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.
If this in-depth educational content is useful for you, subscribe to our AI research mailing list to be alerted when we release new material.
I will use the implementation of Bert from one of the best NLP library out there — HuggingFace Transformers. More specifically, I will show the inner working of Bert For Sequence Classification.
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.nnyou 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/sImports
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 thegeluactivation function.BertLayerNorm– Shortcut for calling the PyTorch normalization layertorch.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.LayerNormDefine 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 : %s\n'%(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: geluBert 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.
Start with the BertForSequenceClassification found in transformers/src/transformers/models/bert/modeling_bert.py#L1449.
The forward pass uses the following layers:
- BertModel layer:
self.bert = BertModel(config)
- torch.nn.Dropout layer for dropout:
self.dropout = nn.Dropout(config.hidden_dropout_prob)
- torch.nn.Linear layer used for classification:
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
BertModel
This is the core Bert model that can be found at: transformers/src/transformers/models/bert/modeling_bert.py#L815.
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:
- BertEmbeddings layer:
self.embeddings = BertEmbeddings(config)
- BertEncoder layer:
self.encoder = BertEncoder(config)
- BertPooler layer:
self.pooler = BertPooler(config)
Bert Embeddings
This is where we feed the input_sequences created under Bert Tokenizer and get our first embeddings.
Implementation can be found at: transformers/src/transformers/models/bert/modeling_bert.py#L165.
This layer contains actual PyTorch layers. I won’t go into farther details since this is how far we need to go.
The forward pass uses following layers:
- torch.nn.Embedding layer for word embeddings:
self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
- torch.nn.Embedding layer for position embeddings:
self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
- torch.nn.Embedding for token type embeddings:
self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
- torch.nn.LayerNorm layer for normalization:
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
- torch.nn.Dropout layer for dropout:
self.dropout = nn.Dropout(config.hidden_dropout_prob)
class BertEmbeddings(torch.nn.Module):
"""Construct the embeddings from word, position and token_type embeddings."""
def __init__(self, config):
super().__init__()
self.word_embeddings = torch.nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
self.position_embeddings = torch.nn.Embedding(config.max_position_embeddings, config.hidden_size)
self.token_type_embeddings = torch.nn.Embedding(config.type_vocab_size, config.hidden_size)
# self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
# any TensorFlow checkpoint file
self.LayerNorm = torch.nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
def forward(
self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
):
if input_ids is not None:
input_shape = input_ids.size()
else:
input_shape = inputs_embeds.size()[:-1]
seq_length = input_shape[1]
if position_ids is None:
position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
# ADDED
print('Created Tokens Positions IDs:\n', position_ids)
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
if inputs_embeds is None:
inputs_embeds = self.word_embeddings(input_ids)
token_type_embeddings = self.token_type_embeddings(token_type_ids)
# ADDED
print('\nTokens IDs:\n', input_ids.shape)
print('\nTokens Type IDs:\n', token_type_ids.shape)
print('\nWord Embeddings:\n', inputs_embeds.shape)
embeddings = inputs_embeds + token_type_embeddings
if self.position_embedding_type == "absolute":
position_embeddings = self.position_embeddings(position_ids)
# ADDED
print('\nPosition Embeddings:\n', position_embeddings.shape)
embeddings += position_embeddings
# ADDED
print('\nToken Types Embeddings:\n', token_type_embeddings.shape)
print('\nSum Up All Embeddings:\n', embeddings.shape)
embeddings = self.LayerNorm(embeddings)
# ADDED
print('\nEmbeddings Layer Nromalization:\n', embeddings.shape)
embeddings = self.dropout(embeddings)
# ADDED
print('\nEmbeddings Dropout Layer:\n', embeddings.shape)
return embeddings
# Create Bert embedding layer.
bert_embeddings_block = BertEmbeddings(bert_configuraiton)
# Perform a forward pass.
embedding_output = bert_embeddings_block.forward(input_ids=input_sequences['input_ids'], token_type_ids=input_sequences['token_type_ids'])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 Encoder
This layer contains the core of the bert model where the self-attention happens.
The implementation can be found at: transformers/src/transformers/models/bert/modeling_bert.py#L512.
The forward pass uses:
- 12 of the BertLayer layers ( in this setup
config.num_hidden_layers=12):
self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
BERT LAYER
This layer contains basic components of the self-attention implementation.
