Embedding in pytorch

Question

Does Embedding make similar words closer to each other? And do I just need to give to it all the sentences? Or it is just a lookup table and I need to code the model?

Answer 1

nn.Embedding holds a Tensor of dimension (vocab_size, vector_size), i.e. of the size of the vocabulary x the dimension of each vector embedding, and a method that does the lookup.

When you create an embedding layer, the Tensor is initialised randomly. It is only when you train it when this similarity between similar words should appear. Unless you have overwritten the values of the embedding with a previously trained model, like GloVe or Word2Vec, but that's another story.

So, once you have the embedding layer defined, and the vocabulary defined and encoded (i.e. assign a unique number to each word in the vocabulary) you can use the instance of the nn.Embedding class to get the corresponding embedding.

For example:

import torch
from torch import nn
embedding = nn.Embedding(1000,128)
embedding(torch.LongTensor([3,4]))

will return the embedding vectors corresponding to the word 3 and 4 in your vocabulary. As no model has been trained, they will be random.

Answer 2

You could treat nn.Embedding as a lookup table where the key is the word index and the value is the corresponding word vector. However, before using it you should specify the size of the lookup table, and initialize the word vectors yourself. Following is a code example demonstrating this.

import torch.nn as nn 

# vocab_size is the number of words in your train, val and test set
# vector_size is the dimension of the word vectors you are using
embed = nn.Embedding(vocab_size, vector_size)

# intialize the word vectors, pretrained_weights is a 
# numpy array of size (vocab_size, vector_size) and 
# pretrained_weights[i] retrieves the word vector of
# i-th word in the vocabulary
embed.weight.data.copy_(torch.fromnumpy(pretrained_weights))

# Then turn the word index into actual word vector
vocab = {"some": 0, "words": 1}
word_indexes = [vocab[w] for w in ["some", "words"]] 
word_vectors = embed(word_indexes)

Answer 3

torch.nn.Embedding just creates a Lookup Table, to get the word embedding given a word index.

from collections import Counter
import torch.nn as nn

# Let's say you have 2 sentences(lowercased, punctuations removed) :
sentences = "i am new to PyTorch i am having fun"

words = sentences.split(' ')
    
vocab = Counter(words) # create a dictionary
vocab = sorted(vocab, key=vocab.get, reverse=True)
vocab_size = len(vocab)

# map words to unique indices
word2idx = {word: ind for ind, word in enumerate(vocab)} 

# word2idx = {'i': 0, 'am': 1, 'new': 2, 'to': 3, 'pytorch': 4, 'having': 5, 'fun': 6}

encoded_sentences = [word2idx[word] for word in words]

# encoded_sentences = [0, 1, 2, 3, 4, 0, 1, 5, 6]

# let's say you want embedding dimension to be 3
emb_dim = 3 

Now, embedding layer can be initialized as :

emb_layer = nn.Embedding(vocab_size, emb_dim)
word_vectors = emb_layer(torch.LongTensor(encoded_sentences))

This initializes embeddings from a standard Normal distribution(that is 0 mean and unit variance). Thus, these word vectors don't have any sense of 'relatedness'.

word_vectors is a torch tensor of size (9,3). (since there are 9 words in our data)

emb_layer has one trainable parameter called weight, which is, by default, set to be trained. You can check it by :

emb_layer.weight.requires_grad

which returns True. If you don't want to train your embeddings during model training(say, when you are using pre-trained embeddings), you can set them to False by :

emb_layer.weight.requires_grad = False

If your vocabulary size is 10,000 and you wish to initialize embeddings using pre-trained embeddings(of dim 300), say, Word2Vec, do it as :

emb_layer = nn.Embedding(10000, 300)
emb_layer.load_state_dict({'weight': torch.from_numpy(emb_mat)})

here, emb_mat is a Numpy matrix of size (10,000, 300) containing 300-dimensional Word2vec word vectors for each of the 10,000 words in your vocabulary.

Now, the embedding layer is loaded with Word2Vec word representations.

Answer 4

Agh! I think this part is still missing. Showcasing that when you set the embedding layer you automatically get the weights, that you may later alter with nn.Embedding.from_pretrained(weight)

import torch
import torch.nn as nn

embedding = nn.Embedding(10, 4)
print(type(embedding))
print(embedding)

t1 = embedding(torch.LongTensor([0,1,2,3,4,5,6,7,8,9])) # adding, 10 won't work
print(t1.shape)
print(t1)


t2 = embedding(torch.LongTensor([1,2,3]))
print(t2.shape)
print(t2)

#predefined weights
weight = torch.FloatTensor([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]])
print(weight.shape)
embedding = nn.Embedding.from_pretrained(weight)
# get embeddings for ind 0 and 1
embedding(torch.LongTensor([0, 1]))

Output:

<class 'torch.nn.modules.sparse.Embedding'>
Embedding(10, 4)
torch.Size([10, 4])
tensor([[-0.7007,  0.0169, -0.9943, -0.6584],
        [-0.7390, -0.6449,  0.1481, -1.4454],
        [-0.1407, -0.1081,  0.6704, -0.9218],
        [-0.2738, -0.2832,  0.7743,  0.5836],
        [ 0.4950, -1.4879,  0.4768,  0.4148],
        [ 0.0826, -0.7024,  1.2711,  0.7964],
        [-2.0595,  2.1670, -0.1599,  2.1746],
        [-2.5193,  0.6946, -0.0624, -0.1500],
        [ 0.5307, -0.7593, -1.7844,  0.1132],
        [-0.0371, -0.5854, -1.0221,  2.3451]], grad_fn=<EmbeddingBackward>)
torch.Size([3, 4])
tensor([[-0.7390, -0.6449,  0.1481, -1.4454],
        [-0.1407, -0.1081,  0.6704, -0.9218],
        [-0.2738, -0.2832,  0.7743,  0.5836]], grad_fn=<EmbeddingBackward>)
torch.Size([2, 3])

tensor([[0.1000, 0.2000, 0.3000],
        [0.4000, 0.5000, 0.6000]])

And the last part is that the Embedding layer weights can be learned with the gradient descent.