代写HW5留学生、代做Java/Python编程设计、代写artwork、代写CS, C/C++设计

日期：2018-11-13 10:17

HW5

November 11, 2018

0.1 Dataset

WikiArt is an amazing resource containing centuries of artwork. Since such datasets are wonderful

for deep learning, Kaggle has hosted a challenge to characterize the ’fingerprints’ of various artists.

The Kaggle dataset contains metadata and also a set of images that have been resized so that the

shorter dimension is 256 pixels. To make this homework reasonably fast even for those without

GPUs, we have further reduced the images to 64 x 64. CNNs and neural networks in general prefer

to have consistent sizes. To achieve this, we cut the center 256 pixels from the longer dimension

and then shrunk the images by a factor of 4. This isn’t a perfect solution since it did cut off a few

heads as you will see.

The selected images are for portraits and landscapes. No, we’re not talking about the orientation

but rather the content of the images. Thanks to help from Rashmi and Dave, we have a small

enough data set that should give reasonable results in a timely manner even on just a CPU.

The data were originally divided into a training and a test set. We have further divided the

training set into a train and validation set. In this homework you will be using the training set and

validation set to train and assess your deep learning models. At the final step, you will see how

well your final training worked on the test set. In each of these directories, there is a truth.txt file

that has the image name and whether it is a portrait or landscape scene.

0.2 Problem 1

Read in and display the first 5 portraits and the first 5 landscapes. Note, if you are using

the OpenCV tools, then the color may be distorted. The cvtColor() method using

cv2.COLOR_BGR2RGB may be useful. However, it is likely easier to use the generator and

plot_strip example from section.

Construct a baseline CNN classifier using Keras for the training set and assess the validation

set performance at each epoch. The goal is to correctly classify portraits from landscapes. Plot the

resulting performance on the training and validation set as a function of epoch using the criteria

over which you are optimizing. You should run at least 20 epochs for this problem.

From the pattern of training and validation curves, describe what is good/bad and what you

plan to do next to improve the result.

0.3 Problem 2

This step is where we want you to do most of your personal learning. Your goal is to improve

the network using a combination of architecture choices, parameter tuning, and experimenting

with different optimizers/dropout/regularization/etc. Treat each of these as separate optimization/exploration

steps for now. We would like to see 3 separate steps that cover different areas.

1

The format of the 3 steps should be as follows: * State the hypothesis/strategy for how you will

improve/explore a particular aspect. * Describe what types of tests you are running and why (i.e.

what range of parameters are you choosing and why) * Include the code and results * State your

interpretation of the results

We’re not looking for research in deep learning, but we want you to gain some hands-on experience

working with Keras and figuring out what works. A good example may be comparing

strategies to overcome overfitting, or comparing a few different CNN architectures in terms of

performance and speed, or comparing data augmentation types and results.

0.4 Problem 3

Assess your best model on the test data. Plot the corresponding AUC curve from the results (since

we’ve provided the truth). This was not directly covered in section, but will require a prediction

using images in the same format as the training. We suggest referring to the Keras API else use a

Google to search to find how to make predictions.

Display the 5 images [worst] misclassified images for each class. Worst is in brackets since

certain architectures may only make a binary decision rather than a score. In that case, plot 5 of

each.