Implementation can be found at transformers/src/transformers/models/bert/modeling_bert.py#L429.
The forward pass uses:
- BertAttention layer:
self.attention = BertAttention(config)
- BertIntermediate layer:
self.intermediate = BertIntermediate(config)
- BertOutput layer:
self.output = BertOutput(config)
Bert Attention
This layer contains basic components of the self-attention implementation.
Implementation can be found at transformers/src/transformers/models/bert/modeling_bert.py#L351.
The forward pass uses:
- BertSelfAttention layer:
self.self = BertSelfAttention(config)
- BertSelfOutput layer:
self.output = BertSelfOutput(config)
BertSelfAttention
This layer contains the torch.nn basic components of the self-attention implementation.
Implementation can be found at transformers/src/transformers/models/bert/modeling_bert.py#L212.
The forward pass uses:
- torch.nn.Linear used for the Query layer:
self.query = nn.Linear(config.hidden_size, self.all_head_size)
- torch.nn.Linear used for the Key layer:
self.key = nn.Linear(config.hidden_size, self.all_head_size)
- torch.nn.Linear used for the Value layer:
self.value = nn.Linear(config.hidden_size, self.all_head_size)
- torch.nn.Dropout layer for dropout:
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
class BertSelfAttention(torch.nn.Module):
def __init__(self, config):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
raise ValueError(
"The hidden size (%d) is not a multiple of the number of attention "
"heads (%d)" % (config.hidden_size, config.num_attention_heads)
)
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
# ADDED
print('Attention Head Size:\n', self.attention_head_size)
print('\nCombined Attentions Head Size:\n', self.all_head_size)
self.query = torch.nn.Linear(config.hidden_size, self.all_head_size)
self.key = torch.nn.Linear(config.hidden_size, self.all_head_size)
self.value = torch.nn.Linear(config.hidden_size, self.all_head_size)
self.dropout = torch.nn.Dropout(config.attention_probs_dropout_prob)
self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
self.max_position_embeddings = config.max_position_embeddings
self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
self.is_decoder = config.is_decoder
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
):
# ADDED
print('\nHidden States:\n', hidden_states.shape)
mixed_query_layer = self.query(hidden_states)
# If this is instantiated as a cross-attention module, the keys
# and values come from an encoder; the attention mask needs to be
# such that the encoder's padding tokens are not attended to.
is_cross_attention = encoder_hidden_states is not None
if is_cross_attention and past_key_value is not None:
# ADDED
print('\nQuery Linear Layer:\n', mixed_query_layer.shape)
print('\nKey Linear Layer:\n', past_key_value[0].shape)
print('\nValue Linear Layer:\n', past_key_value[1].shape)
# reuse k,v, cross_attentions
key_layer = past_key_value[0]
value_layer = past_key_value[1]
attention_mask = encoder_attention_mask
elif is_cross_attention:
# ADDED
print('\nQuery Linear Layer:\n', mixed_query_layer.shape)
print('\nKey Linear Layer:\n', self.key(encoder_hidden_states).shape)
print('\nValue Linear Layer:\n', self.value(encoder_hidden_states).shape)
key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
attention_mask = encoder_attention_mask
elif past_key_value is not None:
# ADDED
print('\nQuery Linear Layer:\n', mixed_query_layer.shape)
print('\nKey Linear Layer:\n', self.key(hidden_states).shape)
print('\nValue Linear Layer:\n', self.value(hidden_states).shape)
key_layer = self.transpose_for_scores(self.key(hidden_states))
value_layer = self.transpose_for_scores(self.value(hidden_states))
key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
else:
# ADDED
print('\nQuery Linear Layer:\n', mixed_query_layer.shape)
print('\nKey Linear Layer:\n', self.key(hidden_states).shape)
print('\nValue Linear Layer:\n', self.value(hidden_states).shape)
key_layer = self.transpose_for_scores(self.key(hidden_states))
value_layer = self.transpose_for_scores(self.value(hidden_states))
query_layer = self.transpose_for_scores(mixed_query_layer)
# ADDED
print('\nQuery:\n', query_layer.shape)
print('\nKey:\n', key_layer.shape)
print('\nValue:\n', value_layer.shape)
if self.is_decoder:
# if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
# Further calls to cross_attention layer can then reuse all cross-attention
# key/value_states (first "if" case)
# if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
# all previous decoder key/value_states. Further calls to uni-directional self-attention
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
# if encoder bi-directional self-attention `past_key_value` is always `None`
past_key_value = (key_layer, value_layer)
# ADDED
print('\nKey Transposed:\n', key_layer.transpose(-1, -2).shape)
