog_mask.png', face_with_mask)
      
      if __name__ == '__main__':
          main()
      

      Save the script and run it:

      • python step_4_dog_mask_simple.py

      You'll have the following picture of a child with a dog mask in outputs/child_with_dog_mask.png:

      Picture of child with dog mask on

      You now have a utility that applies dog masks to faces. Now let's use what you've built to add the dog mask in real time.

      We'll pick up from where we left off in Step 3. Copy step_3_camera_face_detect.py to step_4_dog_mask.py.

      • cp step_3_camera_face_detect.py step_4_dog_mask.py

      Open your new script.

      First, import the NumPy library at the top of the script:

      step_4_dog_mask.py

      import numpy as np
      ...
      

      Then add the apply_mask function from your previous work into this new file above the main function:

      step_4_dog_mask.py

      def apply_mask(face: np.array, mask: np.array) -> np.array:
          """Add the mask to the provided face, and return the face with mask."""
          mask_h, mask_w, _ = mask.shape
          face_h, face_w, _ = face.shape
      
          # Resize the mask to fit on face
          factor = min(face_h / mask_h, face_w / mask_w)
          new_mask_w = int(factor * mask_w)
          new_mask_h = int(factor * mask_h)
          new_mask_shape = (new_mask_w, new_mask_h)
          resized_mask = cv2.resize(mask, new_mask_shape)
      
          # Add mask to face - ensure mask is centered
          face_with_mask = face.copy()
          non_white_pixels = (resized_mask < 250).all(axis=2)
          off_h = int((face_h - new_mask_h) / 2)  
          off_w = int((face_w - new_mask_w) / 2)
          face_with_mask[off_h: off_h+new_mask_h, off_w: off_w+new_mask_w][non_white_pixels] = 
               resized_mask[non_white_pixels]
      
          return face_with_mask
      ...
      

      Second, locate this line in the main function:

      step_4_dog_mask.py

          cap = cv2.VideoCapture(0)
      

      Add this code after that line to load the dog mask:

      step_4_dog_mask.py

          cap = cv2.VideoCapture(0)
      
          # load mask
          mask = cv2.imread('assets/dog.png')
          ...
      

      Next, in the while loop, locate this line:

      step_4_dog_mask.py

              ret, frame = cap.read()
      

      Add this line after it to extract the image's height and width:

      step_4_dog_mask.py

              ret, frame = cap.read()
              frame_h, frame_w, _ = frame.shape
              ...
      

      Next, delete the line in main that draws bounding boxes. You'll find this line in the for loop that iterates over detected faces:

      step_4_dog_mask.py

              for x, y, w, h in rects:
              ...
                  cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2) # DELETE ME
              ...
      

      In its place, add this code which crops the frame. For aesthetic purposes, we crop an area slightly larger than the face.

      step_4_dog_mask.py

              for x, y, w, h in rects:
                  # crop a frame slightly larger than the face
                  y0, y1 = int(y - 0.25*h), int(y + 0.75*h)
                  x0, x1 = x, x + w
      

      Introduce a check in case the detected face is too close to the edge.

      step_4_dog_mask.py

                  # give up if the cropped frame would be out-of-bounds
                  if x0 < 0 or y0 < 0 or x1 > frame_w or y1 > frame_h:
                      continue
      

      Finally, insert the face with a mask into the image.

      step_4_dog_mask.py

                  # apply mask
                  frame[y0: y1, x0: x1] = apply_mask(frame[y0: y1, x0: x1], mask)
      

      Verify that your script looks like this:

      step_4_dog_mask.py

      """Real-time dog filter
      
      Move your face around and a dog filter will be applied to your face if it is not out-of-bounds. With the test frame in focus, hit `q` to exit. Note that typing `q` into your terminal will do nothing.
      """
      
      import numpy as np
      import cv2
      
      
      def apply_mask(face: np.array, mask: np.array) -> np.array:
          """Add the mask to the provided face, and return the face with mask."""
          mask_h, mask_w, _ = mask.shape
          face_h, face_w, _ = face.shape
      
          # Resize the mask to fit on face
          factor = min(face_h / mask_h, face_w / mask_w)
          new_mask_w = int(factor * mask_w)
          new_mask_h = int(factor * mask_h)
          new_mask_shape = (new_mask_w, new_mask_h)
          resized_mask = cv2.resize(mask, new_mask_shape)
      
          # Add mask to face - ensure mask is centered
          face_with_mask = face.copy()
          non_white_pixels = (resized_mask < 250).all(axis=2)
          off_h = int((face_h - new_mask_h) / 2)
          off_w = int((face_w - new_mask_w) / 2)
          face_with_mask[off_h: off_h+new_mask_h, off_w: off_w+new_mask_w][non_white_pixels] = 
               resized_mask[non_white_pixels]
      
          return face_with_mask
      
      def main():
          cap = cv2.VideoCapture(0)
      
          # load mask
          mask = cv2.imread('assets/dog.png')
      
          # initialize front face classifier
          cascade = cv2.CascadeClassifier("assets/haarcascade_frontalface_default.xml")
      
          while(True):
              # Capture frame-by-frame
              ret, frame = cap.read()
              frame_h, frame_w, _ = frame.shape
      
              # Convert to black-and-white
              gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
              blackwhite = cv2.equalizeHist(gray)
      
              # Detect faces
              rects = cascade.detectMultiScale(
                  blackwhite, scaleFactor=1.3, minNeighbors=4, minSize=(30, 30),
                  flags=cv2.CASCADE_SCALE_IMAGE)
      
              # Add mask to faces
              for x, y, w, h in rects:
                  # crop a frame slightly larger than the face
                  y0, y1 = int(y - 0.25*h), int(y + 0.75*h)
                  x0, x1 = x, x + w
      
                  # give up if the cropped frame would be out-of-bounds
                  if x0 < 0 or y0 < 0 or x1 > frame_w or y1 > frame_h:
                      continue
      
                  # apply mask
                  frame[y0: y1, x0: x1] = apply_mask(frame[y0: y1, x0: x1], mask)
      
              # Display the resulting frame
              cv2.imshow('frame', frame)
              if cv2.waitKey(1) & 0xFF == ord('q'):
                  break
      
          # When everything done, release the capture
          cap.release()
          cv2.destroyAllWindows()
      
      
      if __name__ == '__main__':
          main()
      

      Save the file and exit your editor. Then run the script.

      • python step_4_dog_mask.py

      You now have a real-time dog filter running. The script will also work with multiple faces in the picture, so you can get your friends together for some automatic dog-ification.

      GIF for working dog filter

      This concludes our first primary objective in this tutorial, which is to create a Snapchat-esque dog filter. Now let's use facial expression to determine the dog mask applied to a face.

      Step 5 — Build a Basic Face Emotion Classifier using Least Squares

      In this section you'll create an emotion classifier to apply different masks based on displayed emotions. If you smile, the filter will apply a corgi mask. If you frown, it will apply a pug mask. Along the way, you'll explore the least-squares framework, which is fundamental to understanding and discussing machine learning concepts.

      To understand how to process our data and produce predictions, we'll first briefly explore machine learning models.

      We need to ask two questions for each model that we consider. For now, these two questions will be sufficient to differentiate between models:

      1. Input: What information is the model given?
      2. Output: What is the model trying to predict?

      At a high-level, the goal is to develop a model for emotion classification. The model is:

      1. Input: given images of faces.
      2. Output: predicts the corresponding emotion.
      model: face -> emotion
      

      The approach we'll use is least squares; we take a set of points, and we find a line of best fit. The line of best fit, shown in the following image, is our model.

      Least Squares

      Consider the input and output for our line:

      1. Input: given x coordinates.
      2. Output: predicts the corresponding $y$ coordinate.
      least squares line: x -> y
      

      Our input x must represent faces and our output y must represent emotion, in order for us to use least squares for emotion classification:

      • x -> face: Instead of using one number for x, we will use a vector of values for x. Thus, x can represent images of faces. The article Ordinary Least Squares explains why you can use a vector of values for x.
      • y -> emotion: Each emotion will correspond to a number. For example, "angry" is 0, "sad" is 1, and "happy" is 2. In this way, y can represent emotions. However, our line is not constrained to output the y values 0, 1, and 2. It has an infinite number of possible y values–it could be 1.2, 3.5, or 10003.42. How do we translate those y values to integers corresponding to classes? See the article One-Hot Encoding for more detail and explanation.

      Armed with this background knowledge, you will build a simple least-squares classifier using vectorized images and one-hot encoded labels. You'll accomplish this in three steps:

      1. Preprocess the data: As explained at the start of this section, our samples are vectors where each vector encodes an image of a face. Our labels are integers corresponding to an emotion, and we'll apply one-hot encoding to these labels.
      2. Specify and train the model: Use the closed-form least squares solution, w^*.
      3. Run a prediction using the model: Take the argmax of Xw^* to obtain predicted emotions.

      Let's get started.

      First, set up a directory to contain the data:

      Then download the data, curated by Pierre-Luc Carrier and Aaron Courville, from a 2013 Face Emotion Classification competition on Kaggle.

      • wget -O data/fer2013.tar https://bitbucket.org/alvinwan/adversarial-examples-in-computer-vision-building-then-fooling/raw/babfe4651f89a398c4b3fdbdd6d7a697c5104cff/fer2013.tar

      Navigate to the data directory and unpack the data.

      • cd data
      • tar -xzf fer2013.tar

      Now we'll create a script to run the least-squares model. Navigate to the root of your project:

      Create a new file for the script:

      Add Python boilerplate and import the packages you will need:

      step_5_ls_simple.py

      """Train emotion classifier using least squares."""
      
      import numpy as np
      
      def main():
          pass
      
      if __name__ == '__main__':
          main()
      

      Next, load the data into memory. Replace pass in your main function with the following code:

      step_5_ls_simple.py

      
          # load data
          with np.load('data/fer2013_train.npz') as data:
              X_train, Y_train = data['X'], data['Y']
      
          with np.load('data/fer2013_test.npz') as data:
              X_test, Y_test = data['X'], data['Y']
      

      Now one-hot encode the labels. To do this, construct the identity matrix with numpy and then index into this matrix using our list of labels:

      step_5_ls_simple.py

          # one-hot labels
          I = np.eye(6)
          Y_oh_train, Y_oh_test = I[Y_train], I[Y_test]
      

      Here, we use the fact that the i-th row in the identity matrix is all zero, except for the i-th entry. Thus, the i-th row is the one-hot encoding for the label of class i. Additionally, we use numpy's advanced indexing, where [a, b, c, d][[1, 3]] = [b, d].

