Instructions
User Manual:
Open the PDF directly: View PDF 
.
Page Count: 3
Project: Image Classification and Regression
MVA - CentraleSupelec
Vincent Lepetit
•The goal of this project is to learn how to implement simple image classification and regression
in Keras.
•You can refer to the introductory slides on Keras:
https://www.labri.fr/perso/vlepetit/teaching/deep_learning_mva/keras_intro.pdf
and of course the online documentation on Keras https://keras.io/.
1 Sending your Answers
1. Create an account on GitHub or equivalent if you don’t have one;
2. Use jupyter to do the assignment. You should also add plain text to explain your solution. Put
the jupyter file on your github account;
3. Send the link to your jupyter file on your github account by email to vincent.lepetit@
u-bordeaux.fr with MVA-MP1, MVA-MP2, or MVA-MP3 and your last name and first name
in the email subject. Example:
Subject: MVA-MP1 Doe, John
Body: link_to_your_jupyter_file.ipynb
No exception! No .py, no .zip, no .pdf
4. No acknowledgment for receiving the email, sorry.
2 Getting Started
Download the code provided at
https://www.labri.fr/perso/vlepetit/teaching/deep_learning_mva/mp1.py
Take some time to read it and understand it.
3 Simple Classification
You can generate a training set of images of simple geometric shapes (rectangle, disk, triangle)
centered in the images by calling the function:
[X_train, Y_train] = generate_dataset_classification(300, 20)
1
†Build and train a linear classifier in Keras to classify a image into one of the three possible
categories (i.e. rectangle, disk, triangle). Try using the stochastic gradient descent optimizer, then
the Adam optimizer.
Hints: You will have to use the following functions: Sequential,add,Dense (do not forget the
activation), compile,fit,np_utils.to_categorical. For the Adam optimizer, I used a batch size
of 32. You should use a small number of epochs when debugging to see if the optimization seems to
converge correctly.
You can check your classifier using the following code (for example):
X_test = generate_a_disk()
X_test = X_test.reshape(1, X_test.shape[0])
model.predict(X_test)
4 Visualization of the Solution
We would like to visualize the weights of the linear classifier. Check the output of the function
model.get_weights(): The first part corresponds to the matrix of the classifier. Its columns have
the same size as the input images, because Keras uses vector-matrix multiplications instead of matrix-
vector multiplications.
†Visualize the 3 columns as images.
Hint: Only two (short) lines of code are required to visualize one column.
5 A More Difficult Classification Problem
Now, the shapes are allowed to move within the images and change dimensions. You can generate
the new training set with:
[X_train, Y_train] = generate_dataset_classification(300, 20, True)
Retrain your linear classifier on this new training set. Add the metrics=[’accuracy’] parameter
when calling the compile function to get the classification error in addition to the loss value.
You can generate a test set by calling:
[X_test, Y_test] = generate_test_set_classification()
and evaluate your classifier on this test set by calling:
model.evaluate(X_test, Y_test)
†Train a convolutional (not-to-)deep network on this new dataset. What is the value of the loss
function on this test set when using your deep network?
Hints: You can limit yourself to 1 convolutional layer with 16 5×5 filters, 1 pooling layer, and
one fully connected layer, but you are free to use any other architecture. You are allowed to increase
the number of training samples if you want to.
6 A Regression Problem
The task now is to predict the image locations of the vertices of a triangle, given an image of this
triangle. You can generate a training set by calling:
[X_train, Y_train] = generate_dataset_regression(300, 20)
2
You can visualize a training sample (or a prediction) by calling the visualize_prediction
function:
visualize_prediction(X_train[0], Y_train[0])
†Build and train a regressor on this data. Evaluate your solution on the test set generated by
[X_test, Y_test] = generate_test_set_regression()
Hint: You may have to normalize somehow the vertices in Y_train and Y_test before training
and testing...
7 Image Denoising
Implement a hourglass network for denoising: Modifying the generate_a_* functions to generate
pairs of images, where one image has noise with random amplitude, and the second image has the
same content but without the noise. Train your network to predict a noise-free image given a noisy
image as input.
3
