mmdetection/mmdet/models/utils/vlfuse_helper.py

# Copyright (c) OpenMMLab. All rights reserved.
import math

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
from mmcv.cnn.bricks import DropPath
from torch import Tensor

try:
    from transformers import BertPreTrainedModel
    from transformers.modeling_utils import apply_chunking_to_forward
    from transformers.models.bert.modeling_bert import \
        BertAttention as HFBertAttention
    from transformers.models.bert.modeling_bert import \
        BertIntermediate as HFBertIntermediate
    from transformers.models.bert.modeling_bert import \
        BertOutput as HFBertOutput
except ImportError:
    BertPreTrainedModel = object
    apply_chunking_to_forward = None
    HFBertAttention = object
    HFBertIntermediate = object
    HFBertOutput = object

MAX_CLAMP_VALUE = 50000


def permute_and_flatten(layer, N, A, C, H, W):
    layer = layer.view(N, A, C, H, W)
    layer = layer.permute(0, 3, 4, 1, 2)
    layer = layer.reshape(N, -1, C)
    return layer


def clamp_values(vector):
    vector = torch.clamp(vector, min=-MAX_CLAMP_VALUE, max=MAX_CLAMP_VALUE)
    return vector


class BiMultiHeadAttention(nn.Module):
    """Bidirectional fusion Multi-Head Attention layer."""

    def __init__(self,
                 v_dim: int,
                 l_dim: int,
                 embed_dim: int,
                 num_heads: int,
                 dropout: float = 0.1):
        super(BiMultiHeadAttention, self).__init__()

        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.head_dim = embed_dim // num_heads
        self.v_dim = v_dim
        self.l_dim = l_dim

        assert (
            self.head_dim * self.num_heads == self.embed_dim
        ), 'embed_dim must be divisible by num_heads ' \
           f'(got `embed_dim`: {self.embed_dim} ' \
           f'and `num_heads`: {self.num_heads}).'
        self.scale = self.head_dim**(-0.5)
        self.dropout = dropout

        self.v_proj = nn.Linear(self.v_dim, self.embed_dim)
        self.l_proj = nn.Linear(self.l_dim, self.embed_dim)
        self.values_v_proj = nn.Linear(self.v_dim, self.embed_dim)
        self.values_l_proj = nn.Linear(self.l_dim, self.embed_dim)

        self.out_v_proj = nn.Linear(self.embed_dim, self.v_dim)
        self.out_l_proj = nn.Linear(self.embed_dim, self.l_dim)

        self.stable_softmax_2d = False
        self.clamp_min_for_underflow = True
        self.clamp_max_for_overflow = True

        self._reset_parameters()

    def _shape(self, tensor: Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads,
                           self.head_dim).transpose(1, 2).contiguous()

    def _reset_parameters(self):
        nn.init.xavier_uniform_(self.v_proj.weight)
        self.v_proj.bias.data.fill_(0)
        nn.init.xavier_uniform_(self.l_proj.weight)
        self.l_proj.bias.data.fill_(0)
        nn.init.xavier_uniform_(self.values_v_proj.weight)
        self.values_v_proj.bias.data.fill_(0)
        nn.init.xavier_uniform_(self.values_l_proj.weight)
        self.values_l_proj.bias.data.fill_(0)
        nn.init.xavier_uniform_(self.out_v_proj.weight)
        self.out_v_proj.bias.data.fill_(0)
        nn.init.xavier_uniform_(self.out_l_proj.weight)
        self.out_l_proj.bias.data.fill_(0)

    def forward(self, vision: Tensor, lang: Tensor, attention_mask_l=None):
        bsz, tgt_len, _ = vision.size()

        query_states = self.v_proj(vision) * self.scale
        key_states = self._shape(self.l_proj(lang), -1, bsz)
        value_v_states = self._shape(self.values_v_proj(vision), -1, bsz)
        value_l_states = self._shape(self.values_l_proj(lang), -1, bsz)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len,
                                   bsz).view(*proj_shape)
        key_states = key_states.view(*proj_shape)
        value_v_states = value_v_states.view(*proj_shape)
        value_l_states = value_l_states.view(*proj_shape)

        src_len = key_states.size(1)
        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))