      Computing (X^TX)^{-1} would take too long on commodity hardware, as X^TX is a 2304x2304 matrix with over four million values, so we'll reduce this time by selecting only the first 100 features. Add this code:

      step_5_ls_simple.py

      ...
          # select first 100 dimensions
          A_train, A_test = X_train[:, :100], X_test[:, :100]
      

      Next, add this code to evaluate the closed-form least-squares solution:

      step_5_ls_simple.py

      ...
          # train model
          w = np.linalg.inv(A_train.T.dot(A_train)).dot(A_train.T.dot(Y_oh_train))
      

      Then define an evaluation function for training and validation sets. Place this before your main function:

      step_5_ls_simple.py

      def evaluate(A, Y, w):
          Yhat = np.argmax(A.dot(w), axis=1)
          return np.sum(Yhat == Y) / Y.shape[0]
      

      To estimate labels, we take the inner product with each sample and get the indices of the maximum values using np.argmax. Then we compute the average number of correct classifications. This final number is your accuracy.

      Finally, add this code to the end of the main function to compute the training and validation accuracy using the evaluate function you just wrote:

      step_5_ls_simple.py

          # evaluate model
          ols_train_accuracy = evaluate(A_train, Y_train, w)
          print('(ols) Train Accuracy:', ols_train_accuracy)
          ols_test_accuracy = evaluate(A_test, Y_test, w)
          print('(ols) Test Accuracy:', ols_test_accuracy)
      

      Double-check that your script matches the following:

      step_5_ls_simple.py

      """Train emotion classifier using least squares."""
      
      import numpy as np
      
      
      def evaluate(A, Y, w):
          Yhat = np.argmax(A.dot(w), axis=1)
          return np.sum(Yhat == Y) / Y.shape[0]
      
      def main():
      
          # load data
          with np.load('data/fer2013_train.npz') as data:
              X_train, Y_train = data['X'], data['Y']
      
          with np.load('data/fer2013_test.npz') as data:
              X_test, Y_test = data['X'], data['Y']
      
          # one-hot labels
          I = np.eye(6)
          Y_oh_train, Y_oh_test = I[Y_train], I[Y_test]
      
          # select first 100 dimensions
          A_train, A_test = X_train[:, :100], X_test[:, :100]
      
          # train model
          w = np.linalg.inv(A_train.T.dot(A_train)).dot(A_train.T.dot(Y_oh_train))
      
          # evaluate model
          ols_train_accuracy = evaluate(A_train, Y_train, w)
          print('(ols) Train Accuracy:', ols_train_accuracy)
          ols_test_accuracy = evaluate(A_test, Y_test, w)
          print('(ols) Test Accuracy:', ols_test_accuracy)
      
      
      if __name__ == '__main__':
          main()
      

      Save your file, exit your editor, and run the Python script.

      • python step_5_ls_simple.py

      You'll see the following output:

      Output

      (ols) Train Accuracy: 0.4748918316507146 (ols) Test Accuracy: 0.45280545359202934

      Our model gives 47.5% train accuracy. We repeat this on the validation set to obtain 45.3% accuracy. For a three-way classification problem, 45.3% is reasonably above guessing, which is 33%​. This is our starting classifier for emotion detection, and in the next step, you'll build off of this least-squares model to improve accuracy. The higher the accuracy, the more reliably your emotion-based dog filter can find the appropriate dog filter for each detected emotion.

      Step 6 — Improving Accuracy by Featurizing the Inputs

      We can use a more expressive model to boost accuracy. To accomplish this, we featurize our inputs.

      The original image tells us that position (0, 0) is red, (1, 0) is brown, and so on. A featurized image may tell us that there is a dog to the top-left of the image, a person in the middle, etc. Featurization is powerful, but its precise definition is beyond the scope of this tutorial.

      We'll use an approximation for the radial basis function (RBF) kernel, using a random Gaussian matrix. We won't go into detail in this tutorial. Instead, we'll treat this as a black box that computes higher-order features for us.

      We'll continue where we left off in the previous step. Copy the previous script so you have a good starting point:

      • cp step_5_ls_simple.py step_6_ls_simple.py

      Open the new file in your editor:

      We'll start by creating the featurizing random matrix. Again, we'll use only 100 features in our new feature space.

      Locate the following line, defining A_train and A_test:

      step_6_ls_simple.py

          # select first 100 dimensions
          A_train, A_test = X_train[:, :100], X_test[:, :100]
      

      Directly above this definition for A_train and A_test, add a random feature matrix:

      step_6_ls_simple.py

          d = 100
          W = np.random.normal(size=(X_train.shape[1], d))
          # select first 100 dimensions
          A_train, A_test = X_train[:, :100], X_test[:, :100]  ...
      

      Then replace the definitions for A_train and A_test. We redefine our matrices, called design matrices, using this random featurization.

      step_6_ls_simple.py

          A_train, A_test = X_train.dot(W), X_test.dot(W)
      

      Save your file and run the script.

      • python step_6_ls_simple.py

      You'll see the following output:

      Output

      (ols) Train Accuracy: 0.584174642717 (ols) Test Accuracy: 0.584425799685

      This featurization now offers 58.4% train accuracy and 58.4% validation accuracy, a 13.1% improvement in validation results. We trimmed the X matrix to be 100 x 100, but the choice of 100 was arbirtary. We could also trim the X matrix to be 1000 x 1000 or 50 x 50. Say the dimension of x is d x d. We can test more values of d by re-trimming X to be d x d and recomputing a new model.

      Trying more values of d, we find an additional 4.3% improvement in test accuracy to 61.7%. In the following figure, we consider the performance of our new classifier as we vary d. Intuitively, as d increases, the accuracy should also increase, as we use more and more of our original data. Rather than paint a rosy picture, however, the graph exhibits a negative trend:

      Performance of featurized ordinary least squares

      As we keep more of our data, the gap between the training and validation accuracies increases as well. This is clear evidence of overfitting, where our model is learning representations that are no longer generalizable to all data. To combat overfitting, we'll regularize our model by penalizing complex models.

      We amend our ordinary least-squares objective function with a regularization term, giving us a new objective. Our new objective function is called ridge regression and it looks like this:

      min_w |Aw- y|^2 + lambda |w|^2
      

      In this equation, lambda is a tunable hyperparameter. Plug lambda = 0 into the equation and ridge regression becomes least-squares. Plug lambda = infinity into the equation, and you'll find the best w must now be zero, as any non-zero w incurs infinite loss. As it turns out, this objective yields a closed-form solution as well:

      w^* = (A^TA + lambda I)^{-1}A^Ty
      

      Still using the featurized samples, retrain and reevaluate the model once more.

      Open step_6_ls_simple.py again in your editor:

      This time, increase the dimensionality of the new feature space to d=1000​. Change the value of d from 100 to 1000 as shown in the following code block:

      step_6_ls_simple.py

      ...
          d = 1000
          W = np.random.normal(size=(X_train.shape[1], d))
      ...
      

      Then apply ridge regression using a regularization of lambda = 10^{10}. Replace the line defining w with the following two lines:

      step_6_ls_simple.py

      ...
          # train model
          I = np.eye(A_train.shape[1])
          w = np.linalg.inv(A_train.T.dot(A_train) + 1e10 * I).dot(A_train.T.dot(Y_oh_train))
      

      Then locate this block:

      step_6_ls_simple.py

      ...
        ols_train_accuracy = evaluate(A_train, Y_train, w)
        print('(ols) Train Accuracy:', ols_train_accuracy)
        ols_test_accuracy = evaluate(A_test, Y_test, w)
        print('(ols) Test Accuracy:', ols_test_accuracy)
      

      Replace it with the following:

      step_6_ls_simple.py

      ...
      
        print('(ridge) Train Accuracy:', evaluate(A_train, Y_train, w))
        print('(ridge) Test Accuracy:', evaluate(A_test, Y_test, w))
      

      The completed script should look like this:

      step_6_ls_simple.py

      """Train emotion classifier using least squares."""
      
      import numpy as np
      
      def evaluate(A, Y, w):
          Yhat = np.argmax(A.dot(w), axis=1)
          return np.sum(Yhat == Y) / Y.shape[0]
      
      def main():
          # load data
          with np.load('data/fer2013_train.npz') as data:
              X_train, Y_train = data['X'], data['Y']
      
          with np.load('data/fer2013_test.npz') as data:
              X_test, Y_test = data['X'], data['Y']
      
          # one-hot labels
          I = np.eye(6)
          Y_oh_train, Y_oh_test = I[Y_train], I[Y_test]
          d = 1000
          W = np.random.normal(size=(X_train.shape[1], d))
          # select first 100 dimensions
          A_train, A_test = X_train.dot(W), X_test.dot(W)
      
          # train model
          I = np.eye(A_train.shape[1])
          w = np.linalg.inv(A_train.T.dot(A_train) + 1e10 * I).dot(A_train.T.dot(Y_oh_train))
      
          # evaluate model
          print('(ridge) Train Accuracy:', evaluate(A_train, Y_train, w))
          print('(ridge) Test Accuracy:', evaluate(A_test, Y_test, w))
      
      if __name__ == '__main__':
          main()
      

      Save the file, exit your editor, and run the script:

      • python step_6_ls_simple.py

      You'll see the following output:

      Output

      (ridge) Train Accuracy: 0.651173462698 (ridge) Test Accuracy: 0.622181436812

      There's an additional improvement of 0.4% in validation accuracy to 62.2%, as train accuracy drops to 65.1%. Once again reevaluating across a number of different d, we see a smaller gap between training and validation accuracies for ridge regression. In other words, ridge regression was subject to less overfitting.

      Performance of featurized ols and ridge regression

      Baseline performance for least squares, with these extra enhancements, performs reasonably well. The training and inference times, all together, take no more than 20 seconds for even the best results. In the next section, you'll explore even more complex models.

      Step 7 — Building the Face-Emotion Classifier Using a Convolutional Neural Network in PyTorch

      In this section, you'll build a second emotion classifier using neural networks instead of least squares. Again, our goal is to produce a model that accepts faces as input and outputs an emotion. Eventually, this classifier will then determine which dog mask to apply.

      For a brief neural network visualization and introduction, see the article Understanding Neural Networks. Here, we will use a deep-learning library called PyTorch. There are a number of deep-learning libraries in widespread use, and each has various pros and cons. PyTorch is a particularly good place to start. To impliment this neural network classifier, we again take three steps, as we did with the least-squares classifier:

      1. Preprocess the data: Apply one-hot encoding and then apply PyTorch abstractions.
      2. Specify and train the model: Set up a neural network using PyTorch layers. Define optimization hyperparameters and run stochastic gradient descent.
      3. Run a prediction using the model: Evaluate the neural network.

