PSHuman / lib /pymafx /models /attention.py
fffiloni's picture
Migrated from GitHub
2252f3d verified
raw
history blame
15.8 kB
"""
Copyright (c) Microsoft Corporation.
Licensed under the MIT license.
"""
from __future__ import absolute_import, division, print_function, unicode_literals
import logging
import math
import os
import code
import torch
from torch import nn
from .transformers.bert.modeling_bert import BertPreTrainedModel, BertEmbeddings, BertPooler, BertIntermediate, BertOutput, BertSelfOutput
# import src.modeling.data.config as cfg
# from src.modeling._gcnn import GraphConvolution, GraphResBlock
from .transformers.bert.modeling_utils import prune_linear_layer
LayerNormClass = torch.nn.LayerNorm
BertLayerNorm = torch.nn.LayerNorm
from .transformers.bert import BertConfig
class BertSelfAttention(nn.Module):
def __init__(self, config):
super(BertSelfAttention, self).__init__()
if config.hidden_size % config.num_attention_heads != 0:
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.output_attentions = config.output_attentions
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
self.query = nn.Linear(config.hidden_size, self.all_head_size)
self.key = nn.Linear(config.hidden_size, self.all_head_size)
self.value = nn.Linear(config.hidden_size, self.all_head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
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, head_mask=None, history_state=None):
if history_state is not None:
raise
x_states = torch.cat([history_state, hidden_states], dim=1)
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(x_states)
mixed_value_layer = self.value(x_states)
else:
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(hidden_states)
mixed_value_layer = self.value(hidden_states)
# print('mixed_query_layer', mixed_query_layer.shape, mixed_key_layer.shape, mixed_value_layer.shape)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
# print('query_layer', query_layer.shape, key_layer.shape, value_layer.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))
attention_scores = attention_scores / math.sqrt(self.attention_head_size)
# 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 = nn.Softmax(dim=-1)(attention_scores)
# 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)
# Mask heads if we want to
if head_mask is not None:
raise
attention_probs = attention_probs * head_mask
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size, )
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer, )
return outputs
class BertAttention(nn.Module):
def __init__(self, config):
super(BertAttention, self).__init__()
self.self = BertSelfAttention(config)
self.output = BertSelfOutput(config)
def prune_heads(self, heads):
if len(heads) == 0:
return
mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
for head in heads:
mask[head] = 0
mask = mask.view(-1).contiguous().eq(1)
index = torch.arange(len(mask))[mask].long()
# 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
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
def forward(self, input_tensor, attention_mask, head_mask=None, history_state=None):
self_outputs = self.self(input_tensor, attention_mask, head_mask, history_state)
attention_output = self.output(self_outputs[0], input_tensor)
outputs = (attention_output, ) + self_outputs[1:] # add attentions if we output them
return outputs
class AttLayer(nn.Module):
def __init__(self, config):
super(AttLayer, self).__init__()
self.attention = BertAttention(config)
self.intermediate = BertIntermediate(config)
self.output = BertOutput(config)
def MHA(self, hidden_states, attention_mask, head_mask=None, history_state=None):
attention_outputs = self.attention(hidden_states, attention_mask, head_mask, history_state)
attention_output = attention_outputs[0]
# print('attention_output', hidden_states.shape, attention_output.shape)
intermediate_output = self.intermediate(attention_output)
# print('intermediate_output', intermediate_output.shape)
layer_output = self.output(intermediate_output, attention_output)
# print('layer_output', layer_output.shape)
outputs = (layer_output, ) + attention_outputs[1:] # add attentions if we output them
return outputs
def forward(self, hidden_states, attention_mask, head_mask=None, history_state=None):
return self.MHA(hidden_states, attention_mask, head_mask, history_state)
class AttEncoder(nn.Module):
def __init__(self, config):
super(AttEncoder, self).__init__()
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.layer = nn.ModuleList([AttLayer(config) for _ in range(config.num_hidden_layers)])
def forward(self, hidden_states, attention_mask, head_mask=None, encoder_history_states=None):
all_hidden_states = ()
all_attentions = ()
for i, layer_module in enumerate(self.layer):
if self.output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states, )
history_state = None if encoder_history_states is None else encoder_history_states[i]
layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i], history_state)
hidden_states = layer_outputs[0]
if self.output_attentions:
all_attentions = all_attentions + (layer_outputs[1], )
# Add last layer
if self.output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states, )
outputs = (hidden_states, )
if self.output_hidden_states:
outputs = outputs + (all_hidden_states, )
if self.output_attentions:
outputs = outputs + (all_attentions, )
return outputs # outputs, (hidden states), (attentions)
class EncoderBlock(BertPreTrainedModel):
def __init__(self, config):
super(EncoderBlock, self).__init__(config)
self.config = config
# self.embeddings = BertEmbeddings(config)
self.encoder = AttEncoder(config)
