File size: 5,150 Bytes
844f7c0 |
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 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# Position embedding utils
# --------------------------------------------------------
import numpy as np
import torch
# --------------------------------------------------------
# 2D sine-cosine position embedding
# References:
# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py
# MoCo v3: https://github.com/facebookresearch/moco-v3
# --------------------------------------------------------
def get_2d_sincos_pos_embed(embed_dim, grid_size_h, grid_size_w, cls_token=False):
"""
grid_size: int of the grid height and width
return:
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
grid_h = np.arange(grid_size_h, dtype=np.float32)
grid_w = np.arange(grid_size_w, dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_size_h, grid_size_w])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token:
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
return pos_embed
def get_nd_sincos_pos_embed(embed_dim,dims, cls_token=False):
"""
grid_size: int of the grid height and width
return:
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
grid_1ds= [np.arange(d, dtype=np.float32)for d in dims]
grid = np.meshgrid(*grid_1ds) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid[:,None]
embed_dims = np.array([embed_dim//(2*len(dims))*2] * len(dims))
embed_dims[-1] += embed_dim - sum(embed_dims)
pos_embeds = [get_1d_sincos_pos_embed_from_grid(d, g) for d,g in zip(embed_dims,grid)]
pos_embed = np.concatenate(pos_embeds, axis=1) # (H*W, D)
pos_embed = pos_embed.reshape(embed_dim,*dims)
assert not cls_token # nnot implemennted
if cls_token:
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
return pos_embed
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
assert embed_dim % 2 == 0
# use half of dimensions to encode grid_h
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
return emb
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
"""
embed_dim: output dimension for each position
pos: a list of positions to be encoded: size (M,)
out: (M, D)
"""
assert embed_dim % 2 == 0
omega = np.arange(embed_dim // 2, dtype=np.float32)
omega /= embed_dim / 2.0
omega = 1.0 / 10000**omega # (D/2,)
pos = pos.reshape(-1) # (M,)
out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product
emb_sin = np.sin(out) # (M, D/2)
emb_cos = np.cos(out) # (M, D/2)
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
return emb
# --------------------------------------------------------
# Interpolate position embeddings for high-resolution
# References:
# DeiT: https://github.com/facebookresearch/deit
# --------------------------------------------------------
def interpolate_pos_embed(model, checkpoint_model, new_size=(64, 128)):
if "net.pos_embed" in checkpoint_model:
pos_embed_checkpoint = checkpoint_model["net.pos_embed"]
embedding_size = pos_embed_checkpoint.shape[-1]
orig_num_patches = pos_embed_checkpoint.shape[-2]
patch_size = model.patch_size
w_h_ratio = 2
orig_h = int((orig_num_patches // w_h_ratio) ** 0.5)
orig_w = w_h_ratio * orig_h
orig_size = (orig_h, orig_w)
new_size = (new_size[0] // patch_size, new_size[1] // patch_size)
# print (orig_size)
# print (new_size)
if orig_size[0] != new_size[0]:
print("Interpolate PEs from %dx%d to %dx%d" % (orig_size[0], orig_size[1], new_size[0], new_size[1]))
pos_tokens = pos_embed_checkpoint.reshape(-1, orig_size[0], orig_size[1], embedding_size).permute(
0, 3, 1, 2
)
new_pos_tokens = torch.nn.functional.interpolate(
pos_tokens, size=(new_size[0], new_size[1]), mode="bicubic", align_corners=False
)
new_pos_tokens = new_pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
checkpoint_model["net.pos_embed"] = new_pos_tokens
def interpolate_channel_embed(checkpoint_model, new_len):
if "net.channel_embed" in checkpoint_model:
channel_embed_checkpoint = checkpoint_model["net.channel_embed"]
old_len = channel_embed_checkpoint.shape[1]
if new_len <= old_len:
checkpoint_model["net.channel_embed"] = channel_embed_checkpoint[:, :new_len]
|