      Create a new file, named step_7_fer_simple.py

      • nano step_7_fer_simple.py

      Import the necessary utilities and create a Python class that will hold your data. For data processing here, you will create the train and test datasets. To do these, implement PyTorch's Dataset interface, which lets you load and use PyTorch's built-in data pipeline for the face-emotion recognition dataset:

      step_7_fer_simple.py

      from torch.utils.data import Dataset
      from torch.autograd import Variable
      import torch.nn as nn
      import torch.nn.functional as F
      import torch.optim as optim
      import numpy as np
      import torch
      import cv2
      import argparse
      
      
      class Fer2013Dataset(Dataset):
          """Face Emotion Recognition dataset.
      
          Utility for loading FER into PyTorch. Dataset curated by Pierre-Luc Carrier
          and Aaron Courville in 2013.
      
          Each sample is 1 x 1 x 48 x 48, and each label is a scalar.
          """
          pass
      

      Delete the pass placeholder in the Fer2013Dataset class. In its place, add a function that will initialize our data holder:

      step_7_fer_simple.py

          def __init__(self, path: str):
              """
              Args:
                  path: Path to `.np` file containing sample nxd and label nx1
              """
              with np.load(path) as data:
                  self._samples = data['X']
                  self._labels = data['Y']
              self._samples = self._samples.reshape((-1, 1, 48, 48))
      
              self.X = Variable(torch.from_numpy(self._samples)).float()
              self.Y = Variable(torch.from_numpy(self._labels)).float()
      ...
      

      This function starts by loading the samples and labels. Then it wraps the data in PyTorch data structures.

      Directly after the __init__ function, add a __len__ function, as this is needed to implement the Dataset interface PyTorch expects:

      step_7_fer_simple.py

      ...
          def __len__(self):
              return len(self._labels)
      

      Finally, add a __getitem__ method, which returns a dictionary containing the sample and the label:

      step_7_fer_simple.py

          def __getitem__(self, idx):
              return {'image': self._samples[idx], 'label': self._labels[idx]}
      

      Double-check that your file looks like the following:

      step_7_fer_simple.py

      from torch.utils.data import Dataset
      from torch.autograd import Variable
      import torch.nn as nn
      import torch.nn.functional as F
      import torch.optim as optim
      import numpy as np
      import torch
      import cv2
      import argparse
      
      
      class Fer2013Dataset(Dataset):
          """Face Emotion Recognition dataset.
          Utility for loading FER into PyTorch. Dataset curated by Pierre-Luc Carrier
          and Aaron Courville in 2013.
          Each sample is 1 x 1 x 48 x 48, and each label is a scalar.
          """
      
          def __init__(self, path: str):
              """
              Args:
                  path: Path to `.np` file containing sample nxd and label nx1
              """
              with np.load(path) as data:
                  self._samples = data['X']
                  self._labels = data['Y']
              self._samples = self._samples.reshape((-1, 1, 48, 48))
      
              self.X = Variable(torch.from_numpy(self._samples)).float()
              self.Y = Variable(torch.from_numpy(self._labels)).float()
      
          def __len__(self):
              return len(self._labels)
      
          def __getitem__(self, idx):
              return {'image': self._samples[idx], 'label': self._labels[idx]}
      

      Next, load the Fer2013Dataset dataset. Add the following code to the end of your file after the Fer2013Dataset class:

      step_7_fer_simple.py

      trainset = Fer2013Dataset('data/fer2013_train.npz')
      trainloader = torch.utils.data.DataLoader(trainset, batch_size=32, shuffle=True)
      
      testset = Fer2013Dataset('data/fer2013_test.npz')
      testloader = torch.utils.data.DataLoader(testset, batch_size=32, shuffle=False)
      

      This code initializes the dataset using the Fer2013Dataset class you created. Then for the train and validation sets, it wraps the dataset in a DataLoader. This translates the dataset into an iterable to use later.

      As a sanity check, verify that the dataset utilities are functioning. Create a sample dataset loader using DataLoader and print the first element of that loader. Add the following to the end of your file:

      step_7_fer_simple.py

      if __name__ == '__main__':
          loader = torch.utils.data.DataLoader(trainset, batch_size=2, shuffle=False)
          print(next(iter(loader)))
      

      Verify that your completed script looks like this:

      step_7_fer_simple.py

      from torch.utils.data import Dataset
      from torch.autograd import Variable
      import torch.nn as nn
      import torch.nn.functional as F
      import torch.optim as optim
      import numpy as np
      import torch
      import cv2
      import argparse
      
      
      class Fer2013Dataset(Dataset):
          """Face Emotion Recognition dataset.
          Utility for loading FER into PyTorch. Dataset curated by Pierre-Luc Carrier
          and Aaron Courville in 2013.
          Each sample is 1 x 1 x 48 x 48, and each label is a scalar.
          """
      
          def __init__(self, path: str):
              """
              Args:
                  path: Path to `.np` file containing sample nxd and label nx1
              """
              with np.load(path) as data:
                  self._samples = data['X']
                  self._labels = data['Y']
              self._samples = self._samples.reshape((-1, 1, 48, 48))
      
              self.X = Variable(torch.from_numpy(self._samples)).float()
              self.Y = Variable(torch.from_numpy(self._labels)).float()
      
          def __len__(self):
              return len(self._labels)
      
          def __getitem__(self, idx):
              return {'image': self._samples[idx], 'label': self._labels[idx]}
      
      trainset = Fer2013Dataset('data/fer2013_train.npz')
      trainloader = torch.utils.data.DataLoader(trainset, batch_size=32, shuffle=True)
      
      testset = Fer2013Dataset('data/fer2013_test.npz')
      testloader = torch.utils.data.DataLoader(testset, batch_size=32, shuffle=False)
      
      if __name__ == '__main__':
          loader = torch.utils.data.DataLoader(trainset, batch_size=2, shuffle=False)
          print(next(iter(loader)))
      

      Exit your editor and run the script.

      • python step_7_fer_simple.py

      This outputs the following pair of tensors. Our data pipeline outputs two samples and two labels. This indicates that our data pipeline is up and ready to go:

      Output

      {'image': (0 ,0 ,.,.) = 24 32 36 ... 173 172 173 25 34 29 ... 173 172 173 26 29 25 ... 172 172 174 ... ⋱ ... 159 185 157 ... 157 156 153 136 157 187 ... 152 152 150 145 130 161 ... 142 143 142 ⋮ (1 ,0 ,.,.) = 20 17 19 ... 187 176 162 22 17 17 ... 195 180 171 17 17 18 ... 203 193 175 ... ⋱ ... 1 1 1 ... 106 115 119 2 2 1 ... 103 111 119 2 2 2 ... 99 107 118 [torch.LongTensor of size 2x1x48x48] , 'label': 1 1 [torch.LongTensor of size 2] }

      Now that you've verified that the data pipeline works, return to step_7_fer_simple.py to add the neural network and optimizer. Open step_7_fer_simple.py.

      • nano step_7_fer_simple.py

      First, delete the last three lines you added in the previous iteration:

      step_7_fer_simple.py

      # Delete all three lines
      if __name__ == '__main__':
          loader = torch.utils.data.DataLoader(trainset, batch_size=2, shuffle=False)
          print(next(iter(loader)))
      

      In their place, define a PyTorch neural network that includes three convolutional layers, followed by three fully connected layers. Add this to the end of your existing script:

      step_7_fer_simple.py

      class Net(nn.Module):
          def __init__(self):
              super(Net, self).__init__()
              self.conv1 = nn.Conv2d(1, 6, 5)
              self.pool = nn.MaxPool2d(2, 2)
              self.conv2 = nn.Conv2d(6, 6, 3)
              self.conv3 = nn.Conv2d(6, 16, 3)
              self.fc1 = nn.Linear(16 * 4 * 4, 120)
              self.fc2 = nn.Linear(120, 48)
              self.fc3 = nn.Linear(48, 3)
      
          def forward(self, x):
              x = self.pool(F.relu(self.conv1(x)))
              x = self.pool(F.relu(self.conv2(x)))
              x = self.pool(F.relu(self.conv3(x)))
              x = x.view(-1, 16 * 4 * 4)
              x = F.relu(self.fc1(x))
              x = F.relu(self.fc2(x))
              x = self.fc3(x)
              return x
      

      Now initialize the neural network, define a loss function, and define optimization hyperparameters by adding the following code to the end of the script:

      step_7_fer_simple.py

      net = Net().float()
      criterion = nn.CrossEntropyLoss()
      optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
      

      We'll train for two epochs. For now, we define an epoch to be an iteration of training where every training sample has been used exactly once.

      First, extract image and label from the dataset loader and then wrap each in a PyTorch Variable. Second, run the forward pass and then backpropagate through the loss and neural network. Add the following code to the end of your script to do that:

      step_7_fer_simple.py

      for epoch in range(2):  # loop over the dataset multiple times
      
          running_loss = 0.0
          for i, data in enumerate(trainloader, 0):
              inputs = Variable(data['image'].float())
              labels = Variable(data['label'].long())
              optimizer.zero_grad()
      
              # forward + backward + optimize
              outputs = net(inputs)
              loss = criterion(outputs, labels)
              loss.backward()
              optimizer.step()
      
              # print statistics
              running_loss += loss.data[0]
              if i % 100 == 0:
                  print('[%d, %5d] loss: %.3f' % (epoch, i, running_loss / (i + 1)))
      

      Your script should now look like this:

      step_7_fer_simple.py

      from torch.utils.data import Dataset
      from torch.autograd import Variable
      import torch.nn as nn
      import torch.nn.functional as F
      import torch.optim as optim
      import numpy as np
      import torch
      import cv2
      import argparse
      
      
      class Fer2013Dataset(Dataset):
          """Face Emotion Recognition dataset.
      
          Utility for loading FER into PyTorch. Dataset curated by Pierre-Luc Carrier
          and Aaron Courville in 2013.
      