# Take the dot product between "query" and "key" to get the raw attention scores.
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
# ADDED
print('\nAttention Scores:\n', attention_scores.shape)
if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
seq_length = hidden_states.size()[1]
position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
distance = position_ids_l - position_ids_r
positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
positional_embedding = positional_embedding.to(dtype=query_layer.dtype) # fp16 compatibility
if self.position_embedding_type == "relative_key":
relative_position_scores = torch.einsum("bhld,lrd-&amp;amp;gt;bhlr", query_layer, positional_embedding)
attention_scores = attention_scores + relative_position_scores
elif self.position_embedding_type == "relative_key_query":
relative_position_scores_query = torch.einsum("bhld,lrd-&amp;amp;gt;bhlr", query_layer, positional_embedding)
relative_position_scores_key = torch.einsum("bhrd,lrd-&amp;amp;gt;bhlr", key_layer, positional_embedding)
attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
attention_scores = attention_scores / math.sqrt(self.attention_head_size)
# ADDED
print('\nAttention Scores Divided by Scalar:\n', attention_scores.shape)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in BertModel forward() function)
attention_scores = attention_scores + attention_mask
# Normalize the attention scores to probabilities.
attention_probs = torch.nn.Softmax(dim=-1)(attention_scores)
# ADDED
print('\nAttention Probabilities Softmax Layer:\n', attention_probs.shape)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs = self.dropout(attention_probs)
# ADDED
print('\nAttention Probabilities Dropout Layer:\n', attention_probs.shape)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = torch.matmul(attention_probs, value_layer)
# ADDED
print('\nContext:\n', context_layer.shape)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
# ADDED
print('\nContext Permute:\n', context_layer.shape)
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
# ADDED
print('\nContext Reshaped:\n', context_layer.shape)
outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
if self.is_decoder:
outputs = outputs + (past_key_value,)
return outputs
# Create bert self attention layer.
bert_selfattention_block = BertSelfAttention(bert_configuraiton)
# Perform a forward pass.
context_embedding = bert_selfattention_block.forward(hidden_states=embedding_output)Attention Head Size:
64
Combined Attentions Head Size:
768
Hidden States:
torch.Size([2, 9, 768])
Query Linear Layer:
torch.Size([2, 9, 768])
Key Linear Layer:
torch.Size([2, 9, 768])
Value Linear Layer:
torch.Size([2, 9, 768])
Query:
torch.Size([2, 12, 9, 64])
Key:
torch.Size([2, 12, 9, 64])
Value:
torch.Size([2, 12, 9, 64])
Key Transposed:
torch.Size([2, 12, 64, 9])
Attention Scores:
torch.Size([2, 12, 9, 9])
Attention Scores Divided by Scalar:
torch.Size([2, 12, 9, 9])
Attention Probabilities Softmax Layer:
torch.Size([2, 12, 9, 9])
Attention Probabilities Dropout Layer:
torch.Size([2, 12, 9, 9])
Context:
torch.Size([2, 12, 9, 64])
Context Permute:
torch.Size([2, 9, 12, 64])
Context Reshaped:
torch.Size([2, 9, 768])BertSelfOutput
This layer contains the torch.nn basic components of the self-attention implementation.
Implementation can be found at transformers/src/transformers/models/bert/modeling_bert.py#L337.
