CS 5043: HW5: Complex Convolutional Neural Networks
Assignment notes:
- Deadline: Thursday, April 8th @11:59pm.
- Hand-in procedure: submit a pdf to Gradescope
- This work is to be done on your own. While general discussion
about Python, Keras and Tensorflow is encouraged, sharing
solution-specific code is inappropriate. Likewise, downloading
solution-specific code is not allowed.
- Do not submit MSWord documents.
Problem
Our objective with this homework is to explore non-serial deep network
architectures, constructed using the Keras Model API.
- The data set and classification problem is the same as in HW 4.
- Implement one more more non-serial deep networks. Possibilities include:
- An inception-style network.
- A network with more than one input image at the same
time (e.g., two input images separated by 10 time
steps).
- A network with K input images, each processed by the
same single-image recognizer and the outputs combined
using a recurrent neural network.
Hints / Notes
- For networks with multiple input images, you will need to alter
either the data loading process or the generator process. The
key is to make sure that all of the images that are presented
as part of one example all come from the same object. It
is fine to share this code with the rest of the class if you
have a nice solution.
What to Hand In
Hand in a PDF file that contains:
- Code for generating and training the network. Some useful UNIX
command line programs:
- enscript: translate code (e.g., py files) into postscript files
- ps2pdf: translate postscript files into pdf files
- pdfunite: merge several pdf files together
- The same learning curve and confusion matrix figures as
produced in HW 4
- A report of the mean multi-class AUC
- Your batch files used to run the experiments for your shallow and deep networks
- A statistical test of the mean validation set performance of your
best model from HW4 and your model for this assignment
Grades
- 50 pts: Model generation code. Is it correct? clean? documented?
- 50 pts: Model figures and performance. To achieve full credit
you must show a statistically significant improvement over your
best HW4 model.
- 10 pts: You solved the bonus problem of achieving average of 0.9 on the
validation data set your model across the five rotations, or
you beat the best of the TA/instructor solutions
andrewhfagg -- gmail.com
Last modified: Wed Mar 24 23:00:04 2021