          Each sample is 1 x 1 x 48 x 48, and each label is a scalar.
          """
          def __init__(self, path: str):
              """
              Args:
                  path: Path to `.np` file containing sample nxd and label nx1
              """
              with np.load(path) as data:
                  self._samples = data['X']
                  self._labels = data['Y']
              self._samples = self._samples.reshape((-1, 1, 48, 48))
      
              self.X = Variable(torch.from_numpy(self._samples)).float()
              self.Y = Variable(torch.from_numpy(self._labels)).float()
      
          def __len__(self):
              return len(self._labels)
      
      
          def __getitem__(self, idx):
              return {'image': self._samples[idx], 'label': self._labels[idx]}
      
      
      trainset = Fer2013Dataset('data/fer2013_train.npz')
      trainloader = torch.utils.data.DataLoader(trainset, batch_size=32, shuffle=True)
      
      testset = Fer2013Dataset('data/fer2013_test.npz')
      testloader = torch.utils.data.DataLoader(testset, batch_size=32, shuffle=False)
      
      
      class Net(nn.Module):
          def __init__(self):
              super(Net, self).__init__()
              self.conv1 = nn.Conv2d(1, 6, 5)
              self.pool = nn.MaxPool2d(2, 2)
              self.conv2 = nn.Conv2d(6, 6, 3)
              self.conv3 = nn.Conv2d(6, 16, 3)
              self.fc1 = nn.Linear(16 * 4 * 4, 120)
              self.fc2 = nn.Linear(120, 48)
              self.fc3 = nn.Linear(48, 3)
      
          def forward(self, x):
              x = self.pool(F.relu(self.conv1(x)))
              x = self.pool(F.relu(self.conv2(x)))
              x = self.pool(F.relu(self.conv3(x)))
              x = x.view(-1, 16 * 4 * 4)
              x = F.relu(self.fc1(x))
              x = F.relu(self.fc2(x))
              x = self.fc3(x)
              return x
      
      net = Net().float()
      criterion = nn.CrossEntropyLoss()
      optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
      
      
      for epoch in range(2):  # loop over the dataset multiple times
      
          running_loss = 0.0
          for i, data in enumerate(trainloader, 0):
              inputs = Variable(data['image'].float())
              labels = Variable(data['label'].long())
              optimizer.zero_grad()
      
              # forward + backward + optimize
              outputs = net(inputs)
              loss = criterion(outputs, labels)
              loss.backward()
              optimizer.step()
      
              # print statistics
              running_loss += loss.data[0]
              if i % 100 == 0:
                  print('[%d, %5d] loss: %.3f' % (epoch, i, running_loss / (i + 1)))
      

      Save the file and exit the editor once you've verified your code. Then, launch this proof-of-concept training:

      • python step_7_fer_simple.py

      You'll see output similar to the following as the neural network trains:

      Output

      [0, 0] loss: 1.094 [0, 100] loss: 1.049 [0, 200] loss: 1.009 [0, 300] loss: 0.963 [0, 400] loss: 0.935 [1, 0] loss: 0.760 [1, 100] loss: 0.768 [1, 200] loss: 0.775 [1, 300] loss: 0.776 [1, 400] loss: 0.767

      You can then augment this script using a number of other PyTorch utilities to save and load models, output training and validation accuracies, fine-tune a learning-rate schedule, etc. After training for 20 epochs with a learning rate of 0.01 and momentum of 0.9, our neural network attains a 87.9% train accuracy and a 75.5% validation accuracy, a further 6.8% improvement over the most successful least-squares approach thus far at 66.6%. We'll include these additional bells and whistles in a new script.

      Create a new file to hold the final face emotion detector which your live camera feed will use. This script contains the code above along with a command-line interface and an easy-to-import version of our code that will be used later. Additionally, it contains the hyperparameters tuned in advance, for a model with higher accuracy.

      Start with the following imports. This matches our previous file but additionally includes OpenCV as import cv2.

      step_7_fer.py

      from torch.utils.data import Dataset
      from torch.autograd import Variable
      import torch.nn as nn
      import torch.nn.functional as F
      import torch.optim as optim
      import numpy as np
      import torch
      import cv2
      import argparse
      

      Directly beneath these imports, reuse your code from step_7_fer_simple.py to define the neural network:

      step_7_fer.py

      class Net(nn.Module):
          def __init__(self):
              super(Net, self).__init__()
              self.conv1 = nn.Conv2d(1, 6, 5)
              self.pool = nn.MaxPool2d(2, 2)
              self.conv2 = nn.Conv2d(6, 6, 3)
              self.conv3 = nn.Conv2d(6, 16, 3)
              self.fc1 = nn.Linear(16 * 4 * 4, 120)
              self.fc2 = nn.Linear(120, 48)
              self.fc3 = nn.Linear(48, 3)
      
          def forward(self, x):
              x = self.pool(F.relu(self.conv1(x)))
              x = self.pool(F.relu(self.conv2(x)))
              x = self.pool(F.relu(self.conv3(x)))
              x = x.view(-1, 16 * 4 * 4)
              x = F.relu(self.fc1(x))
              x = F.relu(self.fc2(x))
              x = self.fc3(x)
              return x
      

      Again, reuse the code for the Face Emotion Recognition dataset from step_7_fer_simple.py and add it to this file:

      step_7_fer.py

      class Fer2013Dataset(Dataset):
          """Face Emotion Recognition dataset.
          Utility for loading FER into PyTorch. Dataset curated by Pierre-Luc Carrier
          and Aaron Courville in 2013.
          Each sample is 1 x 1 x 48 x 48, and each label is a scalar.
          """
      
          def __init__(self, path: str):
              """
              Args:
                  path: Path to `.np` file containing sample nxd and label nx1
              """
              with np.load(path) as data:
                  self._samples = data['X']
                  self._labels = data['Y']
              self._samples = self._samples.reshape((-1, 1, 48, 48))
      
              self.X = Variable(torch.from_numpy(self._samples)).float()
              self.Y = Variable(torch.from_numpy(self._labels)).float()
      
          def __len__(self):
              return len(self._labels)
      
          def __getitem__(self, idx):
              return {'image': self._samples[idx], 'label': self._labels[idx]}
      

      Next, define a few utilities to evaluate the neural network's performance. First, add an evaluate function which compares the neural network's predicted emotion to the true emotion for a single image:

      step_7_fer.py

      def evaluate(outputs: Variable, labels: Variable, normalized: bool=True) -> float:
          """Evaluate neural network outputs against non-one-hotted labels."""
          Y = labels.data.numpy()
          Yhat = np.argmax(outputs.data.numpy(), axis=1)
          denom = Y.shape[0] if normalized else 1
          return float(np.sum(Yhat == Y) / denom)
      

      Then add a function called batch_evaluate which applies the first function to all images:

      step_7_fer.py

      def batch_evaluate(net: Net, dataset: Dataset, batch_size: int=500) -> float:
          """Evaluate neural network in batches, if dataset is too large."""
          score = 0.0
          n = dataset.X.shape[0]
          for i in range(0, n, batch_size):
              x = dataset.X[i: i + batch_size]
              y = dataset.Y[i: i + batch_size]
              score += evaluate(net(x), y, False)
          return score / n
      

      Now, define a function called get_image_to_emotion_predictor that takes in an image and outputs a predicted emotion, using a pretrained model:

      step_7_fer.py

      def get_image_to_emotion_predictor(model_path='assets/model_best.pth'):
          """Returns predictor, from image to emotion index."""
          net = Net().float()
          pretrained_model = torch.load(model_path)
          net.load_state_dict(pretrained_model['state_dict'])
      
          def predictor(image: np.array):
              """Translates images into emotion indices."""
              if image.shape[2] > 1:
                  image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
              frame = cv2.resize(image, (48, 48)).reshape((1, 1, 48, 48))
              X = Variable(torch.from_numpy(frame)).float()
              return np.argmax(net(X).data.numpy(), axis=1)[0]
          return predictor
      

      Finally, add the following code to define the main function to leverage the other utilities:

      step_7_fer.py

      def main():
          trainset = Fer2013Dataset('data/fer2013_train.npz')
          testset = Fer2013Dataset('data/fer2013_test.npz')
          net = Net().float()
      
          pretrained_model = torch.load("assets/model_best.pth")
          net.load_state_dict(pretrained_model['state_dict'])
      
          train_acc = batch_evaluate(net, trainset, batch_size=500)
          print('Training accuracy: %.3f' % train_acc)
          test_acc = batch_evaluate(net, testset, batch_size=500)
          print('Validation accuracy: %.3f' % test_acc)
      
      
      if __name__ == '__main__':
          main()
      

      This loads a pretrained neural network and evaluates its performance on the provided Face Emotion Recognition dataset. Specifically, the script outputs accuracy on the images we used for training, as well as a separate set of images we put aside for testing purposes.

      Double-check that your file matches the following:

      step_7_fer.py

      from torch.utils.data import Dataset
      from torch.autograd import Variable
      import torch.nn as nn
      import torch.nn.functional as F
      import torch.optim as optim
      import numpy as np
      import torch
      import cv2
      import argparse
      
      class Net(nn.Module):
          def __init__(self):
              super(Net, self).__init__()
              self.conv1 = nn.Conv2d(1, 6, 5)
              self.pool = nn.MaxPool2d(2, 2)
              self.conv2 = nn.Conv2d(6, 6, 3)
              self.conv3 = nn.Conv2d(6, 16, 3)
              self.fc1 = nn.Linear(16 * 4 * 4, 120)
              self.fc2 = nn.Linear(120, 48)
              self.fc3 = nn.Linear(48, 3)
      
          def forward(self, x):
              x = self.pool(F.relu(self.conv1(x)))
              x = self.pool(F.relu(self.conv2(x)))
              x = self.pool(F.relu(self.conv3(x)))
              x = x.view(-1, 16 * 4 * 4)
              x = F.relu(self.fc1(x))
              x = F.relu(self.fc2(x))
              x = self.fc3(x)
              return x
      
      
      class Fer2013Dataset(Dataset):
          """Face Emotion Recognition dataset.
          Utility for loading FER into PyTorch. Dataset curated by Pierre-Luc Carrier
          and Aaron Courville in 2013.
          Each sample is 1 x 1 x 48 x 48, and each label is a scalar.
          """
      
          def __init__(self, path: str):
              """
              Args:
                  path: Path to `.np` file containing sample nxd and label nx1
              """
              with np.load(path) as data:
                  self._samples = data['X']
                  self._labels = data['Y']
              self._samples = self._samples.reshape((-1, 1, 48, 48))
      
              self.X = Variable(torch.from_numpy(self._samples)).float()
              self.Y = Variable(torch.from_numpy(self._labels)).float()
      
          def __len__(self):
              return len(self._labels)
      
          def __getitem__(self, idx):
              return {'image': self._samples[idx], 'label': self._labels[idx]}
      
      
      def evaluate(outputs: Variable, labels: Variable, normalized: bool=True) -> float:
          """Evaluate neural network outputs against non-one-hotted labels."""
          Y = labels.data.numpy()
          Yhat = np.argmax(outputs.data.numpy(), axis=1)
          denom = Y.shape[0] if normalized else 1
          return float(np.sum(Yhat == Y) / denom)
      
      
      def batch_evaluate(net: Net, dataset: Dataset, batch_size: int=500) -> float:
          """Evaluate neural network in batches, if dataset is too large."""
          score = 0.0
          n = dataset.X.shape[0]
          for i in range(0, n, batch_size):
              x = dataset.X[i: i + batch_size]
              y = dataset.Y[i: i + batch_size]
              score += evaluate(net(x), y, False)
          return score / n
      
      
      def get_image_to_emotion_predictor(model_path='assets/model_best.pth'):
          """Returns predictor, from image to emotion index."""
          net = Net().float()
          pretrained_model = torch.load(model_path)
          net.load_state_dict(pretrained_model['state_dict'])
      
          def predictor(image: np.array):
              """Translates images into emotion indices."""
              if image.shape[2] > 1:
                  image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
              frame = cv2.resize(image, (48, 48)).reshape((1, 1, 48, 48))
              X = Variable(torch.from_numpy(frame)).float()
              return np.argmax(net(X).data.numpy(), axis=1)[0]
          return predictor
      
      
      def main():
          trainset = Fer2013Dataset('data/fer2013_train.npz')
          testset = Fer2013Dataset('data/fer2013_test.npz')
          net = Net().float()
      
          pretrained_model = torch.load("assets/model_best.pth")
          net.load_state_dict(pretrained_model['state_dict'])
      
          train_acc = batch_evaluate(net, trainset, batch_size=500)
          print('Training accuracy: %.3f' % train_acc)
          test_acc = batch_evaluate(net, testset, batch_size=500)
          print('Validation accuracy: %.3f' % test_acc)
      
      
      if __name__ == '__main__':
          main(
      

      Save the file and exit your editor.

