Sentiment Analysis
Your task is to develop a deep learning model that takes each sentence from movie reviews and classifies it into one of the 5 sentiments: 0 – very negative, 1 – negative, 2 – neutral, 3 – positive, 4 – very positive.
Task 1
Copy and paste the hw2/ directory under the cs571 project to your own cs571 repo. You should see the following 3 files under hw2/res/ :
sst.trn.gold.tsv : training set.
sst.dev.gold.tsv : development set. sst.tst.unlabeled.tsv : evaluation set (unlabeled).
Download a pre-trained FastText model and put it under hw2/res/ :
fasttext-50-180614.bin : 50 dimensional word embeddings. fasttext-100-180614.bin : 100 dimensional word embeddings. fasttext-200-180614.bin : 200 dimensional word embeddings.
Update the following functions in src/hw2.py :
__init__ : initialize the FastText model as well as your neural network model. load : loads a pre-trained model saved by the save function.
save : saves a pre-trained model trained by the train function.
train : trains a model using training and development sets.
decode : predicts sentiment labels for the input data.
Your minimum required approach is the Convolutional Neural Network model introduced by Kim 2014.
Task 2
Write a report, res/hw2.pdf , describing your approach. Submission
Push all your changes to Github and make sure the following files are properly updated:
src/hw2.py
res/hw2.pdf
line ::=
sentiment ::= 0|1|2|3|4
document ::=
res/hw2-model (your final model; should be able to be loaded by the load function) Do not add the FastText model to git.
Submission: https://canvas.emory.edu/courses/54027/assignments/209131 Development
Important
Here are things you should notice before you start.
Each homework is structured as a python package. If you are using IDE, such as PyCharm, please open a new project and use hw2 as your project ROOT directory; otherwise, the package import path will be incorrect.
Develop everything based on the provided class. DO NOT delete existing methods, modify inheritance, or remove the given parameters. You can add new functions either outside or inside the class. Make sure everything can be run in the container successfully.
Please list all your dependencies in requirements.txt. A fixed library version is required (ref). GPU usage
You might want to use a GPU machine for this homework (using AWS is recommended). We have created an EC2 AMI with all the prerequisites for this homework. Please change your AWS region to Oregon(us-west-2) and launch an EC2 p2.xlarge instance with AMI id: ami- 04cd8c7e3716b3284 . It is okay to launch instance of your own. Make sure you instance can run nvidia-docker without sudo permission.
CUDA version for deep learning framework
Here let’s assume we want to use tenforflow==1.12 as our deep learning framework. Go to the hardware configuration page to check which GPU drivers are needed. It requires CUDA==9.0 and
cuDNN>=7.2 .
Next we need to change the source of Docker. List of CUDA and cuDNN dockers can be founded on Nvidia Docker Hub. We choose Ubuntu 16.04 for Operating System and find docker with
CUDA==9.0 and cuDNN>=7.2 in this section.
9.0-cudnn7-devel, 9.0-cudnn7-devel-ubuntu16.04 (9.0/devel/cudnn7/Dockerfile)
And replace the first line in the Dockerfile from
Run
Build a docker image:
– FROM nvidia/cuda:9.2-cudnn7-devel
+ FROM nvidia/cuda:9.0-cudnn7-devel
$ cd cs571/hw2
$ docker build -t hw2 .
Run the docker image:
Remember to replace /home/ubuntu/cs571/hw2 to the path where you locate res directory on your machine.
$ docker run -v /home/ubuntu/cs571/hw2:/mnt:rw -e “RESOURCE=/mnt/res/” hw2
Reviews
There are no reviews yet.