I created this notebook to better understand the inner workings of Bert. I followed a lot of tutorials to try to understand the architecture, but I was never able to really understand what was happening under the hood. For me it always helps to see the actual code instead of just simple abstract diagrams that a lot of times don’t match the actual implementation. If you’re like me than this tutorial will help!
I went as deep as you can go with Deep Learning — all the way to the tensor level. For me it helps to see the code and how the tensors move between layers. I feel like this level of abstraction is close enough to the core of the model to perfectly understand the inner workings.
The term forward pass is used in Neural Networks and it refers to the calculations involved from the input sequence all the way to output of the last layer. It’s basically the flow of data from input to output.
I will follow the code from an example input sequence all the way to the final output prediction.
What should I know for this notebook?
Some prior knowledge of Bert is needed. I won’t go into any details of how Bert works. For this there is plenty of information out there.
Since I am using the PyTorch implementation of Bert any knowledge on PyTorch is very useful.
Knowing a little bit about the transformers library helps too.
How deep are we going?
I think the best way to understand such a complex model as Bert is to see the actual layer components that are used. I will dig in the code until I see the actual PyTorch layers used torch.nn. In my opinion there is no need to go deeper than the torch.nn layers.
Tutorial Structure
Each section contains multiple subsections.
The order of each section matches the order of the model’s layers from input to output.
At the beginning of each section of code I created a diagram to illustrate the flow of tensors of that particular code.
I created the diagrams following the model’s implementation.
The major section Bert For Sequence Classification starts with the Class Call that shows how we normally create the Bert model for sequence classification and perform a forward pass. Class Components contains the components of BertForSequenceClassification implementation.
At the end of each major section, I assemble all components from that section and show the output and diagram.
At the end of the notebook, I have all the code parts and diagrams assembled.
Terminology
I will use regular deep learning terminology found in most Bert tutorials. I’m using some terms in a slightly different way:
Layer and layers: In this tutorial when I mention layer it can be an abstraction of a group of layers or just a single layer. When I reach torch.nn you know I refer to a single layer.
torch.nn: I’m referring to any PyTorch layer module. This is the deepest I will go in this tutorial.
How to use this notebook?
The purpose of this notebook is purely educational. This notebook is to be used to align known information on how Bert woks with the code implementation of Bert. I used the Bert implementation from Transformers. My contribution is on arranging the code implementation and creating associated diagrams.
Dataset
For simplicity I will only use two sentences as our data input: I love cats! and He hates pineapple pizza.. I’ll pretend to do binary sentiment classification on these two sentences.
Coding
Now let’s do some coding! We will go through each coding cell in the notebook and describe what it does, what’s the code, and when is relevant — show the output.
I made this format to be easy to follow if you decide to run each code cell in your own python notebook.
When I learn from a tutorial, I always try to replicate the results. I believe it’s easy to follow along if you have the code next to the explanations.
Installs
transformers library needs to be installed to use all the awesome code from Hugging Face. To get the latest version I will install it straight from GitHub.
# install the transformers library
!pip install -q git+https://github.com/huggingface/transformers.git
Installing build dependencies ... done
Getting requirements to build wheel ... done
Preparing wheel metadata ... done |████████████████████████████████| 2.9MB 6.7MB/s |████████████████████████████████| 890kB 48.9MB/s |████████████████████████████████| 1.1MB 49.0MB/s Building wheel for transformers (PEP 517) ... done
Building wheel for sacremoses (setup.py) ... done |████████████████████████████████| 71kB 5.2MB/s
Imports
Import all needed libraries for this notebook.
Declare parameters used for this notebook:
set_seed(123) – Always good to set a fixed seed for reproducibility.
n_labels – How many labels are we using in this dataset. This is used to decide size of classification head.
ACT2FN – Dictionary for special activation functions used in Bert. We’ll only need the gelu activation function.
BertLayerNorm – Shortcut for calling the PyTorch normalization layer torch.nn.LayerNorm.
import math
import torch
from transformers.activations import gelu
from transformers import (BertTokenizer, BertConfig, BertForSequenceClassification, BertPreTrainedModel, apply_chunking_to_forward, set_seed, )
from transformers.modeling_outputs import (BaseModelOutputWithPastAndCrossAttentions, BaseModelOutputWithPoolingAndCrossAttentions, SequenceClassifierOutput, ) # Set seed for reproducibility.
set_seed(123) # How many labels are we using in training.
# This is used to decide size of classification head.
n_labels = 2 # GELU Activation function.
