The Adventure Works data science team wants to use deep learning techniques to build an image classification model that can identify each class of product the company sells. Specifically, the team wants to develop a convolutional neural network (CNN) model. You can use your choice of deep learning framework from PyTorch or Keras.
- A Data Science Virtual Machine (DSVM)
- The resized gear image data fom the previous challenges.
There are three tasks in this challenge:
- Explore a sample convolutional neural network model.
- Create and train a convolutional neural network (CNN) model.
- Use your model with new data.
Decide as a team which deep learning framework(s) you want to explore - your coach can help you. If you wish, different team members the team can explore PyTorch and Keras; but the important thing is that as a team, you must end up with a solution that meets the challenge success critera - so decide how you want to divide the work and create a shared solution!
Explore the notes and code in the 02-Image Classification (framework).ipynb notebook in the ready2019/notebooks folder to see an example of building and using a CNN.
In a a new notebook, create a CNN that predicts the class of an image based on the resized gear images you created in the previous challenge.
-
The architecture of your model should consist of a series of convolutional, pooling, drop, and fully-connected layers that you define.
-
The input layer of your model must match the size and shape of the training image arrays.
-
The output layer of your model must include an output for each class the model is designed to predict.
-
Randomly split the data into training and validation subsets with which to train and validate the model.
-
For each epoch in your training process, you should record the average loss for both the training and validation data; and when training is complete you should plot the training and loss values like this:
- Use the Python 3.5 kernel in Jupyterhub on your DSVM.
- Base your initial solution on the code in the sample notebook.
- To improve the model's performance, try adding more convolutional and pool layers, or using more training epochs.
- Try to avoid overfitting your model to the training data. One sign of this is that after your training and validation loss metrics converge, the training loss continues to drop but your validation loss stays the same or rises (as shown in the image above). The end result is a model that performs well when predicting the classes of images that it has been trained on, but which does not generalize well to new images.
- Techniques to help avoid overfitting include:
- Including drop layers to randomly remove some features from the model.
- Augmenting the data with re-oriented, skewed, or otherwise altered versions of training images.
Use the model to predict the class of at least five relevant images that are not included in the gear dataset. You can find these images by using Bing to search for appropriate terms, for example:
- https://www.bing.com/images/search?q=ski+helmet
- https://www.bing.com/images/search?q=climbing+axe
- https://www.bing.com/images/search?q=tent
- https://www.bing.com/images/search?q=carabiner
- https://www.bing.com/images/search?q=insulated+jacket
- You can download the new images to your DSVM by using the
curlcommand. - You should resize and pre-process the new images to match the images with which the model was trained.
-
Successfully train a convolutional neural network model
-
Plot the average training and validation loss observed when training your model
-
Achieve model accuracy of 0.90 (90%) or greater using your test data set.
-
Show predictions for the five images you identified in the Challenge section, like this:
(Note: Your model is not required to predict the correct class for all of the images, but it would be good if it does!)
When your coach has verified your team's solution, you can proceed to the next challenge.
- An Intuitive Explanation of Convolutional Neural Networks
- How Convolutional Neural Networks work (video)
- Demystifying AI (video)