      As before, with the face detector, download pre-trained model parameters and save them to your assets folder with the following command:

      • wget -O assets/model_best.pth https://github.com/alvinwan/emotion-based-dog-filter/raw/master/src/assets/model_best.pth

      Run the script to use and evaluate the pre-trained model:

      This will output the following:

      Output

      Training accuracy: 0.879 Validation accuracy: 0.755

      At this point, you've built a pretty accurate face-emotion classifier. In essence, our model can correctly disambiguate between faces that are happy, sad, and surprised eight out of ten times. This is a reasonably good model, so you can now move on to using this face-emotion classifier to determine which dog mask to apply to faces.

      Step 8 — Finishing the Emotion-Based Dog Filter

      Before integrating our brand-new face-emotion classifier, we will need animal masks to pick from. We'll use a Dalmation mask and a Sheepdog mask:

      Dalmation mask
      Sheepdog mask

      Execute these commands to download both masks to your assets folder:

      • wget -O assets/dalmation.png https://xpresservers.com/wp-content/uploads/2019/04/1554419827_591_How-To-Apply-Computer-Vision-to-Build-an-Emotion-Based-Dog-Filter-in-Python-3.png # dalmation
      • wget -O assets/sheepdog.png https://xpresservers.com/wp-content/uploads/2019/04/1554419827_102_How-To-Apply-Computer-Vision-to-Build-an-Emotion-Based-Dog-Filter-in-Python-3.png # sheepdog

      Now let's use the masks in our filter. Start by duplicating the step_4_dog_mask.py file:

      • cp step_4_dog_mask.py step_8_dog_emotion_mask.py

      Open the new Python script.

      • nano step_8_dog_emotion_mask.py

      Insert a new line at the top of the script to import the emotion predictor:

      step_8_dog_emotion_mask.py

      from step_7_fer import get_image_to_emotion_predictor
      ...
      

      Then, in the main() function, locate this line:

      step_8_dog_emotion_mask.py

          mask = cv2.imread('assets/dog.png')
      

      Replace it with the following to load the new masks and aggregate all masks into a tuple:

      step_8_dog_emotion_mask.py

          mask0 = cv2.imread('assets/dog.png')
          mask1 = cv2.imread('assets/dalmation.png')
          mask2 = cv2.imread('assets/sheepdog.png')
          masks = (mask0, mask1, mask2)
      

      Add a line break, and then add this code to create the emotion predictor.

      step_8_dog_emotion_mask.py

      
          # get emotion predictor
          predictor = get_image_to_emotion_predictor()
      

      Your main function should now match the following:

      step_8_dog_emotion_mask.py

      def main():
          cap = cv2.VideoCapture(0)
      
          # load mask
          mask0 = cv2.imread('assets/dog.png')
          mask1 = cv2.imread('assets/dalmation.png')
          mask2 = cv2.imread('assets/sheepdog.png')
          masks = (mask0, mask1, mask2)
      
          # get emotion predictor
          predictor = get_image_to_emotion_predictor()
      
          # initialize front face classifier
          ...
      

      Next, locate these lines:

      step_8_dog_emotion_mask.py

      
                  # apply mask
                  frame[y0: y1, x0: x1] = apply_mask(frame[y0: y1, x0: x1], mask)
      

      Insert the following line below the # apply mask line to select the appropriate mask by using the predictor:

      step_8_dog_emotion_mask.py

                  # apply mask
                  mask = masks[predictor(frame[y:y+h, x: x+w])]
                  frame[y0: y1, x0: x1] = apply_mask(frame[y0: y1, x0: x1], mask)
      
      

      The completed file should look like this:

      step_8_dog_emotion_mask.py

      """Test for face detection"""
      
      from step_7_fer import get_image_to_emotion_predictor
      import numpy as np
      import cv2
      
      def apply_mask(face: np.array, mask: np.array) -> np.array:
          """Add the mask to the provided face, and return the face with mask."""
          mask_h, mask_w, _ = mask.shape
          face_h, face_w, _ = face.shape
      
          # Resize the mask to fit on face
          factor = min(face_h / mask_h, face_w / mask_w)
          new_mask_w = int(factor * mask_w)
          new_mask_h = int(factor * mask_h)
          new_mask_shape = (new_mask_w, new_mask_h)
          resized_mask = cv2.resize(mask, new_mask_shape)
      
          # Add mask to face - ensure mask is centered
          face_with_mask = face.copy()
          non_white_pixels = (resized_mask < 250).all(axis=2)
          off_h = int((face_h - new_mask_h) / 2)
          off_w = int((face_w - new_mask_w) / 2)
          face_with_mask[off_h: off_h+new_mask_h, off_w: off_w+new_mask_w][non_white_pixels] = 
               resized_mask[non_white_pixels]
      
          return face_with_mask
      
      def main():
      
          cap = cv2.VideoCapture(0)
          # load mask
          mask0 = cv2.imread('assets/dog.png')
          mask1 = cv2.imread('assets/dalmation.png')
          mask2 = cv2.imread('assets/sheepdog.png')
          masks = (mask0, mask1, mask2)
      
          # get emotion predictor
          predictor = get_image_to_emotion_predictor()
      
          # initialize front face classifier
          cascade = cv2.CascadeClassifier("assets/haarcascade_frontalface_default.xml")
      
          while True:
              # Capture frame-by-frame
              ret, frame = cap.read()
              frame_h, frame_w, _ = frame.shape
      
              # Convert to black-and-white
              gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
              blackwhite = cv2.equalizeHist(gray)
      
              rects = cascade.detectMultiScale(
                  blackwhite, scaleFactor=1.3, minNeighbors=4, minSize=(30, 30),
                  flags=cv2.CASCADE_SCALE_IMAGE)
      
              for x, y, w, h in rects:
                  # crop a frame slightly larger than the face
                  y0, y1 = int(y - 0.25*h), int(y + 0.75*h)
                  x0, x1 = x, x + w
                  # give up if the cropped frame would be out-of-bounds
                  if x0 < 0 or y0 < 0 or x1 > frame_w or y1 > frame_h:
                      continue
                  # apply mask
                  mask = masks[predictor(frame[y:y+h, x: x+w])]
                  frame[y0: y1, x0: x1] = apply_mask(frame[y0: y1, x0: x1], mask)
      
              # Display the resulting frame
              cv2.imshow('frame', frame)
              if cv2.waitKey(1) & 0xFF == ord('q'):
                  break
      
          cap.release()
          cv2.destroyAllWindows()
      
      if __name__ == '__main__':
          main()
      

      Save and exit your editor. Now launch the script:

      • python step_8_dog_emotion_mask.py

      Now try it out! Smiling will register as "happy" and show the original dog. A neutral face or a frown will register as "sad" and yield the dalmation. A face of "surprise," with a nice big jaw drop, will yield the sheepdog.

      GIF for emotion-based dog filter

      This concludes our emotion-based dog filter and foray into computer vision.

      Conclusion

      In this tutorial, you built a face detector and dog filter using computer vision and employed machine learning models to apply masks based on detected emotions.

      Machine learning is widely applicable. However, it's up to the practitioner to consider the ethical implications of each application when applying machine learning. The application you built in this tutorial was a fun exercise, but remember that you relied on OpenCV and an existing dataset to identify faces, rather than supplying your own data to train the models. The data and models used have significant impacts on how a program works.

      For example, imagine a job search engine where the models were trained with data about candidates. such as race, gender, age, culture, first language, or other factors. And perhaps the developers trained a model that enforces sparsity, which ends up reducing the feature space to a subspace where gender explains most of the variance. As a result, the model influences candidate job searches and even company selection processes based primarily on gender. Now consider more complex situations where the model is less interpretable and you don't know what a particular feature corresponds to. You can learn more about this in Equality of Opportunity in Machine Learning by Professor Moritz Hardt at UC Berkeley.

      There can be an overwhelming magnitude of uncertainty in machine learning. To understand this randomness and complexity, you'll have to develop both mathematical intuitions and probabilistic thinking skills. As a practitioner, it is up to you to dig into the theoretical underpinnings of machine learning.



      Source link

      How To Detect and Extract Faces from an Image with OpenCV and Python


      The author selected the Open Internet/Free Speech Fund to receive a donation as part of the Write for DOnations program.

      Introduction

      Images make up a large amount of the data that gets generated each day, which makes the ability to process these images important. One method of processing images is via face detection. Face detection is a branch of image processing that uses machine learning to detect faces in images.

      A Haar Cascade is an object detection method used to locate an object of interest in images. The algorithm is trained on a large number of positive and negative samples, where positive samples are images that contain the object of interest. Negative samples are images that may contain anything but the desired object. Once trained, the classifier can then locate the object of interest in any new images.

      In this tutorial, you will use a pre-trained Haar Cascade model from OpenCV and Python to detect and extract faces from an image. OpenCV is an open-source programming library that is used to process images.

      Prerequisites

      Step 1 — Configuring the Local Environment

      Before you begin writing your code, you will first create a workspace to hold the code and install a few dependencies.

      Create a directory for the project with the mkdir command:

      Change into the newly created directory:

      Next, you will create a virtual environment for this project. Virtual environments isolate different projects so that differing dependencies won't cause any disruptions. Create a virtual environment named face_scrapper to use with this project:

      • python3 -m venv face_scrapper

      Activate the isolated environment:

      • source face_scrapper/bin/activate

      You will now see that your prompt is prefixed with the name of your virtual environment:

      Now that you've activated your virtual environment, you will use nano or your favorite text editor to create a requirements.txt file. This file indicates the necessary Python dependencies:

      Next, you need to install three dependencies to complete this tutorial:

      • numpy: numpy is a Python library that adds support for large, multi-dimensional arrays. It also includes a large collection of mathematical functions to operate on the arrays.
      • opencv-utils: This is the extended library for OpenCV that includes helper functions.
      • opencv-python: This is the core OpenCV module that Python uses.

