HW 6

Fine-Tuning for Animal Recognition

For this homework, we'll apply the Fine-Tuning training approach to train Caffe to classify dog and cat images. The images are here. This dataset originally came from a CAPTCHA system called ASIRRA Links to an external site. (click link and scroll down to installation instructions) developed a Microsoft. Once upon a time, people could classify animal images much more accurately than computers, and ASIRRA presented dog/cat images to the user to verify that they were actually human. Now deep learning achieves human-level performance on this task, and ASIRRA was shut down.

Process

Save the train/test data somewhere in your EC2 instance.

Follow the steps in the Caffe fine-tuning tutorial, and apply to this new data. Specifically:

• Copy the finetune ipython notebook from the examples directory to make your own HW6 notebook.
• Deep copy the models/finetune_flickr_style directory to models/dogcat.
• Modify the prototxt files in models/dogcat to point to your new data
• Run finetuning for 2000 iterations on the training data using the original finetune solver params. What accuracy do you get?
• The fine-tuning files by default boost the learning rate only in the last FC layer. Modify the prototxts to train the last *two* FC layers with learning rate boost. Run for 2000 iterations and report the accuracy you get.

Submit

Submit your notebook (with all the cells evaluated) here.

Also answer the questions above through this quiz link.

Shutdown

Don't forget to shut down your EC2 instance when you're done.