Spaces:
Runtime error
Runtime error
File size: 4,961 Bytes
9d0a4ae |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 |
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Various positional encodings for the transformer.
"""
import math
import torch
from torch import nn
import numpy as np
def PositionalEncoding(n_position, d_hid):
def get_position_angle_vec(position, d_hid):
return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)]
sinusoid_table = np.array([get_position_angle_vec(pos_i, d_hid) for pos_i in range(n_position)])
sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i
sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1
return torch.FloatTensor(sinusoid_table) # shape:(1, maxLen(n_position), d_hid)
class TrainablePositionalEncoding(nn.Module):
"""Construct the embeddings from word, position and token_type embeddings.
"""
def __init__(self, max_position_embeddings, hidden_size, dropout=0.1):
super(TrainablePositionalEncoding, self).__init__()
self.position_embeddings = nn.Embedding(max_position_embeddings, hidden_size)
self.LayerNorm = nn.LayerNorm(hidden_size)
self.dropout = nn.Dropout(dropout)
def forward(self, input_feat):
"""
Args:
input_feat: (N, L, D)
"""
bsz, seq_length = input_feat.shape[:2]
position_ids = torch.arange(seq_length, dtype=torch.long, device=input_feat.device)
position_ids = position_ids.unsqueeze(0).repeat(bsz, 1) # (N, L)
position_embeddings = self.position_embeddings(position_ids)
embeddings = self.LayerNorm(input_feat + position_embeddings)
embeddings = self.dropout(embeddings)
return embeddings
class PositionEmbeddingSine(nn.Module):
"""
This is a more standard version of the position embedding, very similar to the one
used by the Attention is all you need paper, generalized to work on images. (To 1D sequences)
"""
def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
super().__init__()
self.num_pos_feats = num_pos_feats
self.temperature = temperature
self.normalize = normalize
if scale is not None and normalize is False:
raise ValueError("normalize should be True if scale is passed")
if scale is None:
scale = 2 * math.pi
self.scale = scale
def forward(self, x, mask):
"""
Args:
x: torch.tensor, (batch_size, L, d)
mask: torch.tensor, (batch_size, L), with 1 as valid
Returns:
"""
assert mask is not None
x_embed = mask.cumsum(1, dtype=torch.float32) # (bsz, L)
if self.normalize:
eps = 1e-6
x_embed = x_embed / (x_embed[:, -1:] + eps) * self.scale
dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
# import pdb; pdb.set_trace()
# dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
dim_t = self.temperature ** (2 * torch.div(dim_t, 2).int() / self.num_pos_feats)
pos_x = x_embed[:, :, None] / dim_t # (bsz, L, num_pos_feats)
pos_x = torch.stack((pos_x[:, :, 0::2].sin(), pos_x[:, :, 1::2].cos()), dim=3).flatten(2) # (bsz, L, num_pos_feats*2)
# import ipdb; ipdb.set_trace()
return pos_x # .permute(0, 2, 1) # (bsz, num_pos_feats*2, L)
class PositionEmbeddingLearned(nn.Module):
"""
Absolute pos embedding, learned.
"""
def __init__(self, num_pos_feats=256):
super().__init__()
self.row_embed = nn.Embedding(50, num_pos_feats)
self.col_embed = nn.Embedding(50, num_pos_feats)
self.reset_parameters()
def reset_parameters(self):
nn.init.uniform_(self.row_embed.weight)
nn.init.uniform_(self.col_embed.weight)
def forward(self, x, mask):
h, w = x.shape[-2:]
i = torch.arange(w, device=x.device)
j = torch.arange(h, device=x.device)
x_emb = self.col_embed(i)
y_emb = self.row_embed(j)
pos = torch.cat([
x_emb.unsqueeze(0).repeat(h, 1, 1),
y_emb.unsqueeze(1).repeat(1, w, 1),
], dim=-1).permute(2, 0, 1).unsqueeze(0).repeat(x.shape[0], 1, 1, 1)
return pos
def build_position_encoding(args):
N_steps = args.hidden_dim
if args.position_embedding in ('v2', 'sine'):
# TODO find a better way of exposing other arguments
position_embedding = PositionEmbeddingSine(N_steps, normalize=True)
# elif args.position_embedding in ('v3', 'learned'):
# position_embedding = PositionEmbeddingLearned(N_steps)
else:
raise ValueError(f"not supported {args.position_embedding}")
txt_pos_embed = TrainablePositionalEncoding(
max_position_embeddings=args.max_q_l,
hidden_size=args.hidden_dim, dropout=args.input_dropout)
return position_embedding, txt_pos_embed
|