      Add the following dependencies to the file:

      requirements.txt

      numpy 
      opencv-utils
      opencv-python
      

      Save and close the file.

      Install the dependencies by passing the requirements.txt file to the Python package manager, pip. The -r flag specifies the location of requirements.txt file.

      • pip install -r requirements.txt

      In this step, you set up a virtual environment for your project and installed the necessary dependencies. You're now ready to start writing the code to detect faces from an input image in next step.

      Step 2 — Writing and Running the Face Detector Script

      In this section, you will write code that will take an image as input and return two things:

      • The number of faces found in the input image.
      • A new image with a rectangular plot around each detected face.

      Start by creating a new file to hold your code:

      In this new file, start writing your code by first importing the necessary libraries. You will import two modules here: cv2 and sys. The cv2 module imports the OpenCV library into the program, and sys imports common Python functions, such as argv, that your code will use.

      app.py

      import cv2
      import sys
      

      Next, you will specify that the input image will be passed as an argument to the script at runtime. The Pythonic way of reading the first argument is to assign the value returned by sys.argv[1] function to an variable:

      app.py

      ...
      imagePath = sys.argv[1]
      

      A common practice in image processing is to first convert the input image to gray scale. This is because detecting luminance, as opposed to color, will generally yield better results in object detection. Add the following code to take an input image as an argument and convert it to grayscale:

      app.py

      ...
      image = cv2.imread(imagePath)
      gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
      

      The .imread() function takes the input image, which is passed as an argument to the script, and converts it to an OpenCV object. Next, OpenCV's .cvtColor() function converts the input image object to a grayscale object.

      Now that you've added the code to load an image, you will add the code that detects faces in the specified image:

      app.py

      ...
      faceCascade = cv2.CascadeClassifier(cv2.data.haarcascades + "haarcascade_frontalface_default.xml")
      faces = faceCascade.detectMultiScale(
              gray,
              scaleFactor=1.3,
              minNeighbors=3,
              minSize=(30, 30)
      ) 
      
      print("Found {0} Faces!".format(len(faces)))
      
      

      This code will create a faceCascade object that will load the Haar Cascade file with the cv2.CascadeClassifier method. This allows Python and your code to use the Haar Cascade.

      Next, the code applies OpenCV's .detectMultiScale() method on the faceCascade object. This generates a list of rectangles for all of the detected faces in the image. The list of rectangles is a collection of pixel locations from the image, in the form of Rect(x,y,w,h).

      Here is a summary of the other parameters your code uses:

      • gray: This specifies the use of the OpenCV grayscale image object that you loaded earlier.
      • scaleFactor: This parameter specifies the rate to reduce the image size at each image scale. Your model has a fixed scale during training, so input images can be scaled down for improved detection. This process stops after reaching a threshold limit, defined by maxSize and minSize.
      • minNeighbors: This parameter specifies how many neighbors, or detections, each candidate rectangle should have to retain it. A higher value may result in less false positives, but a value too high can eliminate true positives.
      • minSize: This allows you to define the minimum possible object size measured in pixels. Objects smaller than this parameter are ignored.

      After generating a list of rectangles, the faces are then counted with the len function. The number of detected faces are then returned as output after running the script.

      Next, you will use OpenCV's .rectangle() method to draw a rectangle around the detected faces:

      app.py

      ...
      for (x, y, w, h) in faces:
          cv2.rectangle(image, (x, y), (x+w, y+h), (0, 255, 0), 2)
      
      

      This code uses a for loop to iterate through the list of pixel locations returned from faceCascade.detectMultiScale method for each detected object. The rectangle method will take four arguments:

      • image tells the code to draw rectangles on the original input image.
      • (x,y), (x+w, y+h) are the four pixel locations for the detected object. rectangle will use these to locate and draw rectangles around the detected objects in the input image.
      • (0, 255, 0) is the color of the shape. This argument gets passed as a tuple for BGR. For example, you would use (255, 0, 0) for blue. We are using green in this case.
      • 2 is the thickness of the line measured in pixels.

      Now that you've added the code to draw the rectangles, use OpenCV's .imwrite() method to write the new image to your local filesystem as faces_detected.jpg. This method will return true if the write was successful and false if it wasn't able to write the new image.

      app.py

      ...
      status = cv2.imwrite('faces_detected.jpg', image)
      

      Finally, add this code to print the return the true or false status of the .imwrite() function to the console. This will let you know if the write was successful after running the script.

      app.py

      ...
      print ("Image faces_detected.jpg written to filesystem: ",status)
      

      The completed file will look like this:

      app.py

      import cv2
      import sys
      
      imagePath = sys.argv[1]
      
      image = cv2.imread(imagePath)
      gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
      
      faceCascade = cv2.CascadeClassifier(cv2.data.haarcascades + "haarcascade_frontalface_default.xml")
      faces = faceCascade.detectMultiScale(
          gray,
          scaleFactor=1.3,
          minNeighbors=3,
          minSize=(30, 30)
      )
      
      print("[INFO] Found {0} Faces!".format(len(faces)))
      
      for (x, y, w, h) in faces:
          cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)
      
      status = cv2.imwrite('faces_detected.jpg', image)
      print("[INFO] Image faces_detected.jpg written to filesystem: ", status)
      

      Once you've verified that everything is entered correctly, save and close the file.

      Note: This code was sourced from the publicly available OpenCV documentation.

      Your code is complete and you are ready to run the script.

      Step 3 — Running the Script

      In this step, you will use an image to test your script. When you find an image you'd like to use to test, save it in the same directory as your app.py script. This tutorial will use the following image:

      Input Image of four people looking at phones

      If you would like to test with the same image, use the following command to download it:

      • curl -O https://xpresservers.com/wp-content/uploads/2019/03/How-To-Detect-and-Extract-Faces-from-an-Image-with-OpenCV-and-Python.png

      Once you have an image to test the script, run the script and provide the image path as an argument:

      • python app.py path/to/input_image

      Once the script finishes running, you will receive output like this:

      Output

      [INFO] Found 4 Faces! [INFO] Image faces_detected.jpg written to filesystem: True

      The true output tells you that the updated image was successfully written to the filesystem. Open the image on your local machine to see the changes on the new file:

      Output Image with detected faces

      You should see that your script detected four faces in the input image and drew rectangles to mark them. In the next step, you will use the pixel locations to extract faces from the image.

      Step 4 — Extracting Faces and Saving them Locally (Optional)

      In the previous step, you wrote code to use OpenCV and a Haar Cascade to detect and draw rectangles around faces in an image. In this section, you will modify your code to extract the detected faces from the image into their own files.

      Start by reopening the app.py file with your text editor:

      Next, add the highlighted lines under the cv2.rectangle line:

      app.py

      ...
      for (x, y, w, h) in faces:
          cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)
          roi_color = image[y:y + h, x:x + w] 
          print("[INFO] Object found. Saving locally.") 
          cv2.imwrite(str(w) + str(h) + '_faces.jpg', roi_color) 
      ...
      

      The roi_color object plots the pixel locations from the faces list on the original input image. The x, y, h, and w variables are the pixel locations for each of the objects detected from faceCascade.detectMultiScale method. The code then prints output stating that an object was found and will be saved locally.

      Once that is done, the code saves the plot as a new image using the cv2.imwrite method. It appends the width and height of the plot to the name of the image being written to. This will keep the name unique in case there are multiple faces detected.

      The updated app.py script will look like this:

      app.py

      import cv2
      import sys
      
      imagePath = sys.argv[1]
      
      image = cv2.imread(imagePath)
      gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
      
      faceCascade = cv2.CascadeClassifier(cv2.data.haarcascades + "haarcascade_frontalface_default.xml")
      faces = faceCascade.detectMultiScale(
          gray,
          scaleFactor=1.3,
          minNeighbors=3,
          minSize=(30, 30)
      )
      
      print("[INFO] Found {0} Faces.".format(len(faces)))
      
      for (x, y, w, h) in faces:
          cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)
          roi_color = image[y:y + h, x:x + w]
          print("[INFO] Object found. Saving locally.")
          cv2.imwrite(str(w) + str(h) + '_faces.jpg', roi_color)
      
      status = cv2.imwrite('faces_detected.jpg', image)
      print("[INFO] Image faces_detected.jpg written to filesystem: ", status)
      

      To summarize, the updated code uses the pixel locations to extract the faces from the image into a new file. Once you have finished updating the code, save and close the file.

      Now that you've updated the code, you are ready to run the script once more:

      • python app.py path/to/image

      You will see the similar output once your script is done processing the image:

      Output

      [INFO] Found 4 Faces. [INFO] Object found. Saving locally. [INFO] Object found. Saving locally. [INFO] Object found. Saving locally. [INFO] Object found. Saving locally. [INFO] Image faces_detected.jpg written to file-system: True

      Depending on how many faces are in your sample image, you may see more or less output.

      Looking at the contents of the working directory after the execution of the script, you'll see files for the head shots of all faces found in the input image.

      Directory Listing

      You will now see head shots extracted from the input image collected in the working directory:

      Extracted Faces

      In this step, you modified your script to extract the detected objects from the input image and save them locally.

      Conclusion

      In this tutorial, you wrote a script that uses OpenCV and Python to detect, count, and extract faces from an input image. You can update this script to detect different objects by using a different pre-trained Haar Cascade from the OpenCV library, or you can learn how to train your own Haar Cascade.



      Source link

      How To Create an OAuth App with the Linode Python API Library


      Updated by Linode

      Contributed by

      Linode

      Linode supports the OAuth 2 authorization protocol. OAuth 2 allows a user to safely grant a third-party app permission to act on their behalf. This means that a user could authorize an app to access data and / or make changes to their Linode account and services that are exposed by the Linode APIv4. For example, an app could create or destroy Linodes, manage a NodeBalancer, or alter a domain.

      This guide will show you how to create a simple OAuth application using Flask and the Linode Python API library. This app allows a user to log in with their Linode account and create a Linode with a StackScript. The complete code for this example is available in the Linode APIv4 Python library example repository.

      Before You Begin

      1. Normally, in order to create an OAuth app with Linode your server must have HTTPS enabled. The only exceptions to this rule are localhost addresses, which can use HTTP. As this guide is just a primer and is not intended to supply production ready code, we will be working with a local workstation, using localhost. If you choose to create an app for production, you will need to generate SSL certificates for HTTPS access.

