Transfer learning with unsupervised domain adaptation on cancer cell lines

1. Preparation

1.1 Clone the repository

Please make sure that you have more than 30 GB storage. In any terminal, execute the following.

git clone https://github.com/AntonioShen/DANN.git

1.2 Download the GDSC and CCLE dataset

Google Drive:

https://drive.google.com/drive/folders/1yz9UeFCF6MK0AeNCmnb-6rOVQILYaptL?usp=sharing

On the above webpage, download all 3 files in the gdsc folder and move them into the following folder.

$yourworkspace$/DANN/data/cleaned/gdsc/

Repeat the previous step for the other 3 files in the ccle folder, and move them into the following folder.

$yourworkspace$/DANN/data/cleaned/ccle/

1.3 Install dependencies

Under the project root, execute the following.

pip install -r requirements.txt

1.4 Run dataset helper

This generates essential median csv files for future processing.

python dataset_helper.py

1.3 Run csv parser

This processes csv files and transfer into pt files for loading as tensors.

python parse_to_data_tensors.py -o <OPTION> -n <NAME>

-h for help, if you are a first-time user, we suggest running the following.

python parse_to_data_tensors.py -o exclude -n all

This process will take some time to complete.

2. Experiments

2.1 Run pipeline

The main.py file includes implicit configuration for training and testing.

python main.py

2.2 Access results

If you did not configure your own W&B visualization before running main.py, you can find results and metrics inside the below folder.

$yourworkspace$/DANN/results/plots/

The embedded visual logger is implemented in logger.py, a typical usage is shown below.

from logger import Logger

# An instance of Logger class which requires a list of names
my_logger = Logger(['epoch', 'metric1', 'metric2'])

# Inside the training loop, record the current results
for epoch in range(1, epochs + 1):

    # Do some training
    y_hat = model(x)
    
    # Compute metrics
    m1 = loss_func1(y_hat, y)
    m2 = loss_func2(y_hat, y)
    
    # Call log() to record data or metrics
    # It requires all fields in names be filled in.
    my_logger.log({'epoch': epoch, 
                   'metric1': m1, 
                   'metric2': m2})

# Outside the loop, call save_plot() to save the trend plot
# The x-axis refers to values under the first name, here it is epoch
PATH = './result_plot.jpg'
my_logger.save_plot(PATH)

xingbpshen/DANN