Spaces:
Running
Running
File size: 3,353 Bytes
56cd6b7 |
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 |
# Copyright (C) 2024-present Naver Corporation. All rights reserved.
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
#
# --------------------------------------------------------
# PatchEmbed implementation for DUST3R,
# in particular ManyAR_PatchEmbed that Handle images with non-square aspect ratio
# --------------------------------------------------------
import torch
import dust3r.utils.path_to_croco # noqa: F401
from models.blocks import PatchEmbed # noqa
def get_patch_embed(patch_embed_cls, img_size, patch_size, enc_embed_dim):
assert patch_embed_cls in ['PatchEmbedDust3R', 'ManyAR_PatchEmbed']
patch_embed = eval(patch_embed_cls)(img_size, patch_size, 3, enc_embed_dim)
return patch_embed
class PatchEmbedDust3R(PatchEmbed):
def forward(self, x, **kw):
B, C, H, W = x.shape # 输入图片的尺寸得是16的倍数
assert H % self.patch_size[0] == 0, f"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]})."
assert W % self.patch_size[1] == 0, f"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]})."
x = self.proj(x) # 这是一个Conv2d卷积，即ViT的Patch_Embedding操作，输出编码的维度为1024，卷积核尺寸and步长都是16
pos = self.position_getter(B, x.size(2), x.size(3), x.device) # PositionGetter，位置编码
if self.flatten:
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC
x = self.norm(x) # nn.Identity()
return x, pos
class ManyAR_PatchEmbed (PatchEmbed):
""" Handle images with non-square aspect ratio.
All images in the same batch have the same aspect ratio.
true_shape = [(height, width) ...] indicates the actual shape of each image.
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True):
self.embed_dim = embed_dim
super().__init__(img_size, patch_size, in_chans, embed_dim, norm_layer, flatten)
def forward(self, img, true_shape):
B, C, H, W = img.shape
assert W >= H, f'img should be in landscape mode, but got {W=} {H=}'
assert H % self.patch_size[0] == 0, f"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]})."
assert W % self.patch_size[1] == 0, f"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]})."
assert true_shape.shape == (B, 2), f"true_shape has the wrong shape={true_shape.shape}"
# size expressed in tokens
W //= self.patch_size[0]
H //= self.patch_size[1]
n_tokens = H * W
height, width = true_shape.T
is_landscape = (width >= height)
is_portrait = ~is_landscape
# allocate result
x = img.new_zeros((B, n_tokens, self.embed_dim))
pos = img.new_zeros((B, n_tokens, 2), dtype=torch.int64)
# linear projection, transposed if necessary
x[is_landscape] = self.proj(img[is_landscape]).permute(0, 2, 3, 1).flatten(1, 2).float()
x[is_portrait] = self.proj(img[is_portrait].swapaxes(-1, -2)).permute(0, 2, 3, 1).flatten(1, 2).float()
pos[is_landscape] = self.position_getter(1, H, W, pos.device)
pos[is_portrait] = self.position_getter(1, W, H, pos.device)
x = self.norm(x)
return x, pos
|