      2. Ensure that Python 3 is installed on your workstation.

      Obtaining a Client ID and a Client Secret

      In order for Linode to verify the identity of your app, called a client, you will need to generate a set of credentials, specifically a client ID and a client secret.

      1. Log in to the Linode Cloud Manager and navigate to your Account Profile.

        OAuth Account Profile

      2. From there, click on the My Apps tab and select Create My App. You will be prompted to supply a label for your app and a callback URL. We will discuss the role of the callback URL in depth later in this guide. For now you can supply the following URL:

        http://localhost:5000/auth_callback
        

        Leave Public unchecked and click Submit.

        OAuth Account Profile

      3. A window will appear with your client secret. Copy this down somewhere secure, as once you exit this window you will not be able to retrieve the client secret again, and will be forced to generate a new one.

        OAuth Account Profile

      4. Once you exit the client secret window your app will appear as part of a list of apps. Note your client ID, as this is the last piece of information you will need to verify your app’s identity.

        OAuth Account Profile

      In summary, you should have these three bits of information, with values similar to the ones provided here:

      OAuth 2 Authentication Exchange

      The OAuth 2 workflow is a series of exchanges between your third-party app and Linode. Below is an explanation of these exchanges.

      1. The end user visits your client application’s website and attempts to login.
      2. Your client application redirects the end user to the authentication server (https://login.linode.com) with your client application’s client ID and requested OAuth scopes, which appear in the URL of the login page.
      3. The end user inputs their username and password to the authorization server and authorizes the login.
      4. The authorization server redirects the end user back to your client application with a temporary authorization code (sometimes called an exchange code) in the URL.
      5. The client application issues a POST request to the authentication server containing the authorization code and the client application’s client secret.
      6. The authentication server responds to the client application with a newly issued OAuth access token.

      In the following sections you will write the code to perform each one of these steps, using the Linode Python API library.

      Setup Your Development Environment

      1. Create a project folder and move into that folder.

        mkdir ~/linode-oauth-project && cd ~/linode-oauth-project
        
      2. For this project, you will need to use pip to download and install the required Python libraries. Install pip if you do not already have it:

        apt install python-pip
        
      3. Install the required Python libraries:

        pip install flask flask-session linode_api4
        

      Configure Your App

      In a text editor, create a file named config.py. Add the following variables and values, being sure to change the values to your own.

      The StackScript used in this example is for demo purposes. To explore other available StackScripts, visit the Linode StackScript Library. Note that the stackscript_id does not have quotation marks around it. The secret key is used for serializing session data, and should be a value only you know.

      config.py
      1
      2
      3
      4
      5
      
      client_id = 'ce571a8cdad1ba4a0a7d'
      client_secret = 'fab8e2222e83b9b2f50a76012122ec20a5acb005ed088f3fccda2c9c2c4e1cbd'
      stackscript_id = 320826
      application_name = 'my-application-name'
      secret_key = 'my-secret-key'

      Author an OAuth2 App

      In this section, you will write the code for the app.

      Include Imports

      Ensure you are in the linode-oauth-project directory and create and open a file called app.py in the text editor of your choice. Include the following libraries:

      app.py
      1
      2
      3
      4
      5
      6
      
      import re
      from flask import Flask, redirect, request, render_template, session, send_from_directory
      from flask_session import Session
      from linode_api4 import (LinodeClient, LinodeLoginClient, StackScript, Image, Region, Type, OAuthScopes)
      
      import config

      Set Up Flask and Session Key

      Copy in the following code to set up Flask and the session secret key:

      app.py
      1
      2
      3
      4
      
      ...
      
      app=Flask(__name__)
      app.config['SECRET_KEY'] = config.secret_key

      Create a Function to Return the Linode Login Client

      In app.py add the following function to return the LinodeLoginClient class. The LinodeLoginClient class is the library’s OAuth interface. Note that we are passing the client_id and client_secret parameters from our config.py file to the class:

      ~/linode-oauth-project/app.py
      1
      2
      3
      4
      
      ...
      
      def get_login_client():
          return LinodeLoginClient(config.client_id, config.client_secret)

      Create an Index Route

      In Flask you can create HTTP endpoints with routes. The index route, defined in the code below at the document root /, will be the route the user will see when they navigate to http://localhost:5000/. This route will be responsible for displaying the available Linode plan types, the available regions, and the StackScript-compatible images that a user will choose from when creating their new Linode.

      To query a list of available plan types and regions you can use the LinodeClient class, which is an interface for Linode’s APIv4. Viewing the Linode plan types and regions does not require any sort of authorization, so you can provide a dummy value of no-token to instantiate the class:

      ~/linode-oauth-project/app.py
       1
       2
       3
       4
       5
       6
       7
       8
       9
      10
      11
      12
      13
      14
      
      ...
      
      @app.route('/')
      def index():
          client = LinodeClient('no-token')
          types = client.linode.types(Type.label.contains("Linode"))
          regions = client.regions()
          stackscript = StackScript(client, config.stackscript_id)
          return render_template('configure.html',
              types=types,
              regions=regions,
              application_name=config.application_name,
              stackscript=stackscript
          )

      It is important to note that the two API queries in the above code are slightly different from one another. The client.regions method is a top-level method, just as it appears in the Linode API. The client.linode.types method, on the other hand, is part of the Linode group, which is a collection of methods that deal with Linodes. Again, this is because Linode endpoints are grouped that way in the API. Some methods in the Linode Python library are top level, such as domain_create, while others, like networking.ip_assign, are part of a group. For more information on the top-level methods and groupings, consult the library documentation.

      In addition to querying the API, the above route also renders the configure.html template by passing it the types, regions, application name, and StackScript object. The StackScript object contains a list of StackScript compatible images. We will cover templating in a later section.

      Create a Login Route

      Next, create a login route in app.py. This route will perform two functions. First, it will serialize the user’s plan type, region, and image selections into the session.

      Second, this route will redirect the user to Linode’s login page where they will be prompted to authorize your client app and the scopes you have requested for it. Scopes are sets of permissions that define the access level of your client app. For instance, to create a Linode, your end user must authorize the OAuthScopes.Linodes.create scope.

      ~/linode-oauth-project/app.py
      1
      2
      3
      4
      5
      6
      7
      8
      9
      
      ...
      
      @app.route('/', methods=["POST"])
      def start_auth():
          login_client = get_login_client()
          session['dc'] = request.form['region']
          session['distro'] = request.form['distribution']
          session['type'] = request.form['type']
          return redirect(login_client.generate_login_url(scopes=OAuthScopes.Linodes.create))

      When the user returns to your app from the Linode login page, they will be directed to the callback URL.

      Note

      Below is a list of available scopes:

      • OAuthScopes.Linodes
      • OAuthScopes.Domains
      • OAuthScopes.StackScripts
      • OAuthScopes.Users
      • OAuthScopes.NodeBalancers
      • OAuthScopes.Tokens
      • OAuthScopes.IPs
      • OAuthScopes.Tickets
      • OAuthScopes.Clients
      • OAuthScopes.Account
      • OAuthScopes.Events
      • OAuthScopes.Volumes

      Each scope is broken into five permissions: view, create, modify, delete, and all. The all permission encompasses the other four permissions.

      Manage the OAuth 2 Callback URL

      The OAuth 2 callback URL has two main responsibilities. Its first responsibility is to help prove the identity of the client application. When a user attempts to log in to Linode through OAuth, instead of redirecting the user back to the page they came from, Linode’s OAuth implementation matches the client ID to the callback URL you have registered with your app on Linode’s system. This ensures that a nefarious third party can’t just steal the client ID, which is public, and attempt to authorize their own app with it.

      The callback URL’s second responsibility is to kick off the process of exchanging an authorization code for an access token. This second process is done over POST, and so it doesn’t require the user to physically leave the page they are returned to after they log in to Linode. Now you will write the code that satisfies this second responsibility.

      In app.py, add the following lines:

      ~/linode-oauth-project/app.py
       1
       2
       3
       4
       5
       6
       7
       8
       9
      10
      11
      12
      13
      14
      15
      16
      17
      18
      19
      20
      21
      
      ...
      
      @app.route('/auth_callback')
      def auth_callback():
          code = request.args.get('code')
          login_client = get_login_client()
          token, scopes, _, _ = login_client.finish_oauth(code)
      
          # ensure we have sufficient scopes
          if not OAuthScopes.Linodes.create in scopes:
              return render_template('error.html', error='Insufficient scopes granted to deploy {}'
                      .format(config.application_name))
      
          (linode, password) = make_instance(token, session['type'], session['dc'], session['distro'])
      
          get_login_client().expire_token(token)
          return render_template('success.html',
              password=password,
              linode=linode,
              application_name=config.application_name
          )

      Let’s take a look at what each of the parts of this section does.

      First, a route is defined for the callback with @app.route(), then a function called auth_callback is defined that will run whenever this route is accessed:

      ~/linode-oauth-project/app.py
      1
      2
      3
      4
      
      ...
      @app.route('/auth_callback')
      def auth_callback():
      ...

      When the user is returned to the callback URL, an authorization code is appended to the URL. The variable code is set to retrieve this value from the URL’s request arguments:

      ~/linode-oauth-project/app.py
      1
      2
      3
      
      ...
          code = request.args.get('code')
      ...

      Then you retrieve an instance of the LinodeLoginClient class:

      ~/linode-oauth-project/app.py
      1
      2
      3
      
      ...
          login_client = get_login_client()
      ...

      Once you have the LinodeLoginClient class, you can pass the authorization code to the finish_oauth method, which is a helper method that will manage the authorization code to OAuth token exchange. This method returns an OAuth token, and the scopes the user has agreed upon.

      ~/linode-oauth-project/app.py
      1
      2
      3
      
      ...
          token, scopes, _, _ = login_client.finish_oauth(code)
      ...

      The next section compares the scopes your app requested from the user to the scopes returned by Linode’s OAuth login page. If the returned scopes do not include the correct scopes, in this case the OAuthScopes.Linode.create scope, then an error template is rendered and an error message is displayed:

      ~/linode-oauth-project/app.py
      1
      2
      3
      4
      5
      6
      
      ...
          # ensure we have sufficient scopes
          if not OAuthScopes.Linodes.create in scopes:
              return render_template('error.html', error='Insufficient scopes granted to deploy {}'
                      .format(config.application_name))
      ...

      Once your app has determined that it has the correct permissions, it creates the Linode using the Linode plan type, the region, and the image that the app serialized into session storage. You will create the make_instance function in the next step. The make_instance function returns the linode object, which contains the Linode’s label, group, and IP address, and the function also returns a randomly generated password:

      ~/linode-oauth-project/app.py
      1
      2
      3
      
      ...
          (linode, password) = make_instance(token, session['type'], session['dc'], session['distro'])
      ...

