Resource retrieval

Get the pre-trained net:

In[1]:=

Out[1]=

NetModel parameters

This model consists of a family of individual nets, each identified by a specific parameter combination. Inspect the available parameters:

In[2]:=

Out[2]=

Pick a non-default net by specifying the parameters:

In[3]:=

Out[3]=

Pick a non-default uninitialized net:

In[4]:=

Out[4]=

Basic usage

Given a piece of text, the BERT net produces a sequence of feature vectors of size 768, which correspond to the sequence of input words or subwords:

In[5]:=

Obtain dimensions of the embeddings:

In[6]:=

Out[6]=

Visualize the embeddings:

In[7]:=

Out[7]=

Transformer architecture

Each input text segment is first tokenized into words or subwords using a word-piece tokenizer and additional text normalization. Integer codes called token indices are generated from these tokens, together with additional segment indices:

In[8]:=

Out[9]=

For each input subword token, the encoder yields a pair of indices that correspond to the token index in the vocabulary, and the index of the sentence within the list of input sentences:

In[10]:=

Out[10]=

The list of tokens always starts with special token index 102, which corresponds to the classification index. Also the special token index 103 is used as a separator between the different text segments. Each subword token is also assigned a positional index:

In[11]:=

Out[11]=

A lookup is done to map these indices to numeric vectors of size 768:

In[12]:=

Out[13]=

For each subword token, these three embeddings are combined by summing elements with ThreadingLayer:

In[14]:=

Out[14]=

The transformer architecture then processes the vectors using 12 structurally identical self-attention blocks stacked in a chain:

In[15]:=

Out[15]=

The key part of these blocks is the attention module comprising of 12 parallel self-attention transformations, also called “attention heads”:

In[16]:=

Out[16]=

Each head uses an AttentionLayer at its core:

In[17]:=

Out[17]=

BERT uses self-attention, where the embedding of a given subword depends on the full input text. The following figure compares self-attention (lower left) to other types of connectivity patterns that are popular in deep learning:

Sentence analogies

Define a sentence embedding that takes the last feature vector from BERT subword embeddings (as an arbitrary choice):

In[18]:=

Out[18]=

Define a list of sentence in two broad categories (food and music):

In[19]:=

Precompute the embeddings for a list of sentences:

In[20]:=

Visualize the similarity between the sentences using the net as a feature extractor:

In[21]:=

Out[21]=

Train a classifier model with the subword embeddings

Get a text-processing dataset:

In[22]:=

View a random sample of the dataset:

In[23]:=

Out[23]=

Precompute the BERT vectors for the training and the validation datasets (if available, GPU is highly recommended):

In[24]:=

Define a network to classify the sequences of subword embeddings, using a max-pooling strategy:

In[25]:=

Out[25]=

Train the network on the precomputed BERT vectors:

In[26]:=

Out[26]=

Check the classification error rate on the validation data:

In[27]:=

Out[27]=

Let’s compare the results with the performance of a classifier trained on context-independent word embeddings. Precompute the GloVe vectors for the training and the validation dataset:

In[28]:=

Train the classifier on the precomputed GloVe vectors:

In[29]:=

Out[29]=

Compare the results obtained with GPT and with GloVe:

In[30]:=

Out[30]=

Net information

Inspect the number of parameters of all arrays in the net:

In[31]:=

Out[31]=

Obtain the total number of parameters:

In[32]:=

Out[32]=

Obtain the layer type counts:

In[33]:=

Out[33]=

Display the summary graphic:

In[34]:=

Out[34]=

Export to MXNet

Export the net into a format that can be opened in MXNet:

In[35]:=

Out[35]=

Export also creates a net.params file containing parameters:

In[36]:=

Out[36]=

Get the size of the parameter file:

In[37]:=

Out[37]=