# self.pooler = BertPooler(config)
self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
self.img_dim = config.img_feature_dim
try:
self.use_img_layernorm = config.use_img_layernorm
except:
self.use_img_layernorm = None
self.img_embedding = nn.Linear(self.img_dim, self.config.hidden_size, bias=True)
# self.dropout = nn.Dropout(config.hidden_dropout_prob)
if self.use_img_layernorm:
self.LayerNorm = LayerNormClass(config.hidden_size, eps=config.img_layer_norm_eps)
self.apply(self.init_weights)
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,
img_feats,
input_ids=None,
token_type_ids=None,
attention_mask=None,
position_ids=None,
head_mask=None
):
batch_size = len(img_feats)
seq_length = len(img_feats[0])
input_ids = torch.zeros([batch_size, seq_length], dtype=torch.long).to(img_feats.device)
if position_ids is None:
position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device)
position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
# print('-------------------')
# print('position_ids', seq_length, position_ids.shape)
# 494 torch.Size([2, 494])
position_embeddings = self.position_embeddings(position_ids)
# print('position_embeddings', position_embeddings.shape, self.config.max_position_embeddings, self.config.hidden_size)
# torch.Size([2, 494, 1024]) 512 1024
# torch.Size([2, 494, 256]) 512 256
if attention_mask is None:
attention_mask = torch.ones_like(input_ids)
else:
raise
if token_type_ids is None:
token_type_ids = torch.zeros_like(input_ids)
else:
raise
if attention_mask.dim() == 2:
extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
elif attention_mask.dim() == 3:
extended_attention_mask = attention_mask.unsqueeze(1)
else:
raise NotImplementedError
# extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
extended_attention_mask = extended_attention_mask.to(
dtype=img_feats.dtype
) # fp16 compatibility
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
if head_mask is not None:
raise
if head_mask.dim() == 1:
head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
elif head_mask.dim() == 2:
head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(
-1
) # We can specify head_mask for each layer
head_mask = head_mask.to(
dtype=next(self.parameters()).dtype
) # switch to fload if need + fp16 compatibility
else:
head_mask = [None] * self.config.num_hidden_layers
# Project input token features to have spcified hidden size
# print('img_feats', img_feats.shape) # torch.Size([2, 494, 2051])
img_embedding_output = self.img_embedding(img_feats)
# print('img_embedding_output', img_embedding_output.shape) # torch.Size([2, 494, 1024])
# We empirically observe that adding an additional learnable position embedding leads to more stable training
embeddings = position_embeddings + img_embedding_output
if self.use_img_layernorm:
embeddings = self.LayerNorm(embeddings)
# embeddings = self.dropout(embeddings)
# print('extended_attention_mask', extended_attention_mask.shape) # torch.Size([2, 1, 1, 494])
encoder_outputs = self.encoder(embeddings, extended_attention_mask, head_mask=head_mask)
sequence_output = encoder_outputs[0]
outputs = (sequence_output, )
if self.config.output_hidden_states:
all_hidden_states = encoder_outputs[1]
outputs = outputs + (all_hidden_states, )
if self.config.output_attentions:
all_attentions = encoder_outputs[-1]
outputs = outputs + (all_attentions, )
return outputs
def get_att_block(
img_feature_dim=2048,
output_feat_dim=512,
hidden_feat_dim=1024,
num_attention_heads=4,
num_hidden_layers=1
):
config_class = BertConfig
config = config_class.from_pretrained('lib/pymafx/models/transformers/bert/bert-base-uncased/')
interm_size_scale = 2
config.output_attentions = False
# config.hidden_dropout_prob = args.drop_out
config.img_feature_dim = img_feature_dim
# config.output_feature_dim = output_feat_dim
config.hidden_size = hidden_feat_dim
config.intermediate_size = int(config.hidden_size * interm_size_scale)
config.num_hidden_layers = num_hidden_layers
config.num_attention_heads = num_attention_heads
config.max_position_embeddings = 900
# init a transformer encoder and append it to a list
assert config.hidden_size % config.num_attention_heads == 0
att_model = EncoderBlock(config=config)
return att_model
class Graphormer(BertPreTrainedModel):
'''
The archtecture of a transformer encoder block we used in Graphormer
'''
def __init__(self, config):
super(Graphormer, self).__init__(config)
self.config = config
self.bert = EncoderBlock(config)
self.cls_head = nn.Linear(config.hidden_size, self.config.output_feature_dim)
self.residual = nn.Linear(config.img_feature_dim, self.config.output_feature_dim)
self.apply(self.init_weights)
def forward(
self,
img_feats,
input_ids=None,
token_type_ids=None,
attention_mask=None,
masked_lm_labels=None,
next_sentence_label=None,
position_ids=None,
head_mask=None
):
'''
# self.bert has three outputs
# predictions[0]: output tokens
# predictions[1]: all_hidden_states, if enable "self.config.output_hidden_states"
# predictions[2]: attentions, if enable "self.config.output_attentions"
'''
predictions = self.bert(
img_feats=img_feats,
input_ids=input_ids,
position_ids=position_ids,
token_type_ids=token_type_ids,
attention_mask=attention_mask,
head_mask=head_mask
)
# We use "self.cls_head" to perform dimensionality reduction. We don't use it for classification.
pred_score = self.cls_head(predictions[0])
res_img_feats = self.residual(img_feats)
pred_score = pred_score + res_img_feats
# print('pred_score', pred_score.shape)
if self.config.output_attentions and self.config.output_hidden_states:
return pred_score, predictions[1], predictions[-1]
else:
return pred_score