      Once the Linode has been created, the app expires the OAuth access token. Expiring tokens after use is a strong security measure but if your app is performing many actions on behalf of the user, you might find that time-based expiration scheme is more suitable to your needs. The app then renders the success template by passing it the linode object, the password, and application name:

      ~/linode-oauth-project/app.py
      1
      2
      3
      4
      5
      6
      7
      
      ...
          get_login_client().expire_token(token)
          return render_template('success.html',
              password=password,
              linode=linode,
              application_name=config.application_name
          )

      Create a Function to Deploy a Linode

      Now, create the make_instance function that you referenced above:

      ~/linode-oauth-project/app.py
       1
       2
       3
       4
       5
       6
       7
       8
       9
      10
      11
      12
      
      ...
      
      def make_instance(token, type_id, region_id, distribution_id):
          client = LinodeClient('{}'.format(token))
          stackscript = StackScript(client, config.stackscript_id)
          (linode, password) = client.linode.instance_create(type_id, region_id,
                  group=config.application_name,
                  image=distribution_id, stackscript=stackscript.id)
      
          if not linode:
              raise RuntimeError("it didn't work")
          return linode, password

      The make_instance function takes an OAuth access token, the type ID, the region ID, and the image (Linux distribution) ID as parameters. It creates an instance of the LinodeClient class, and unlike the instance of LinodeClient used earlier in the guide, this one requires an OAuth token because you will be using it to create a Linode. The function then creates a Linode using the linode.instance_create method, returning the linode object and the password.

      Finally, if there was an error with the creation of the Linode, the if not linode statement will raise a runtime error.

      Set the name Variable

      At the end of your app.py, paste in the following code to make sure you can run your app:

      ~/linode-oauth-project/app.py
      1
      2
      3
      
      if __name__ == '__main__':
          app.debug=True
          app.run()

      Create App Templates

      Now that you have written the backend code for your app, you’ll need to create a frontend user interface. Begin by creating a templates directory in your project directory and moving into it:

      mkdir ~/linode-oauth-project/templates && cd ~/linode-oauth-project/templates
      

      Using your preferred text editor, create and open base.html. This will be the base template from which your other templates will inherit their stylesheets and JavaScript files:

      ~/linode-oauth-project/templates/base.html
       1
       2
       3
       4
       5
       6
       7
       8
       9
      10
      11
      12
      13
      14
      15
      16
      17
      18
      19
      20
      21
      22
      23
      24
      25
      26
      27
      28
      29
      30
      31
      32
      33
      34
      35
      36
      37
      38
      39
      40
      41
      42
      43
      44
      45
      46
      47
      48
      49
      50
      51
      52
      53
      54
      55
      56
      57
      58
      59
      60
      61
      62
      63
      64
      65
      66
      67
      
      <html>
      <head>
          <title>Install On Linode</title>
          <link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.6/css/bootstrap.min.css"
              integrity="sha384-1q8mTJOASx8j1Au+a5WDVnPi2lkFfwwEAa8hDDdjZlpLegxhjVME1fgjWPGmkzs7" crossorigin="anonymous">
          <style>
              body{
                  text-align: center;
                  background-color: #333333;
              }
              .form-group{
                  display: inline-block;
                  text-align: left;
                  width: 250px;
                  border: 1px solid #cccccc;
                  margin: 5px;
                  padding: 5px;
              }
              .form-group label{
                  color: #337ab7;
              }
              .form-group select{
                  font-size: 16px;
                  outline: none;
                  border: 0px solid #000000;
                  box-shadow: inset 0 1px 1px rgba(0,0,0,0);
                  -webkit-box-shadow: inset 0 1px 1px rgba(0,0,0,0);
              }
              .form-group select:focus{
                  box-shadow: inset 0 1px 1px rgba(0,0,0,0);
                  -webkit-box-shadow: inset 0 1px 1px rgba(0,0,0,0);
              }
              .btn-lg{
                  border-radius: 0px;
                  margin-top: 20px;
              }
              .row{
                  margin-bottom: 20px
              }
              .pop{
                  color: #337ab7;
                  font-weight: bold
              }
              code{
                  color: #337ab7;
                  background-color: #eeeeee
              }
              .boxy{
                  border: 1px solid #cccccc;
                  width: 400px;
                  background-color: #f9f9f9;
                  margin: auto;
                  padding: 10px;
              }
          </style>
      </head>
      <body>
          <div class='container' style='background-color: white; border-left: grey; border-right: grey; height: 100%; padding: 20px;'>
              {% block content %}
              {% endblock %}
          </div>
      
          <script src="https://ajax.googleapis.com/ajax/libs/jquery/1.11.3/jquery.min.js"></script>
          <script src="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.6/js/bootstrap.min.js"
              integrity="sha384-0mSbJDEHialfmuBBQP6A4Qrprq5OVfW37PRR3j5ELqxss1yVqOtnepnHVP9aJ7xS"
              crossorigin="anonymous"></script>
      </body>

      The important thing to note in the above template is the Jinja2 templating tags. They are:

      {% block content %}
      {% endblock %}
      

      As you will see, any template that extends the base.html template and includes code between the opening and closing content block, will render the code laid out by base.html.

      Create a file called configure.html, which will be the UI a user will see when they reach the document root endpoint (/). Copy in the following code:

      templates/configure.html
       1
       2
       3
       4
       5
       6
       7
       8
       9
      10
      11
      12
      13
      14
      15
      16
      17
      18
      19
      20
      21
      22
      23
      24
      25
      26
      27
      28
      29
      30
      31
      32
      33
      34
      35
      36
      37
      38
      39
      40
      41
      42
      43
      44
      45
      46
      47
      48
      49
      50
      51
      52
      
      {% extends 'base.html' %}
      {% block content %}
          <form method="POST">
              <div class='row'>
                  <h1>Deploy <span style="color: #337ab7;">{{application_name}}</span> to a Linode</h1>
                  <p>
                      This will create a brand new Linode running {{application_name}} on your
                      account and give you the credentials.
                  </p>
              </div>
              <div class='row'>
                  <div class='form-group'>
                      <label for='type'>Type</label>
                      <select name='type'i id='type' class='form-control'
                          onblur="blurring(this)" onfocus="focusing(this)">
                          {% for s in types %}
                              <option value="{{s.id}}">{{s.label}}</option>
                          {% endfor %}
                      </select>
                  </div>
                  <div class='form-group'>
                      <label for='region'>Region</label>
                      <select name='region' id='region' class='form-control'
                          onblur="blurring(this)" onfocus="focusing(this)">
                          {% for o in regions %}
                              <option value="{{o.id}}">{{o.id}}</option>
                          {% endfor %}
                      </select>
                  </div>
                  <div class='form-group'>
                      <label for='distribution'>Images</label>
                      <select name='distribution' id='distribution' class='form-control'
                          onblur="blurring(this)" onfocus="focusing(this)">
                          {% for d in stackscript.images %}
                              <option value="{{d.id.id}}">{{d.id.id}}</option>
                          {% endfor %}
                      </select>
                  </div>
              </div>
              <div class='row'>
                  <input type="submit" value="Deploy Linode" class='btn btn-primary btn-lg'/>
              </div>
          </form>
          <script>
              function focusing(ele){
                  ele.parentElement.style.borderColor = "#337ab7";
              }
              function blurring(ele){
                  ele.parentElement.style.borderColor = "#cccccc";
              }
          </script>
      {% endblock %}

      Here the template begins with two statements: {% extends 'base.html' %} and a {% block content %} statement. These two tags tell Jinja2 to extend the code within base.html, and to place everything within {% block content %} ... {% endblock %} in configure.html between the corresponding {% block content %} ... {% endblock %} tags in base.html.

      configure.html includes Jinja2 logic, with the inclusion of for statements like {% for o in regions %}. These statements are like for statements in other languages, and are used to iterate over an array or list. In this example, it is iterating over the regions that we passed to the template from the index route. configure.html also contains variables, which are denoted by double curly brackets: {{ s.id }}.

      Create another file called error.html. This will be the template that appears whenever there is an error in the Linode deployment. Copy in the following code:

      templates/error.html
       1
       2
       3
       4
       5
       6
       7
       8
       9
      10
      
      {% extends 'base.html' %}
      {% block content %}
          <div class='row'>
              <h1 class="pop">Error</h1>
              <p>{{error}}</p>
          </div>
          <div class='row' style='margin-top: 20px'>
              <a href='/' class='btn btn-lg btn-default'>Try Again</a>
          </div>
      {% endblock %}

      This template works the same way that configure.html does, by extending base.html and providing its own content block.

      Lastly, create another file called success.html. This file follows the pattern set by configure.html and error.html, and will present the user with a confirmation message whenever a Linode is successfully created. This message includes the Linode’s label, group, IP address, and password:

      templates/success.html
       1
       2
       3
       4
       5
       6
       7
       8
       9
      10
      11
      12
      13
      14
      15
      16
      17
      
      {% extends 'base.html' %}
      {% block content %}
          <div class='row'>
              <h1>Success!</h1>
              <p>{{application_name}} has been deployed to <span class="pop">{{linode.label}}</span> in the {{linode.group}} group.</p>
          </div>
          <div class='row'>
              <div class='boxy'>
                  <p>You can access your Linode with the following command:</p>
                  <code>ssh root@{{linode.ipv4[0]}}</code>
                  <br />
                  <br />
                  <p>Your root password is:</p>
                  <code>{{password}}</code>
              </div>
          </div>
      {% endblock %}

      Run Your App

      You are now ready to run your app. Change back to your project’s main directory:

      cd ~/linode-oauth-project
      

      Run the app.py script:

      python3 app.py
      

      Open your browser to the following URL:

      http://localhost:5000/
      

      You should be greeted with your new app. Select a plan, a region, and an image to deploy a Linode using the Linode API Python library.

      Next Steps

      The app you’ve created shows off some of the aspects of the Linode API Python library. You can use LinodeLoginClient to authorize your OAuth app with the appropriate scopes, and can create Linodes through the use of LinodeClient.

      In extending this app, you might want to add multiple functionalities, like creating NodeBalancers from a list of available Linodes, or managing domains. To achieve this goal you’ll probably want to separate the login logic from the Linode creation logic. One way to do this would be store the OAuth token in the session, implementing a time-based expiration mechanism to expire your tokens instead.

      More Information

      You may wish to consult the following resources for additional information on this topic. While these are provided in the hope that they will be useful, please note that we cannot vouch for the accuracy or timeliness of externally hosted materials.

      Find answers, ask questions, and help others.

      This guide is published under a CC BY-ND 4.0 license.



      Source link