croco/models/pos_embed.py

# Copyright (C) 2022-present Naver Corporation. All rights reserved.
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).


# --------------------------------------------------------
# Position embedding utils
# --------------------------------------------------------



import numpy as np

import torch

# --------------------------------------------------------
# 2D sine-cosine position embedding
# References:
# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
# 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, n_cls_token=0):
    """
    grid_size: tuple (height, width) of the grid
    return:
    pos_embed: [grid_size[0]*grid_size[1], embed_dim] or [n_cls_token+grid_size[0]*grid_size[1], embed_dim] (w/ or w/o cls_token)
    """
    grid_h = np.arange(grid_size[0], dtype=np.float32)
    grid_w = np.arange(grid_size[1], 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[0], grid_size[1]])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if n_cls_token>0:
        pos_embed = np.concatenate([np.zeros([n_cls_token, 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=float)
    omega /= embed_dim / 2.
    omega = 1. / 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:
# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
# DeiT: https://github.com/facebookresearch/deit
# --------------------------------------------------------
def interpolate_pos_embed(model, checkpoint_model):
    keys = ['enc_pos_embed']+(['dec_pos_embed'] if hasattr(model,'dec_blocks') else [])
    img_size = model.patch_embed.img_size
    if isinstance(img_size,int): img_size = (img_size,img_size)
    for k in keys:
        if not k in checkpoint_model: continue
        pos_embed_checkpoint = checkpoint_model[k]
        embedding_size = pos_embed_checkpoint.shape[-1]
        num_extra_tokens = 0 # no cls token
        # height (== width) for the checkpoint position embedding
        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
        new_size = (img_size[0]//model.patch_embed.patch_size[0],img_size[1]//model.patch_embed.patch_size[1])
        if orig_size != new_size[0] or orig_size != new_size[1]:
            print("Position interpolate %s from %dx%d to %dx%d" % (k, orig_size, orig_size, new_size[0], new_size[1]))
            extra_tokens = pos_embed_checkpoint[:num_extra_tokens,:]
            pos_tokens = pos_embed_checkpoint[num_extra_tokens:,:]
            pos_tokens = pos_tokens.reshape(1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
            pos_tokens = torch.nn.functional.interpolate(pos_tokens, size=(new_size[0], new_size[1]), mode='bicubic', align_corners=False)
            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2).squeeze(0)
            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=0)
            checkpoint_model[k] = new_pos_embed.squeeze(0)

#----------------------------------------------------------
# RoPE2D: RoPE implementation in 2D
#----------------------------------------------------------

try:
    from models.curope import cuRoPE2D
    RoPE2D = cuRoPE2D
except ImportError:
    print('Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead')

    class RoPE2D(torch.nn.Module):
        
        def __init__(self, freq=100.0, F0=1.0):
            super().__init__()
            self.base = freq 
            self.F0 = F0
            self.cache = {}

        def get_cos_sin(self, D, seq_len, device, dtype):
            if (D,seq_len,device,dtype) not in self.cache:
                inv_freq = 1.0 / (self.base ** (torch.arange(0, D, 2).float().to(device) / D))
                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
                freqs = torch.einsum("i,j->ij", t, inv_freq).to(dtype)
                freqs = torch.cat((freqs, freqs), dim=-1)
                cos = freqs.cos() # (Seq, Dim)
                sin = freqs.sin()
                self.cache[D,seq_len,device,dtype] = (cos,sin)
            return self.cache[D,seq_len,device,dtype]
            
        @staticmethod
        def rotate_half(x):
            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
            return torch.cat((-x2, x1), dim=-1)
            
        def apply_rope1d(self, tokens, pos1d, cos, sin):
            assert pos1d.ndim==2
            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]
            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]
            return (tokens * cos) + (self.rotate_half(tokens) * sin)
            
        def forward(self, tokens, positions):
            """
            input:
                * tokens: batch_size x nheads x ntokens x dim
                * positions: batch_size x ntokens x 2 (y and x position of each token)
            output:
                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)
            """
            assert tokens.size(3)%2==0, "number of dimensions should be a multiple of two"
            D = tokens.size(3) // 2
            assert positions.ndim==3 and positions.shape[-1] == 2 # Batch, Seq, 2
            cos, sin = self.get_cos_sin(D, int(positions.max())+1, tokens.device, tokens.dtype)
            # split features into two along the feature dimension, and apply rope1d on each half
            y, x = tokens.chunk(2, dim=-1)
            y = self.apply_rope1d(y, positions[:,:,0], cos, sin)
            x = self.apply_rope1d(x, positions[:,:,1], cos, sin)
            tokens = torch.cat((y, x), dim=-1)
            return tokens