DL Frac-1 Assig-1
Budget: ₹650 – ₹700 INR
Assignment-1
Q1 (Programming/Experimentation):
Step-1: Download any CNN model pre-trained for the ImageNet classification task.
Step-2: Download the PASCAL VOC 2011 dataset from the following link:
http://host.robots.ox.ac.uk/pascal/VOC/voc2011/index.html
Step-3: Pick any one category of your choice from the dataset. From the training set, randomly
select 20%-50% (at least 20% and at most 50%) images that belong to this category (call it
category ‘A’), and randomly select 10% images each from all the remaining categories (call it
category “not A”). This will be the training dataset.
Step-4: Represent these images using the output of the last fully-connected layer of the above
CNN model.
Step-5: Use the entire validation set and represent these images in the same way as done in
step-4. Evaluate the classification accuracy of the kNN classifier on this set. Analyze the results
by varying different hyper-parameters such as the choice of distance/similarity function, value of
‘k’, feature normalization, etc. Provide the confusion matrix in all the cases.
Step-6: Submit all the codes along with a detailed report (PDF) containing all the results,
analyses and relevant details. Also include some qualitative results.
Q2 (Handwritten submission) :
Assume we perform a binary classification task in a many-to-one set-up. For this, we run vanilla
RNN for two time-steps, with x1 and x2 being the (vector) inputs at t=1 and t=2 respectively, and
y being a real-valued (scalar) output at the second time-step. To map this output into (-1,1), we
apply the tanh function. On this, we calculate the hinge loss function.
Derive the expression for the total loss (L) and calculate its first derivative w.r.t. to the parameter
matrix Why
.
Note:
(1) The assignment needs to be done individually.
(2) Submit your solutions to all the questions in a single zipped file.
(3) One may use any programming language/platform.
(4) For handwritten submissions, write your answers using pen and paper, and submit a
scanned PDF.
(5) There will be a penalty in case of plagiarism.
(6) For late submissions, there will be a penalty of 20% per day. The time recorded in
google-classroom will be considered.
Q1 (Programming/Experimentation):
Step-1: Download any CNN model pre-trained for the ImageNet classification task.
Step-2: Download the PASCAL VOC 2011 dataset from the following link:
http://host.robots.ox.ac.uk/pascal/VOC/voc2011/index.html
Step-3: Pick any one category of your choice from the dataset. From the training set, randomly
select 20%-50% (at least 20% and at most 50%) images that belong to this category (call it
category ‘A’), and randomly select 10% images each from all the remaining categories (call it
category “not A”). This will be the training dataset.
Step-4: Represent these images using the output of the last fully-connected layer of the above
CNN model.
Step-5: Use the entire validation set and represent these images in the same way as done in
step-4. Evaluate the classification accuracy of the kNN classifier on this set. Analyze the results
by varying different hyper-parameters such as the choice of distance/similarity function, value of
‘k’, feature normalization, etc. Provide the confusion matrix in all the cases.
Step-6: Submit all the codes along with a detailed report (PDF) containing all the results,
analyses and relevant details. Also include some qualitative results.
Q2 (Handwritten submission) :
Assume we perform a binary classification task in a many-to-one set-up. For this, we run vanilla
RNN for two time-steps, with x1 and x2 being the (vector) inputs at t=1 and t=2 respectively, and
y being a real-valued (scalar) output at the second time-step. To map this output into (-1,1), we
apply the tanh function. On this, we calculate the hinge loss function.
Derive the expression for the total loss (L) and calculate its first derivative w.r.t. to the parameter
matrix Why
.
Note:
(1) The assignment needs to be done individually.
(2) Submit your solutions to all the questions in a single zipped file.
(3) One may use any programming language/platform.
(4) For handwritten submissions, write your answers using pen and paper, and submit a
scanned PDF.
(5) There will be a penalty in case of plagiarism.
(6) For late submissions, there will be a penalty of 20% per day. The time recorded in
google-classroom will be considered.