Skip-Gram with Negative Sampling (PyTorch)

Mapping semantically similar words into closer locations in the embedding space.

Loss

Using Negative Sampling (drawing random noise words to form incorrect target pairs), the model tries to minimize the following Loss Function:

Repository Contents

This repository contains:

• SkipGram_NegativeSampling.py : Contains the complete source code for pre-processing and batching data, building the model, training the model, and visualizing the resulting word embeddings
• • util.py : Contains utility functions for text pre-processing
• data/text8.txt : Contains the training text
• SkipGram_NegativeSampling.ipynb : Step-by-step Colab Notebook for pre-processing and batching data, building the model, training the model, and visualizing the resulting word embeddings

List of Hyperparameters

• Number of Center Words in a Batch = 512
  • The actual Batch Size will vary, since there will be a varying number (in range [1, single_window_size]) of context words for each center word
• Threshold for Subsampling = 1e-5
• Single-side Window Size for Context = 5
  • So the whole window contains 5*2+1 = 11 words
• Embedding Dimension = 300
• Number of Negative (Noise) Samples Per Center Word = 5
• Learning Rate = 0.003
• Number of Training Epochs = 5

Sources

I referenced Udacity for building & debugging the final model :

• https://github.com/udacity/deep-learning-v2-pytorch