The forward pass uses:
- torch.nn.Linear layer:
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
- torch.nn.LayerNorm layer for normalization:
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
- torch.nn.Dropout layer for dropout:
self.dropout = nn.Dropout(config.hidden_dropout_prob)
class BertSelfOutput(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.dense = torch.nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
print('Hidden States:\n', hidden_states.shape)
hidden_states = self.dense(hidden_states)
print('\nHidden States Linear Layer:\n', hidden_states.shape)
hidden_states = self.dropout(hidden_states)
print('\nHidden States Dropout Layer:\n', hidden_states.shape)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
print('\nHidden States Normalization Layer:\n', hidden_states.shape)
return hidden_states
# Create Bert self output layer.
bert_selfoutput_block = BertSelfOutput(bert_configuraiton)
# Perform a forward pass - context_embedding[0] because we have tuple.
attention_output = bert_selfoutput_block.forward(hidden_states=context_embedding[0], input_tensor=embedding_output)Hidden States:
torch.Size([2, 9, 768])
Hidden States Linear Layer:
torch.Size([2, 9, 768])
Hidden States Dropout Layer:
torch.Size([2, 9, 768])
Hidden States Normalization Layer:
torch.Size([2, 9, 768])Assemble BertAttention
Put together BertSelfAttention layer and BertSelfOutput layer to create the BertAttention layer.
Now perform a forward pass using previous output layer as input.
class BertAttention(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.self = BertSelfAttention(config)
self.output = BertSelfOutput(config)
self.pruned_heads = set()
def prune_heads(self, heads):
if len(heads) == 0:
return
heads, index = find_pruneable_heads_and_indices(
heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
)
# Prune linear layers
self.self.query = prune_linear_layer(self.self.query, index)
self.self.key = prune_linear_layer(self.self.key, index)
self.self.value = prune_linear_layer(self.self.value, index)
self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
# Update hyper params and store pruned heads
self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
self.pruned_heads = self.pruned_heads.union(heads)
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
):
self_outputs = self.self(
hidden_states,
attention_mask,
head_mask,
encoder_hidden_states,
encoder_attention_mask,
past_key_value,
output_attentions,
)
attention_output = self.output(self_outputs[0], hidden_states)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
return outputs
# Create attention assembled layer.
bert_attention_block = BertAttention(bert_configuraiton)
# Perform a forward pass to wholte Bert Attention layer.
attention_output = bert_attention_block(hidden_states=embedding_output)Attention Head Size:
64
Combined Attentions Head Size:
768
Hidden States:
torch.Size([2, 9, 768])
Query Linear Layer:
torch.Size([2, 9, 768])
Key Linear Layer:
torch.Size([2, 9, 768])
Value Linear Layer:
torch.Size([2, 9, 768])
Query:
torch.Size([2, 12, 9, 64])
Key:
torch.Size([2, 12, 9, 64])
Value:
torch.Size([2, 12, 9, 64])
Key Transposed:
torch.Size([2, 12, 64, 9])
Attention Scores:
torch.Size([2, 12, 9, 9])
Attention Scores Divided by Scalar:
torch.Size([2, 12, 9, 9])
Attention Probabilities Softmax Layer:
torch.Size([2, 12, 9, 9])
Attention Probabilities Dropout Layer:
torch.Size([2, 12, 9, 9])
Context:
torch.Size([2, 12, 9, 64])
Context Permute:
torch.Size([2, 9, 12, 64])
Context Reshaped:
torch.Size([2, 9, 768])
Hidden States:
torch.Size([2, 9, 768])
Hidden States Linear Layer:
torch.Size([2, 9, 768])
Hidden States Dropout Layer:
torch.Size([2, 9, 768])
Hidden States Normalization Layer:
torch.Size([2, 9, 768])BertIntermediate
This layer contains the torch.nn basic components of the Bert model implementation.
Implementation can be found at transformers/src/transformers/models/bert/modeling_bert.py#L400.
The forward pass uses:
- torch.nn.Linear layer:
self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
class BertIntermediate(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.dense = torch.nn.Linear(config.hidden_size, config.intermediate_size)
if isinstance(config.hidden_act, str):
self.intermediate_act_fn = ACT2FN[config.hidden_act]
else:
self.intermediate_act_fn = config.hidden_act
def forward(self, hidden_states):
print('\nHidden States:\n', hidden_states.shape)
hidden_states = self.dense(hidden_states)
print('\nHidden States Linear Layer:\n', hidden_states.shape)
hidden_states = self.intermediate_act_fn(hidden_states)
print('\nHidden States Gelu Activation Function:\n', hidden_states.shape)
return hidden_states
# Create bert intermediate layer.
bert_intermediate_block = BertIntermediate(bert_configuraiton)
# Perform a forward pass - attention_output[0] because we have tuple.
intermediate_output = bert_intermediate_block.forward(hidden_states=attention_output[0])Hidden States:
torch.Size([2, 9, 768])
Hidden States Linear Layer:
torch.Size([2, 9, 3072])
Hidden States Gelu Activation Function:
torch.Size([2, 9, 3072])BertOutput
This layer contains the torch.nn basic components of the Bert model implementation.