        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
            raise ValueError(
                f'Attention weights should be of '
                f'size {(bsz * self.num_heads, tgt_len, src_len)}, '
                f'but is {attn_weights.size()}')

        if self.stable_softmax_2d:
            attn_weights = attn_weights - attn_weights.max()

        if self.clamp_min_for_underflow:
            # Do not increase -50000, data type half has quite limited range
            attn_weights = torch.clamp(attn_weights, min=-MAX_CLAMP_VALUE)
        if self.clamp_max_for_overflow:
            # Do not increase 50000, data type half has quite limited range
            attn_weights = torch.clamp(attn_weights, max=MAX_CLAMP_VALUE)

        attn_weights_T = attn_weights.transpose(1, 2)
        attn_weights_l = (
            attn_weights_T -
            torch.max(attn_weights_T, dim=-1, keepdim=True)[0])
        if self.clamp_min_for_underflow:
            # Do not increase -50000, data type half has quite limited range
            attn_weights_l = torch.clamp(attn_weights_l, min=-MAX_CLAMP_VALUE)
        if self.clamp_max_for_overflow:
            # Do not increase 50000, data type half has quite limited range
            attn_weights_l = torch.clamp(attn_weights_l, max=MAX_CLAMP_VALUE)

        attn_weights_l = attn_weights_l.softmax(dim=-1)

        if attention_mask_l is not None:
            assert (attention_mask_l.dim() == 2)
            attention_mask = attention_mask_l.unsqueeze(1).unsqueeze(1)
            attention_mask = attention_mask.expand(bsz, 1, tgt_len, src_len)
            attention_mask = attention_mask.masked_fill(
                attention_mask == 0, -9e15)

            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError('Attention mask should be of '
                                 f'size {(bsz, 1, tgt_len, src_len)}')
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len,
                                             src_len) + attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len,
                                             src_len)

        attn_weights_v = nn.functional.softmax(attn_weights, dim=-1)

        attn_probs_v = F.dropout(
            attn_weights_v, p=self.dropout, training=self.training)
        attn_probs_l = F.dropout(
            attn_weights_l, p=self.dropout, training=self.training)

        attn_output_v = torch.bmm(attn_probs_v, value_l_states)
        attn_output_l = torch.bmm(attn_probs_l, value_v_states)

        if attn_output_v.size() != (bsz * self.num_heads, tgt_len,
                                    self.head_dim):
            raise ValueError(
                '`attn_output_v` should be of '
                f'size {(bsz, self.num_heads, tgt_len, self.head_dim)}, '
                f'but is {attn_output_v.size()}')

        if attn_output_l.size() != (bsz * self.num_heads, src_len,
                                    self.head_dim):
            raise ValueError(
                '`attn_output_l` should be of size '
                f'{(bsz, self.num_heads, src_len, self.head_dim)}, '
                f'but is {attn_output_l.size()}')

        attn_output_v = attn_output_v.view(bsz, self.num_heads, tgt_len,
                                           self.head_dim)
        attn_output_v = attn_output_v.transpose(1, 2)
        attn_output_v = attn_output_v.reshape(bsz, tgt_len, self.embed_dim)

        attn_output_l = attn_output_l.view(bsz, self.num_heads, src_len,
                                           self.head_dim)
        attn_output_l = attn_output_l.transpose(1, 2)
        attn_output_l = attn_output_l.reshape(bsz, src_len, self.embed_dim)

        attn_output_v = self.out_v_proj(attn_output_v)
        attn_output_l = self.out_l_proj(attn_output_l)

        return attn_output_v, attn_output_l


class BiAttentionBlock(nn.Module):
    """BiAttentionBlock Module:

