File size: 9,755 Bytes
fa0f216 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 |
import numpy as np
import torch
import torch.nn as nn
def get_emb(sin_inp):
"""
Gets a base embedding for one dimension with sin and cos intertwined
"""
emb = torch.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
return torch.flatten(emb, -2, -1)
class PositionalEncoding1D(nn.Module):
def __init__(self, channels):
"""
:param channels: The last dimension of the tensor you want to apply pos emb to.
"""
super(PositionalEncoding1D, self).__init__()
self.org_channels = channels
channels = int(np.ceil(channels / 2) * 2)
self.channels = channels
inv_freq = 1.0 / (10000 ** (torch.arange(0, channels, 2).float() / channels))
self.register_buffer("inv_freq", inv_freq, persistent=False)
self.cached_penc = None
def forward(self, tensor):
"""
:param tensor: A 3d tensor of size (batch_size, x, ch)
:return: Positional Encoding Matrix of size (batch_size, x, ch)
"""
if len(tensor.shape) != 3:
raise RuntimeError("The input tensor has to be 3d!")
if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
return self.cached_penc
self.cached_penc = None
batch_size, x, orig_ch = tensor.shape
pos_x = torch.arange(x, device=tensor.device).type(self.inv_freq.type())
sin_inp_x = torch.einsum("i,j->ij", pos_x, self.inv_freq)
emb_x = get_emb(sin_inp_x)
emb = torch.zeros((x, self.channels), device=tensor.device).type(tensor.type())
emb[:, : self.channels] = emb_x
self.cached_penc = emb[None, :, :orig_ch].repeat(batch_size, 1, 1)
return self.cached_penc
class PositionalEncodingPermute1D(nn.Module):
def __init__(self, channels):
"""
Accepts (batchsize, ch, x) instead of (batchsize, x, ch)
"""
super(PositionalEncodingPermute1D, self).__init__()
self.penc = PositionalEncoding1D(channels)
def forward(self, tensor):
tensor = tensor.permute(0, 2, 1)
enc = self.penc(tensor)
return enc.permute(0, 2, 1)
@property
def org_channels(self):
return self.penc.org_channels
class PositionalEncoding2D(nn.Module):
def __init__(self, channels):
"""
:param channels: The last dimension of the tensor you want to apply pos emb to.
"""
super(PositionalEncoding2D, self).__init__()
self.org_channels = channels
channels = int(np.ceil(channels / 4) * 2)
self.channels = channels
inv_freq = 1.0 / (10000 ** (torch.arange(0, channels, 2).float() / channels))
self.register_buffer("inv_freq", inv_freq)
self.cached_penc = None
def forward(self, tensor):
"""
:param tensor: A 4d tensor of size (batch_size, x, y, ch)
:return: Positional Encoding Matrix of size (batch_size, x, y, ch)
"""
if len(tensor.shape) != 4:
raise RuntimeError("The input tensor has to be 4d!")
if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
return self.cached_penc
self.cached_penc = None
batch_size, x, y, orig_ch = tensor.shape
pos_x = torch.arange(x, device=tensor.device).type(self.inv_freq.type())
pos_y = torch.arange(y, device=tensor.device).type(self.inv_freq.type())
sin_inp_x = torch.einsum("i,j->ij", pos_x, self.inv_freq)
sin_inp_y = torch.einsum("i,j->ij", pos_y, self.inv_freq)
emb_x = get_emb(sin_inp_x).unsqueeze(1)
emb_y = get_emb(sin_inp_y)
emb = torch.zeros((x, y, self.channels * 2), device=tensor.device).type(
tensor.type()
)
emb[:, :, : self.channels] = emb_x
emb[:, :, self.channels : 2 * self.channels] = emb_y
self.cached_penc = emb[None, :, :, :orig_ch].repeat(tensor.shape[0], 1, 1, 1)
return self.cached_penc
class PositionalEncodingPermute2D(nn.Module):
def __init__(self, channels):
"""
Accepts (batchsize, ch, x, y) instead of (batchsize, x, y, ch)
"""
super(PositionalEncodingPermute2D, self).__init__()
self.penc = PositionalEncoding2D(channels)
def forward(self, tensor):
tensor = tensor.permute(0, 2, 3, 1)
enc = self.penc(tensor)
return enc.permute(0, 3, 1, 2)
@property
def org_channels(self):
return self.penc.org_channels
class PositionalEncoding3D(nn.Module):
def __init__(self, channels):
"""
:param channels: The last dimension of the tensor you want to apply pos emb to.
