|
import torch |
|
import torch.nn as nn |
|
import math |
|
|
|
|
|
|
|
|
|
|
|
class PositionEmbeddingLearned(nn.Module): |
|
""" |
|
This is a learned version of the position embedding |
|
""" |
|
def __init__(self, num_pos_feats=256): |
|
super().__init__() |
|
self.row_embed = nn.Embedding(32, num_pos_feats) |
|
self.col_embed = nn.Embedding(32, 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).unsqueeze(0).repeat(h, 1, 1) |
|
y_emb = self.row_embed(j).unsqueeze(1).repeat(1, w, 1) |
|
pos = (x_emb + y_emb).permute(2, 0, 1).unsqueeze(0).repeat(x.shape[0], 1, 1, 1) |
|
return pos |
|
|
|
|
|
class PositionEmbeddingSine(nn.Module): |
|
""" |
|
This is a standard version of the position embedding, very similar to the one used by the |
|
"Attention is all you need" paper, generalized to work on examples |
|
""" |
|
def __init__(self, feats_dim=512, temperature=10000, normalize=False, scale=None): |
|
""" |
|
explicitly encode the position using the sinusoid: |
|
PE(pos,2i) = sin(pos/temperature^(2*i/d_model)) |
|
PE(pos,2i+1) = cos(pos/temperature^(2*i/d_model)) |
|
:param feats_dim: the dimension of features, each dimension of the positional embedding to a sinusoid |
|
:param temperature: wavelengths from a geometric progression from scale |
|
:param normalize: whether to normalize the position to (0,1) |
|
:param scale: scale for the position embedding |
|
""" |
|
super(PositionEmbeddingSine, self).__init__() |
|
self.feats_dim = feats_dim |
|
self.T = temperature |
|
self.norm = normalize |
|
if scale is None: |
|
scale = 2 * math.pi |
|
self.scale = scale |
|
|
|
def forward(self, x, mask): |
|
x_embed = mask.cumsum(1, dtype=torch.float32) |
|
y_embed = mask.cumsum(2, dtype=torch.float32) |
|
if self.norm: |
|
eps = 1e-5 |
|
x_embed = x_embed / (x_embed[:, -1:, :] + eps) * self.scale |
|
y_embed = y_embed / (y_embed[:, :, -1:] + eps) * self.scale |
|
|
|
dim_t = torch.arange(self.feats_dim, dtype=torch.float32, device=x.device) |
|
dim_t = self.T ** (2*(dim_t//2)/self.feats_dim) |
|
pos_x = x_embed[:, :, :, None] / dim_t |
|
pos_y = y_embed[:, :, :, None] / dim_t |
|
|
|
pos_x[:, :, :, 0::2], pos_x[:, :, :, 1::2] = pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos() |
|
pos_y[:, :, :, 0::2], pos_y[:, :, :, 1::2] = pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos() |
|
pos = (pos_x + pos_y).permute(0, 3, 1, 2) * 0.5 |
|
return pos |
|
|
|
|
|
def build_position_embed(embed_type='learned', feats_dim=512, temperature=10000): |
|
if embed_type == 'sine': |
|
pos_embed = PositionEmbeddingSine(feats_dim, temperature, normalize=True) |
|
elif embed_type == 'learned': |
|
pos_embed = PositionEmbeddingLearned(feats_dim) |
|
else: |
|
raise ValueError(f"nor supported {embed_type}") |
|
return pos_embed |
|
|
