Please always first refer to User Guide of Ludwig from https://uber.github.io/ludwig/user_guide/ when you have any questions.
- Create a new folder called:
ludwig_exercisein your Google Drive. - Download zip file:
clickbait.csvandconfig.yamlin to theludwig_exercisefolder created in the previous step. - Make sure you have at least 1GB free space in your Google Drive
- Go to your Google Drive
- Access Google Collab at: https://colab.research.google.com. Click "New Python3 Notebook", the browser will show the interface of a iPython Notebook
- Add GPU: Click menu "Edit" -> "Notebook Settings". In the dialog box, select "GPU" from drop box "Hardware Acceleration"
import tensorflow as tf
device_name = tf.test.gpu_device_name()
if device_name != '/device:GPU:0':
raise SystemError('GPU device not found')
print('Found GPU at: {}'.format(device_name))
!apt-get install libgmp-dev libmpfr-dev libmpc-dev
!pip install ludwig --no-deps tensorflow
import pandas as pd
import numpy as np
from google.colab import drive
drive.mount('/gdrive')
Click the authorization link provided, follow the instruction to get the authorization code. Copy the authorization code back to the Google Collab notebook and press enter. Now it should show: Mounted at /gdrive
%cd /gdrive/My Drive/ludwig_exercise
%ls
Making sure you can see files: clickbait.csv, config.yaml show up
df = pd.read_csv('clickbait.csv', index_col=0)
df.head()
print("total data samples:{}".format(df.shape[0]))
print("label distribution:{}".format(np.bincount(df['class'])))
Now we want to command ludwig to do training experiment on the clickbait dataset clickbait.csv with a deep learning architecture described in config.yaml
!ludwig experiment \
--data_csv 'clickbait.csv' \
--model_definition_file 'config.yaml'
Ludwig will automatically split the whole dataset into 3 parts: train, validation and test sets, in ratio: 75%:15%:10%. You can manually split the dataset by 1) create a new column called split in the dataset, 2) assign the split column to value 0 for training samples, 1 for validataion samples and 2 for test samples.
If at any point in the the Google Collab shows error, you can reset the running time: menu "Runtime" -> "Reset all runtimes…", then run every cells again
quora dataset: I have also prepared two datasets: quora_train.csv and quora_test.csv from this link: https://drive.google.com/drive/folders/1VyX0Tf5UZ3SzWEE8wYiMztRbZWavttaq?usp=sharing. Details of this task can be found at: https://www.kaggle.com/c/quora-question-pairs. Please use the train dataset to train a model to predict question pairs similiarity, and report on test set.
Hint 1: Siamese architecture at https://uber.github.io/ludwig/examples/#one-shot-learning-with-siamese-networks is one example of neural network model that can be used. In this example they want to predict similarity between a pair of images, we need to change the architecture accordingly to accommodate our task.
Hint 2: try different encoders
Hint 3: try pre-trained word-embedding such as glove or word2vec with pretrained_embeddings parameter on the configuration file (prefer to user manual for details)
Hint 4: try different training parameters (batch-size, learning rate, etc.)
Check out examples shown on the ludwig official website: https://uber.github.io/ludwig/examples/