ACT2FN = {"gelu": gelu} # Define BertLayerNorm.
BertLayerNorm = torch.nn.LayerNorm
Define Input
Let’s define some text data on which we will use Bert to classify as positive or negative.
We encoded our positive and negative sentiments into:
0 — for negative sentiments.
1 — for positive sentiments.
# Array of text we want to classify
input_texts = ['I love cats!', "He hates pineapple pizza."] # Senitmen labels
labels = [1, 0]
Bert Tokenizer
Creating the tokenizer is pretty standard when using the Transformers library.
Using our newly created tokenizer we’ll use it on our two sentence dataset and create the input_sequence that will be used as input for our Bert model.
Show Bert Tokenizer Diagram
# Create BertTokenizer.
tokenizer = BertTokenizer.from_pretrained('bert-base-cased') # Create input sequence using tokenizer.
input_sequences = tokenizer(text=input_texts, add_special_tokens=True, padding=True, truncation=True, return_tensors='pt') # Since input_sequence is a dictionary we can also add the labels to it
# want to make sure all values ar tensors.
input_sequences.update({'labels':torch.tensor(labels)}) # The tokenizer will return a dictionary of three: input_ids, attention_mask and token_type_ids.
# Let's do a pretty print.
print('PRETTY PRINT OF `input_sequences` UPDATED WITH `labels`:')
[print('%s : %sn'%(k,v)) for k,v in input_sequences.items()]; # Lets see how the text looks like after Bert Tokenizer.
# We see the special tokens added.
print('ORIGINAL TEXT:')
[print(example) for example in input_texts];
print('nTEXT AFTER USING `BertTokenizer`:')
[print(tokenizer.decode(example)) for example in input_sequences['input_ids'].numpy()];
Downloading: 100% |████████████████████████████████| 213k/213k [00:00<00:00, 278kB/s] PRETTY PRINT OF `input_sequences` UPDATED WITH `labels`:
input_ids : tensor([[ 101, 146, 1567, 11771, 106, 102, 0, 0, 0], [ 101, 1124, 18457, 10194, 11478, 7136, 13473, 119, 102]]) token_type_ids : tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0]]) attention_mask : tensor([[1, 1, 1, 1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1]]) labels : tensor([1, 0]) ORIGINAL TEXT:
I love cats!
He hates pineapple pizza. TEXT AFTER USING `BertTokenizer`:
[CLS] I love cats! [SEP] [PAD] [PAD] [PAD]
[CLS] He hates pineapple pizza. [SEP]
Bert Configuration
Predefined values specific to Bert architecture already defined for us by Hugging Face.
# Create the bert configuration.
bert_configuraiton = BertConfig.from_pretrained('bert-base-cased') # Let's see number of layers.
print('NUMBER OF LAYERS:', bert_configuraiton.num_hidden_layers) # We can also see the size of embeddings inside Bert.
print('EMBEDDING SIZE:', bert_configuraiton.hidden_size) # See which activation function used in hidden layers.
print('ACTIVATIONS:', bert_configuraiton.hidden_act)
Downloading: 100% |████████████████████████████████| 433/433 [00:00<00:00, 15.5kB/s] NUMBER OF LAYERS: 12
EMBEDDING SIZE: 768
ACTIVATIONS: gelu
Bert For Sequence Classification
I will go over the Bert for Sequence Classification model. This is a Bert language model with a classification layer on top.
If you plan on looking at other transformers models his tutorial will be very similar.
Class Call
Let’s start with doing a forward pass using the whole model call from Hugging Face Transformer.
# Let' start with the final model how we normally use.
model = BertForSequenceClassification.from_pretrained('bert-base-cased') # Perform a forward pass. We only care about the output and no gradients.
with torch.no_grad(): output = model.forward(**input_sequences) print() # Let's check how a forward pass output looks like.
print('FORWARD PASS OUTPUT:', output)
Downloading: 100% |████████████████████████████████| 436M/436M [00:07<00:00, 61.3MB/s] Some weights of the model checkpoint at bert-base-cased were not used when initializing BertForSequenceClassification: ...
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. FORWARD PASS OUTPUT: SequenceClassifierOutput(loss=tensor(0.7454), logits=tensor([[ 0.2661, -0.1774], [ 0.2223, -0.0847]]), hidden_states=None, attentions=None)
Class Components
Now let’s look at the code implementation and break down each part of the model and check the outputs.
Hugging Face was nice enough to mention a small summary: The bare Bert Model transformer outputting raw hidden-states without any specific head on top.