    First, multi-level visual features are concat; Then the concat visual
    feature and lang feature are fused by attention; Finally the newly visual
    feature are split into multi levels.
    """

    def __init__(self,
                 v_dim: int,
                 l_dim: int,
                 embed_dim: int,
                 num_heads: int,
                 dropout: float = 0.1,
                 drop_path: float = .0,
                 init_values: float = 1e-4):
        super().__init__()

        # pre layer norm
        self.layer_norm_v = nn.LayerNorm(v_dim)
        self.layer_norm_l = nn.LayerNorm(l_dim)
        self.attn = BiMultiHeadAttention(
            v_dim=v_dim,
            l_dim=l_dim,
            embed_dim=embed_dim,
            num_heads=num_heads,
            dropout=dropout)

        # add layer scale for training stability
        self.drop_path = DropPath(
            drop_path) if drop_path > 0. else nn.Identity()
        self.gamma_v = nn.Parameter(
            init_values * torch.ones(v_dim), requires_grad=True)
        self.gamma_l = nn.Parameter(
            init_values * torch.ones(l_dim), requires_grad=True)

    def forward(self,
                visual_features: list,
                lang_feature: Tensor,
                attention_mask_l=None):

        size_per_level, visual_features_flatten = [], []
        for i, feat_per_level in enumerate(visual_features):
            bs, c, h, w = feat_per_level.shape
            size_per_level.append([h, w])
            feat = permute_and_flatten(feat_per_level, bs, -1, c, h, w)
            visual_features_flatten.append(feat)
        visual_features_flatten = torch.cat(visual_features_flatten, dim=1)
        new_v, new_lang_feature = self.single_attention_call(
            visual_features_flatten,
            lang_feature,
            attention_mask_l=attention_mask_l)
        # [bs, N, C] -> [bs, C, N]
        new_v = new_v.transpose(1, 2).contiguous()

        start = 0
        fusion_visual_features = []
        for (h, w) in size_per_level:
            new_v_per_level = new_v[:, :,
                                    start:start + h * w].view(bs, -1, h,
                                                              w).contiguous()
            fusion_visual_features.append(new_v_per_level)
            start += h * w

        return fusion_visual_features, new_lang_feature

    def single_attention_call(self, visual, lang, attention_mask_l=None):
        visual = self.layer_norm_v(visual)
        lang = self.layer_norm_l(lang)
        delta_v, delta_l = self.attn(
            visual, lang, attention_mask_l=attention_mask_l)
        # visual, lang = visual + delta_v, l + delta_l
        visual = visual + self.drop_path(self.gamma_v * delta_v)
        lang = lang + self.drop_path(self.gamma_l * delta_l)
        return visual, lang


class VLFuse(nn.Module):
    """Early Fusion Module."""

    def __init__(self,
                 v_dim: int = 256,
                 l_dim: int = 768,
                 embed_dim: int = 2048,
                 num_heads: int = 8,
                 dropout: float = 0.1,
                 drop_path: float = 0.0,
                 use_checkpoint: bool = False):
        super().__init__()
        # bi-direction (text->image, image->text)
        self.use_checkpoint = use_checkpoint
        self.b_attn = BiAttentionBlock(
            v_dim=v_dim,
            l_dim=l_dim,
            embed_dim=embed_dim,
            num_heads=num_heads,
            dropout=dropout,
            drop_path=drop_path,
            init_values=1.0 / 6.0)

    def forward(self, x):
        visual_features = x['visual']
        language_dict_features = x['lang']

        if self.use_checkpoint:
            fused_visual_features, language_features = checkpoint.checkpoint(
                self.b_attn, visual_features, language_dict_features['hidden'],
                language_dict_features['masks'])
        else:
            fused_visual_features, language_features = self.b_attn(
                visual_features, language_dict_features['hidden'],
                language_dict_features['masks'])

        language_dict_features['hidden'] = language_features
        fused_language_dict_features = language_dict_features

        features_dict = {
            'visual': fused_visual_features,
            'lang': fused_language_dict_features
        }

        return features_dict


class BertEncoderLayer(BertPreTrainedModel):
    """Modified from transformers.models.bert.modeling_bert.BertLayer."""