Implementation can be found at transformers/src/transformers/models/bert/modeling_bert.py#L415.
The forward pass uses:
- torch.nn.Linear layer:
self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
- torch.nn.LayerNorm layer for normalization:
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
- torch.nn.Dropout layer for dropout:
self.dropout = nn.Dropout(config.hidden_dropout_prob)
class BertOutput(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.dense = torch.nn.Linear(config.intermediate_size, config.hidden_size)
self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
print('\nHidden States:\n', hidden_states.shape)
hidden_states = self.dense(hidden_states)
print('\nHidden States Linear Layer:\n', hidden_states.shape)
hidden_states = self.dropout(hidden_states)
print('\nHidden States Dropout Layer:\n', hidden_states.shape)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
print('\nHidden States Layer Normalization:\n', hidden_states.shape)
return hidden_states
# Create bert output layer.
bert_output_block = BertOutput(bert_configuraiton)
# Perform forward pass - attention_output[0] dealing with tuple.
layer_output = bert_output_block.forward(hidden_states=intermediate_output, input_tensor=attention_output[0])Hidden States:
torch.Size([2, 9, 3072])
Hidden States Linear Layer:
torch.Size([2, 9, 768])
Hidden States Dropout Layer:
torch.Size([2, 9, 768])
Hidden States Layer Normalization:
torch.Size([2, 9, 768])Assemble BertLayer
Put together BertAttention layer, BertIntermediate layer and BertOutput layer to create the BertLayer layer.
Now perform a forward pass using previous output layer as input.
class BertLayer(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.chunk_size_feed_forward = config.chunk_size_feed_forward
self.seq_len_dim = 1
self.attention = BertAttention(config)
self.is_decoder = config.is_decoder
self.add_cross_attention = config.add_cross_attention
if self.add_cross_attention:
assert self.is_decoder, f"{self} should be used as a decoder model if cross attention is added"
self.crossattention = BertAttention(config)
self.intermediate = BertIntermediate(config)
self.output = BertOutput(config)
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
):
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
self_attention_outputs = self.attention(
hidden_states,
attention_mask,
head_mask,
output_attentions=output_attentions,
past_key_value=self_attn_past_key_value,
)
attention_output = self_attention_outputs[0]
# if decoder, the last output is tuple of self-attn cache
if self.is_decoder:
outputs = self_attention_outputs[1:-1]
present_key_value = self_attention_outputs[-1]
else:
outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
cross_attn_present_key_value = None
if self.is_decoder and encoder_hidden_states is not None:
assert hasattr(
self, "crossattention"
), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
# cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
cross_attention_outputs = self.crossattention(
attention_output,
attention_mask,
head_mask,
encoder_hidden_states,
encoder_attention_mask,
cross_attn_past_key_value,
output_attentions,
)
attention_output = cross_attention_outputs[0]
outputs = outputs + cross_attention_outputs[1:-1] # add cross attentions if we output attention weights
# add cross-attn cache to positions 3,4 of present_key_value tuple
cross_attn_present_key_value = cross_attention_outputs[-1]
present_key_value = present_key_value + cross_attn_present_key_value
layer_output = apply_chunking_to_forward(
self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
)
outputs = (layer_output,) + outputs
# if decoder, return the attn key/values as the last output
if self.is_decoder:
outputs = outputs + (present_key_value,)
return outputs
def feed_forward_chunk(self, attention_output):
intermediate_output = self.intermediate(attention_output)
layer_output = self.output(intermediate_output, attention_output)
return layer_output
# Assemble block to create Bert Layer.
bert_layer_block = BertLayer(bert_configuraiton)