"""
super(PositionalEncoding3D, self).__init__()
self.org_channels = channels
channels = int(np.ceil(channels / 6) * 2)
if channels % 2:
channels += 1
self.channels = channels
inv_freq = 1.0 / (10000 ** (torch.arange(0, channels, 2).float() / channels))
self.register_buffer("inv_freq", inv_freq)
self.cached_penc = None
def forward(self, tensor):
"""
:param tensor: A 5d tensor of size (batch_size, x, y, z, ch)
:return: Positional Encoding Matrix of size (batch_size, x, y, z, ch)
"""
if len(tensor.shape) != 5:
raise RuntimeError("The input tensor has to be 5d!")
if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
return self.cached_penc
self.cached_penc = None
batch_size, x, y, z, orig_ch = tensor.shape
pos_x = torch.arange(x, device=tensor.device).type(self.inv_freq.type())
pos_y = torch.arange(y, device=tensor.device).type(self.inv_freq.type())
pos_z = torch.arange(z, device=tensor.device).type(self.inv_freq.type())
sin_inp_x = torch.einsum("i,j->ij", pos_x, self.inv_freq)
sin_inp_y = torch.einsum("i,j->ij", pos_y, self.inv_freq)
sin_inp_z = torch.einsum("i,j->ij", pos_z, self.inv_freq)
emb_x = get_emb(sin_inp_x).unsqueeze(1).unsqueeze(1)
emb_y = get_emb(sin_inp_y).unsqueeze(1)
emb_z = get_emb(sin_inp_z)
emb = torch.zeros((x, y, z, self.channels * 3), device=tensor.device).type(
tensor.type()
)
emb[:, :, :, : self.channels] = emb_x
emb[:, :, :, self.channels : 2 * self.channels] = emb_y
emb[:, :, :, 2 * self.channels :] = emb_z
self.cached_penc = emb[None, :, :, :, :orig_ch].repeat(batch_size, 1, 1, 1, 1)
return self.cached_penc
class PositionalEncodingPermute3D(nn.Module):
def __init__(self, channels):
"""
Accepts (batchsize, ch, x, y, z) instead of (batchsize, x, y, z, ch)
"""
super(PositionalEncodingPermute3D, self).__init__()
self.penc = PositionalEncoding3D(channels)
def forward(self, tensor):
tensor = tensor.permute(0, 2, 3, 4, 1)
enc = self.penc(tensor)
return enc.permute(0, 4, 1, 2, 3)
@property
def org_channels(self):
return self.penc.org_channels
class Summer(nn.Module):
def __init__(self, penc):
"""
:param model: The type of positional encoding to run the summer on.
"""
super(Summer, self).__init__()
self.penc = penc
def forward(self, tensor):
"""
:param tensor: A 3, 4 or 5d tensor that matches the model output size
:return: Positional Encoding Matrix summed to the original tensor
"""
penc = self.penc(tensor)
assert (
tensor.size() == penc.size()
), "The original tensor size {} and the positional encoding tensor size {} must match!".format(
tensor.size(), penc.size()
)
return tensor + penc
class SparsePositionalEncoding2D(PositionalEncoding2D):
def __init__(self, channels, x, y, device='cuda'):
super(SparsePositionalEncoding2D, self).__init__(channels)
self.y, self.x = y, x
self.fake_tensor = torch.zeros((1, x, y, channels), device=device)
def forward(self, coords):
"""
:param coords: A list of list of coordinates (((x1, y1), (x2, y22), ... ), ... )
:return: Positional Encoding Matrix summed to the original tensor
"""
encodings = super().forward(self.fake_tensor)
encodings = encodings.permute(0, 3, 1, 2)
indices = torch.nn.utils.rnn.pad_sequence([torch.LongTensor(c) for c in coords], batch_first=True, padding_value=-1)
indices = indices.unsqueeze(0).to(self.fake_tensor.device)
assert self.x == self.y
indices = (indices + 0.5) / self.x * 2 - 1
indices = torch.flip(indices, (-1, ))
return torch.nn.functional.grid_sample(encodings, indices).squeeze().permute(2, 1, 0)
# all_encodings = []
# for coords_row in coords:
# res_encodings = []
# for xy in coords_row:
# if xy is None:
# res_encodings.append(padding)
# else:
# x, y = xy
# res_encodings.append(encodings[x, y, :])
# all_encodings.append(res_encodings)
# return torch.stack(res_encodings).to(self.fake_tensor.device)
# coords = torch.Tensor(coords).to(self.fake_tensor.device).long()
# assert torch.all(coords[:, 0] < self.x)
# assert torch.all(coords[:, 1] < self.y)
# coords = coords[:, 0] + (coords[:, 1] * self.x)
# encodings = super().forward(self.fake_tensor).reshape((-1, self.org_channels))
# return encodings[coords]
if __name__ == '__main__':
pos = SparsePositionalEncoding2D(10, 10, 20)
pos([[0, 0], [0, 9], [1, 0], [9, 15]])
|