DS222-Distributed-Logistic-Regression

This repository contains a comparative study implementing Logistic Regression in single machine and a distributed environment. The results are summarized in the Report.pdf and the implementation details are as discusstion below.

Running the codes

The above codes are developed in python 2.7 and requires the following libraries:

1. numpy==1.15.1
2. tensorflow==1.10.1
3. scipy==1.15

Before running any code of the codes run: python data_prep.py

Optionally for ease of experimentation the preprocessed data is already present in the repository in stored in sparse format.

In memory

This code can optionally take a input argument taking one of the three arguments: constant,decay and increase corresponding to the three strategies for varying learning rate.

Run the code: python LogitR_memo.py -lr "decay"

Distributed Tensorflow

For all the three settings namely bulk synchronous, stale-synchrnous and asynchronous 2 parameter server nodes and 2 worker nodes are used. To run any of the above codes go to the respective nodes and run the following commands:

1. pc1: python LogitR_stsynchro.py --job_name="ps" --task_index=0
2. pc2: python LogitR_stsynchro.py --job_name="ps" --task_index=1
3. pc3: python LogitR_stsynchro.py --job_name="worker" --task_index=0
4. pc4: python LogitR_stsynchro.py --job_name="worker" --task_index=1

Inside the code the ip address of the above nodes need to be specified in the following section:

parameter_servers = ["10.24.1.218:2222","10.24.1.214:2223"]

workers = ["10.24.1.215:2224", "10.24.1.217:2225"]