    def __init__(self,
                 config,
                 clamp_min_for_underflow: bool = False,
                 clamp_max_for_overflow: bool = False):
        super().__init__(config)
        self.config = config
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1

        self.attention = BertAttention(config, clamp_min_for_underflow,
                                       clamp_max_for_overflow)
        self.intermediate = BertIntermediate(config)
        self.output = BertOutput(config)

    def forward(self, inputs):
        language_dict_features = inputs['lang']
        hidden_states = language_dict_features['hidden']
        attention_mask = language_dict_features['masks']

        device = hidden_states.device
        input_shape = hidden_states.size()[:-1]
        # 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 = self.get_extended_attention_mask(
            attention_mask, input_shape, device)

        self_attention_outputs = self.attention(
            hidden_states,
            extended_attention_mask,
            None,
            output_attentions=False,
            past_key_value=None,
        )
        attention_output = self_attention_outputs[0]
        outputs = self_attention_outputs[
            1:]  # add self attentions if we output attention weights
        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
        hidden_states = outputs[0]

        language_dict_features['hidden'] = hidden_states

        features_dict = {
            'visual': inputs['visual'],
            'lang': language_dict_features
        }

        return features_dict

    def feed_forward_chunk(self, attention_output):
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        return layer_output


# The following code is the same as the Huggingface code,
# with the only difference being the additional clamp operation.
class BertSelfAttention(nn.Module):
    """BERT self-attention layer from Huggingface transformers.

    Compared to the BertSelfAttention of Huggingface, only add the clamp.
    """

    def __init__(self,
                 config,
                 clamp_min_for_underflow: bool = False,
                 clamp_max_for_overflow: bool = False):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and \
                not hasattr(config, 'embedding_size'):
            raise ValueError(f'The hidden size ({config.hidden_size}) is '
                             'not a multiple of the number of attention '
                             f'heads ({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

        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)
        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.clamp_min_for_underflow = clamp_min_for_underflow
        self.clamp_max_for_overflow = clamp_max_for_overflow

        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,
    ):
        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:
            # 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:
            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:
            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:
            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)

        if self.is_decoder:
            past_key_value = (key_layer, value_layer)

        # 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))

        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->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->bhlr', query_layer, positional_embedding)
                relative_position_scores_key = torch.einsum(
                    'bhrd,lrd->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)

        if self.clamp_min_for_underflow:
            attention_scores = torch.clamp(
                attention_scores, min=-MAX_CLAMP_VALUE
            )  # Do not increase -50000, data type half has quite limited range
        if self.clamp_max_for_overflow:
            attention_scores = torch.clamp(
                attention_scores, max=MAX_CLAMP_VALUE
            )  # Do not increase 50000, data type half has quite limited range

        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 = 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:
            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 output_attentions else (context_layer, )

        if self.is_decoder:
            outputs = outputs + (past_key_value, )
        return outputs


class BertAttention(HFBertAttention):
    """BertAttention is made up of self-attention and intermediate+output.

    Compared to the BertAttention of Huggingface, only add the clamp.
    """

    def __init__(self,
                 config,
                 clamp_min_for_underflow: bool = False,
                 clamp_max_for_overflow: bool = False):
        super().__init__(config)
        self.self = BertSelfAttention(config, clamp_min_for_underflow,
                                      clamp_max_for_overflow)


class BertIntermediate(HFBertIntermediate):

    def forward(self, hidden_states):
        hidden_states = self.dense(hidden_states)
        hidden_states = clamp_values(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        hidden_states = clamp_values(hidden_states)
        return hidden_states


class BertOutput(HFBertOutput):

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = clamp_values(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        hidden_states = clamp_values(hidden_states)
        return hidden_states