The forward pass uses the following layers:
self.embeddings = BertEmbeddings(config)
self.encoder = BertEncoder(config)
self.pooler = BertPooler(config)
Bert Embeddings
This is where we feed the input_sequences created under Bert Tokenizer and get our first embeddings.
Now perform a forward pass using previous output layer as input.
BERT Model Diagram
class BertModel(BertPreTrainedModel): """ The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of cross-attention is added between the self-attention layers, following the architecture described in `Attention is all you need &amp;amp;amp;lt;https://arxiv.org/abs/1706.03762&amp;amp;gt;`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. To behave as an decoder the model needs to be initialized with the :obj:`is_decoder` argument of the configuration set to :obj:`True`. To be used in a Seq2Seq model, the model needs to initialized with both :obj:`is_decoder` argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an input to the forward pass. """ def __init__(self, config, add_pooling_layer=True): super().__init__(config) self.config = config self.embeddings = BertEmbeddings(config) self.encoder = BertEncoder(config) self.pooler = BertPooler(config) if add_pooling_layer else None self.init_weights() def get_input_embeddings(self): return self.embeddings.word_embeddings def set_input_embeddings(self, value): self.embeddings.word_embeddings = value def _prune_heads(self, heads_to_prune): """ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel """ for layer, heads in heads_to_prune.items(): self.encoder.layer[layer].attention.prune_heads(heads) def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, encoder_hidden_states=None, encoder_attention_mask=None, past_key_values=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is configured as a decoder. encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`): Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`. use_cache (:obj:`bool`, `optional`): If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up decoding (see :obj:`past_key_values`). """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict if self.config.is_decoder: use_cache = use_cache if use_cache is not None else self.config.use_cache else: use_cache = False if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") elif input_ids is not None: input_shape = input_ids.size() batch_size, seq_length = input_shape elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] batch_size, seq_length = input_shape else: raise ValueError("You have to specify either input_ids or inputs_embeds") device = input_ids.device if input_ids is not None else inputs_embeds.device # past_key_values_length past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 if attention_mask is None: attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device) if token_type_ids is None: token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device) # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length] # ourselves in which case we just need to make it broadcastable to all heads. extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device) # If a 2D or 3D attention mask is provided for the cross-attention # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length] if self.config.is_decoder and encoder_hidden_states is not None: encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size() encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length) if encoder_attention_mask is None: encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device) encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask) else: encoder_extended_attention_mask = None # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head # attention_probs has shape bsz x n_heads x N x N # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length] head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers) embedding_output = self.embeddings( input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds, past_key_values_length=past_key_values_length, ) encoder_outputs = self.encoder( embedding_output, attention_mask=extended_attention_mask, head_mask=head_mask, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_extended_attention_mask, past_key_values=past_key_values, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) sequence_output = encoder_outputs[0] pooled_output = self.pooler(sequence_output) if self.pooler is not None else None if not return_dict: return (sequence_output, pooled_output) + encoder_outputs[1:] return BaseModelOutputWithPoolingAndCrossAttentions( last_hidden_state=sequence_output, pooler_output=pooled_output, past_key_values=encoder_outputs.past_key_values, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, cross_attentions=encoder_outputs.cross_attentions, ) # Create bert model.
bert_model = BertModel(bert_configuraiton) # Perform forward pass on entire model.
hidden_states = bert_model.forward(input_ids=input_sequences['input_ids'], attention_mask=input_sequences['attention_mask'], token_type_ids=input_sequences['token_type_ids'])
Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Attention Head Size: 64 Combined Attentions Head Size: 768 Created Tokens Positions IDs: tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8]]) Tokens IDs: torch.Size([2, 9]) Tokens Type IDs: torch.Size([2, 9]) Word Embeddings: torch.Size([2, 9, 768]) Position Embeddings: torch.Size([1, 9, 768]) Token Types Embeddings: torch.Size([2, 9, 768]) Sum Up All Embeddings: torch.Size([2, 9, 768]) Embeddings Layer Nromalization: torch.Size([2, 9, 768]) Embeddings Dropout Layer: torch.Size([2, 9, 768]) ----------------- BERT LAYER 1 ----------------- ... ----------------- BERT LAYER 12 ----------------- ... Hidden States: torch.Size([2, 9, 768]) First Token [CLS]: torch.Size([2, 768]) First Token [CLS] Linear Layer: torch.Size([2, 768]) First Token [CLS] Tanh Activation Function: torch.Size([2, 768])