# Perform feed forward on a whole Bert Layer.
layer_output = bert_layer_block.forward(hidden_states=embedding_output)Attention Head Size:
64
Combined Attentions Head Size:
768
Hidden States:
torch.Size([2, 9, 768])
Query Linear Layer:
torch.Size([2, 9, 768])
Key Linear Layer:
torch.Size([2, 9, 768])
Value Linear Layer:
torch.Size([2, 9, 768])
Query:
torch.Size([2, 12, 9, 64])
Key:
torch.Size([2, 12, 9, 64])
Value:
torch.Size([2, 12, 9, 64])
Key Transposed:
torch.Size([2, 12, 64, 9])
Attention Scores:
torch.Size([2, 12, 9, 9])
Attention Scores Divided by Scalar:
torch.Size([2, 12, 9, 9])
Attention Probabilities Softmax Layer:
torch.Size([2, 12, 9, 9])
Attention Probabilities Dropout Layer:
torch.Size([2, 12, 9, 9])
Context:
torch.Size([2, 12, 9, 64])
Context Permute:
torch.Size([2, 9, 12, 64])
Context Reshaped:
torch.Size([2, 9, 768])
Hidden States:
torch.Size([2, 9, 768])
Hidden States Linear Layer:
torch.Size([2, 9, 768])
Hidden States Dropout Layer:
torch.Size([2, 9, 768])
Hidden States Normalization Layer:
torch.Size([2, 9, 768])
Hidden States:
torch.Size([2, 9, 768])
Hidden States Linear Layer:
torch.Size([2, 9, 3072])
Hidden States Gelu Activation Function:
torch.Size([2, 9, 3072])
Hidden States:
torch.Size([2, 9, 3072])
Hidden States Linear Layer:
torch.Size([2, 9, 768])
Hidden States Dropout Layer:
torch.Size([2, 9, 768])
Hidden States Layer Normalization:
torch.Size([2, 9, 768])Assemble BertEncoder
Put together 12 of the BertLayer layers ( in this setup config.num_hidden_layers=12) to create the BertEncoder layer.
Now perform a forward pass using previous output layer as input.
class BertEncoder(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.layer = torch.nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_values=None,
use_cache=None,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
):
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
next_decoder_cache = () if use_cache else None
for i, layer_module in enumerate(self.layer):
# ADDED
print('\n----------------- BERT LAYER %d -----------------'%(i+1))
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_head_mask = head_mask[i] if head_mask is not None else None
past_key_value = past_key_values[i] if past_key_values is not None else None
if getattr(self.config, "gradient_checkpointing", False):
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs, past_key_value, output_attentions)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(layer_module),
hidden_states,
attention_mask,
layer_head_mask,
encoder_hidden_states,
encoder_attention_mask,
)
else:
layer_outputs = layer_module(
hidden_states,
attention_mask,
layer_head_mask,
encoder_hidden_states,
encoder_attention_mask,
past_key_value,
output_attentions,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[-1],)
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
if self.config.add_cross_attention:
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(
v
for v in [
hidden_states,
next_decoder_cache,
all_hidden_states,
all_self_attentions,
all_cross_attentions,
]
if v is not None
)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=next_decoder_cache,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
cross_attentions=all_cross_attentions,
)
# create bert encoder block by stacking 12 layers
bert_encoder_block = BertEncoder(bert_configuraiton)
# perform forward pass on entire Bert Encoder
encoder_embedding = bert_encoder_block.forward(hidden_states=embedding_output)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
----------------- BERT LAYER 1 -----------------
Hidden States:
torch.Size([2, 9, 768])
Query Linear Layer:
torch.Size([2, 9, 768])
Key Linear Layer:
torch.Size([2, 9, 768])
Value Linear Layer:
torch.Size([2, 9, 768])
Query:
torch.Size([2, 12, 9, 64])
Key:
torch.Size([2, 12, 9, 64])
Value:
torch.Size([2, 12, 9, 64])
Key Transposed:
torch.Size([2, 12, 64, 9])
Attention Scores:
torch.Size([2, 12, 9, 9])
Attention Scores Divided by Scalar:
torch.Size([2, 12, 9, 9])
Attention Probabilities Softmax Layer:
torch.Size([2, 12, 9, 9])
Attention Probabilities Dropout Layer:
torch.Size([2, 12, 9, 9])
Context:
torch.Size([2, 12, 9, 64])
Context Permute:
torch.Size([2, 9, 12, 64])
Context Reshaped:
torch.Size([2, 9, 768])
Hidden States:
torch.Size([2, 9, 768])
Hidden States Linear Layer:
torch.Size([2, 9, 768])
Hidden States Dropout Layer:
torch.Size([2, 9, 768])
Hidden States Normalization Layer:
torch.Size([2, 9, 768])
Hidden States:
torch.Size([2, 9, 768])
Hidden States Linear Layer:
torch.Size([2, 9, 3072])
Hidden States Gelu Activation Function:
torch.Size([2, 9, 3072])
Hidden States:
torch.Size([2, 9, 3072])
Hidden States Linear Layer:
torch.Size([2, 9, 768])
Hidden States Dropout Layer:
torch.Size([2, 9, 768])
Hidden States Layer Normalization:
torch.Size([2, 9, 768])
----------------- BERT LAYER 2 -----------------
...
