bidulki-99's picture
Add files using upload-large-folder tool
f3c3632 verified
Raw History Blame Contribute Delete
4 kB
import torch
from torch import nn
class PatchEmbed2D(nn.Module):
"""
2D Image to Patch Embedding.
Args:
img_size (tuple[int]): Image size.
patch_size (tuple[int]): Patch token size.
in_chans (int): Number of input image channels.
embed_dim(int): Number of projection output channels.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(self, img_size, patch_size, in_chans, embed_dim, norm_layer=None):
super().__init__()
self.img_size = img_size
height, width = img_size
h_patch_size, w_path_size = patch_size
padding_left = padding_right = padding_top = padding_bottom = 0
h_remainder = height % h_patch_size
w_remainder = width % w_path_size
if h_remainder:
h_pad = h_patch_size - h_remainder
padding_top = h_pad // 2
padding_bottom = int(h_pad - padding_top)
if w_remainder:
w_pad = w_path_size - w_remainder
padding_left = w_pad // 2
padding_right = int(w_pad - padding_left)
self.pad = nn.ZeroPad2d((padding_left, padding_right, padding_top, padding_bottom))
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x: torch.Tensor):
B, C, H, W = x.shape
assert H == self.img_size[0] and W == self.img_size[1], \
f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
x = self.pad(x)
x = self.proj(x)
if self.norm is not None:
x = self.norm(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
return x
class PatchEmbed3D(nn.Module):
"""
3D Image to Patch Embedding.
Args:
img_size (tuple[int]): Image size.
patch_size (tuple[int]): Patch token size.
in_chans (int): Number of input image channels.
embed_dim(int): Number of projection output channels.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(self, img_size, patch_size, in_chans, embed_dim, norm_layer=None):
super().__init__()
self.img_size = img_size
level, height, width = img_size
l_patch_size, h_patch_size, w_patch_size = patch_size
padding_left = padding_right = padding_top = padding_bottom = padding_front = padding_back = 0
l_remainder = level % l_patch_size
h_remainder = height % l_patch_size
w_remainder = width % w_patch_size
if l_remainder:
l_pad = l_patch_size - l_remainder
padding_front = l_pad // 2
padding_back = l_pad - padding_front
if h_remainder:
h_pad = h_patch_size - h_remainder
padding_top = h_pad // 2
padding_bottom = h_pad - padding_top
if w_remainder:
w_pad = w_patch_size - w_remainder
padding_left = w_pad // 2
padding_right = w_pad - padding_left
self.pad = nn.ZeroPad3d(
(padding_left, padding_right, padding_top, padding_bottom, padding_front, padding_back)
)
self.proj = nn.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x: torch.Tensor):
B, C, L, H, W = x.shape
assert L == self.img_size[0] and H == self.img_size[1] and W == self.img_size[2], \
f"Input image size ({L}*{H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}*{self.img_size[2]})."
x = self.pad(x)
x = self.proj(x)
if self.norm:
x = self.norm(x.permute(0, 2, 3, 4, 1)).permute(0, 4, 1, 2, 3)
return x