Description
1
FaceScrub Dataset
This assignment will be done using the FaceScrub1
dataset. We will be using a tiny version of
this, with 6 celebrities and cropped images of 32-by-32. The target labels are the actor/actress
name, encoded as integers, as well as the gender, encoded as ’0’ and ’1’. You are provided with
two .npy files which have 936 rows of images and labels, and you should divide the dataset into
80/10/10% for training, validation and test, respectively.
1http://vintage.winklerbros.net/facescrub.html
v2
2
The name (ID) of the actors: ‘Lorraine Bracco’, ‘Gerard Butler’, ‘Peri Gilpin’, ‘Angie Harmon’, ‘Daniel Radcliffe’, and ‘Michael Vartan’ are encoded as ‘0’, ‘1’, ‘2’, ‘3’, ‘4’, and ‘5’,
respectively.
The gender of the actors: ‘Male’ and ‘Female’ are encoded as ‘0’ and ‘1’, respectively.
You should use the following code to load the dataset.
def data_segmentation(data_path, target_path, task):
# task = 0 >> select the name ID targets for face recognition task
# task = 1 >> select the gender ID targets for gender recognition task
data = np.load(data_path)/255
data = np.reshape(data, [-1, 32*32])
target = np.load(target_path)
np.random.seed(45689)
rnd_idx = np.arange(np.shape(data)[0])
np.random.shuffle(rnd_idx)
trBatch = int(0.8*len(rnd_idx))
validBatch = int(0.1*len(rnd_idx))
trainData, validData, testData = data[rnd_idx[1:trBatch],:], \
data[rnd_idx[trBatch+1:trBatch + validBatch],:],\
data[rnd_idx[trBatch + validBatch+1:-1],:]
trainTarget, validTarget, testTarget = target[rnd_idx[1:trBatch], task], \
target[rnd_idx[trBatch+1:trBatch + validBatch], task],\
target[rnd_idx[trBatch + validBatch + 1:-1], task]
return trainData, validData, testData, trainTarget, validTarget, testTarget
3
1 Convolutional Neural Networks [10 pt.]
Implement a convolutional neural network with one convolutional layer, one max-pooling
layer, and two layers of hidden units for classification of the FaceScrub dataset. Train the
model over the dataset for all of the provided training data. You should use the same Xavier
initialization of the weight matrices as before. A CNN is an efficient way to share weights in the
model and reduce the amount of parameters in a deep pipeline that provides an intuitive
interpretation of learning to recognize and compose image patches.
1. Convolutional layer: Write code that instantiates a 5-by-5 kernel of 32 filters for the
image, and perform a 2d convolution of the image with stride 1 in each direction. Define a bias
variable of shape [32] for the output of the kernel, and add the bias to the output of the 2d
convolution. What will be the output tensor dimensions if we had used 64 filters of 5-by-5 and a
stride of 1? What will it be if we used 32 filters of 7-by-7 and a stride of 2?
2. Max pooling layer: Write code that instantiates a max pooling layer of size 3-by-3 and
a stride of 2 in each direction.
3.Fully connected layer: Now flatten the output of the previous layer and pass through two
hidden layers with ReLU activations of size 384 and 192 with a dropout rate of 0.5. At the final
layer, output the probabilities of predicting each celebrity ID.
4. Learning: Use cross entropy as the loss, as you’ve done before, and train the model. Plot the
training and validation loss, as well as the training and validation accuracy over 50 epochs of
training. Run on 3 different hyperparameter settings (i.e. change the learning rate, weight
decay coefficient, dropout) and report your results.
5. Visualization: To get more insight into what a convolutional neural network achieves
with its architecture, you will visualize the function that the convolutional layer provides.
Visu-alize 8 of the 5-by-5 kernels trained in question 1, comment on what the network is
trained to recognize with these kernels and how further layers of convolution may improve
the per-formance.