----------------- BERT LAYER 12 -----------------BertPooler
This layer contains the core of the bert model where the self-attention happens.
The implementation can be found at: transformers/src/transformers/models/bert/modeling_bert.py#L601.
The forward pass uses:
- torch.nn.Linear layer:
self.dense = torch.nn.Linear(config.hidden_size, config.hidden_size)
- torch.nn.Tanh activation function layer:
self.activation = torch.nn.Tanh()
class BertPooler(torch.nn.Module):
def __init__(self, config):
super().__init__()
self.dense = torch.nn.Linear(config.hidden_size, config.hidden_size)
self.activation = torch.nn.Tanh()
def forward(self, hidden_states):
# We "pool" the model by simply taking the hidden state corresponding
# to the first token.
print('\nHidden States:\n', hidden_states.shape)
first_token_tensor = hidden_states[:, 0]
print('\nFirst Token [CLS]:\n', first_token_tensor.shape)
pooled_output = self.dense(first_token_tensor)
print('\nFirst Token [CLS] Linear Layer:\n', pooled_output.shape)
pooled_output = self.activation(pooled_output)
print('\nFirst Token [CLS] Tanh Activation Function:\n', pooled_output.shape)
return pooled_output
# Create bert pooler block.
bert_pooler_block = BertPooler(bert_configuraiton)
# Perform forward pass - encoder_embedding[0] because it is a tuple.
pooled_output = bert_pooler_block(hidden_states=encoder_embedding[0])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])Assemble BertModel
Put together BertEmbeddings layer, BertEncoder layer and BertPooler layer to create the BertModel layer.
Now perform a forward pass using previous output layer as input.
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])Assemble Components
Put together BertModel layer, torch.nn.Dropout layer and torch.nn.Linear layer to create the BertForSequenceClassification model.
Now perform a forward pass using previous output layer as input.
class BertForSequenceClassification(BertPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.bert = BertModel(config)
self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
If :obj:`config.num_labels &amp;amp;gt; 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.bert(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pooled_output = outputs[1]
pooled_output = self.dropout(pooled_output)
logits = self.classifier(pooled_output)
loss = None
if labels is not None:
if self.num_labels == 1:
# We are doing regression
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), labels.view(-1))
else:
loss_fct = torch.nn.CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
# create Bert model with classification layer - BertForSequenceClassificatin
bert_for_sequence_classification_model = BertForSequenceClassification(bert_configuraiton)
# perform forward pass on entire model
outputs = bert_for_sequence_classification_model(**input_sequences)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])Complete Diagram
- If you want a .pdf version of this diagram: bert_inner_workings.pdf.
- If you want a .png version of this diagram: bert_inner_workings.png.
Final Note
If you made it this far Congrats! 🎊 and Thank you! 🙏 for your interest in my tutorial!
I’ve been using this code for a while now and I feel it got to a point where is nicely documented and easy to follow.
Of course is easy for me to follow because I built it. That is why any feedback is welcome and it helps me improve my future tutorials!
If you see something wrong please let me know by opening an issue on my ml_things GitHub repository!
A lot of tutorials out there are mostly a one-time thing and are not being maintained. I plan on keeping my tutorials up to date as much as I can.
This article was originally published on George Mihaila’s personal blog and re-published to TOPBOTS with permission from the author.
Enjoy this article? Sign up for more AI updates.
We’ll let you know when we release more technical education.
Leave a Reply
You must be logged in to post a comment.