Upload Trace-Anything (transformers port, original code)
653a5eb verified 19 days ago
29 kB
# ruff: noqa
# fmt: off
"""Trace Anything, vendored verbatim from ByteDance-Seed/TraceAnything @ 54677b5e7bf11510c2e8c917a509988ad379f8eb (trace_anything/): the files below
concatenated, package-internal imports replaced by relative imports of the other vendored files (generated by scripts/model_ports/common/vendor_flat.py;
do not edit by hand). Code license: Apache-2.0."""


# ==============================================================================================================
# trace_anything/layers/blocks.py
# ==============================================================================================================
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# 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.

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


# --------------------------------------------------------
# Main encoder/decoder blocks
# --------------------------------------------------------
# References:
# timm
# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/helpers.py
# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py
# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/mlp.py
# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/patch_embed.py


import collections.abc
from itertools import repeat
import math

import torch
import torch.nn as nn
from torch.nn.functional import scaled_dot_product_attention
from torch.nn.attention import SDPBackend


def _ntuple(n):
    def parse(x):
        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
            return x
        return tuple(repeat(x, n))

    return parse


to_2tuple = _ntuple(2)


def drop_path(
    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
    if drop_prob == 0.0 or not training:
        return x
    keep_prob = 1 - drop_prob
    shape = (x.shape[0],) + (1,) * (
        x.ndim - 1
    )  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
    if keep_prob > 0.0 and scale_by_keep:
        random_tensor.div_(keep_prob)
    return x * random_tensor


class DropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""

    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob
        self.scale_by_keep = scale_by_keep

    def forward(self, x):
        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)

    def extra_repr(self):
        return f"drop_prob={round(self.drop_prob,3):0.3f}"


class Mlp(nn.Module):
    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""

    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        act_layer=nn.GELU,
        bias=True,
        drop=0.0,
    ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        bias = to_2tuple(bias)
        drop_probs = to_2tuple(drop)

        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])
        self.act = act_layer()
        self.drop1 = nn.Dropout(drop_probs[0])
        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])
        self.drop2 = nn.Dropout(drop_probs[1])

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop1(x)
        x = self.fc2(x)
        x = self.drop2(x)
        return x


class Attention(nn.Module):
    def __init__(
        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0,
        attn_mask=None, is_causal=False, attn_implementation="pytorch_naive",
        attn_bias_for_inference_enabled=False,
    ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = head_dim**-0.5
        # use attention biasing to accommodate for longer sequences than during training
        self.attn_bias_for_inference_enabled = attn_bias_for_inference_enabled
        gamma = 1.0
        train_seqlen = 20
        inference_seqlen = 137
        self.attn_bias_scale = head_dim**-0.5 * (gamma * math.log(inference_seqlen) / math.log(train_seqlen))**0.5
        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.dropout_p = attn_drop
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
        self.rope = rope
        self.attn_mask = attn_mask
        self.is_causal = is_causal
        self.attn_implementation = attn_implementation

    def forward(self, x, xpos):
        B, N, C = x.shape

        qkv = (
            self.qkv(x)
            .reshape(B, N, 3, self.num_heads, C // self.num_heads)
            .transpose(1, 3)
        )
        q, k, v = [qkv[:, :, i] for i in range(3)]
        # q,k,v = qkv.unbind(2)  # make torchscript happy (cannot use tensor as tuple)

        if self.rope is not None:
            with torch.autocast(device_type=next(self.parameters()).device.type, dtype=torch.float32):  # FIXME: for some reason Lightning didn't pick up torch.cuda.amp.custom_fwd when using bf16-true
                q = self.rope(q, xpos) if xpos is not None else q
                k = self.rope(k, xpos) if xpos is not None else k

        if not self.training and self.attn_bias_for_inference_enabled:
            scale = self.attn_bias_scale
        else:
            scale = self.scale

        # Important: For the fusion Transformer, we forward through the attention with bfloat16 precision
        # If you are not using this block for the fusion Transformer, you should double check the precision of the input and output
        if self.attn_implementation == "pytorch_naive":
            assert self.attn_mask is None, "attn_mask not supported for pytorch_naive implementation of scaled dot product attention"
            assert self.is_causal is False, "is_causal not supported for pytorch_naive implementation of scaled dot product attention"
            dtype = k.dtype
            with torch.autocast("cuda", dtype=torch.bfloat16):
                x = (q @ k.transpose(-2, -1)) * scale
                x = x.softmax(dim=-1)
                x = self.attn_drop(x)
            if dtype == torch.float32:  # if input was FP32, cast back to FP32
                x = x.to(torch.float32)
            x = (x @ v).transpose(1, 2).reshape(B, N, C)
            x = self.proj(x)
            x = self.proj_drop(x)  
        elif self.attn_implementation == "flash_attention":
            with torch.nn.attention.sdpa_kernel(SDPBackend.FLASH_ATTENTION):
                dtype = k.dtype
                with torch.autocast("cuda", dtype=torch.bfloat16):
                    x = scaled_dot_product_attention(q, k, v, attn_mask=self.attn_mask, dropout_p=self.dropout_p, is_causal=self.is_causal, scale=scale)
                if dtype == torch.float32:  # if input was FP32, cast back to FP32
                    x = x.to(torch.float32)
                x = x.transpose(1, 2).reshape(B, N, C)
                x = self.proj(x)
                x = self.proj_drop(x)
        elif self.attn_implementation == "pytorch_auto":
            with torch.nn.attention.sdpa_kernel([SDPBackend.EFFICIENT_ATTENTION,]):
                dtype = k.dtype
                with torch.autocast("cuda", dtype=torch.bfloat16):
                    x = scaled_dot_product_attention(q, k, v, attn_mask=self.attn_mask, dropout_p=self.dropout_p, is_causal=self.is_causal, scale=scale)
                if dtype == torch.float32:  # if input was FP32, cast back to FP32
                    x = x.to(torch.float32)
                x = x.transpose(1, 2).reshape(B, N, C)
                x = self.proj(x)
                x = self.proj_drop(x)
        else:
            raise ValueError(f"Unknown attn_implementation: {self.attn_implementation}")

        return x


class Block(nn.Module):
    def __init__(
        self,
        dim,
        num_heads,
        mlp_ratio=4.0,
        qkv_bias=False,
        drop=0.0,
        attn_drop=0.0,
        drop_path=0.0,
        act_layer=nn.GELU,
        norm_layer=nn.LayerNorm,
        rope=None,
        attn_implementation="pytorch_naive",
        attn_bias_for_inference_enabled=False,
    ):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim,
            rope=rope,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=drop,
            attn_implementation=attn_implementation,
            attn_bias_for_inference_enabled=attn_bias_for_inference_enabled,
        )
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
        )

    def forward(self, x, xpos):
        x = x + self.drop_path(self.attn(self.norm1(x), xpos))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x


class CrossAttention(nn.Module):
    def __init__(
        self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0, attn_mask=None, is_causal=False, attn_implementation="pytorch_naive"
    ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = head_dim**-0.5

        self.projq = nn.Linear(dim, dim, bias=qkv_bias)
        self.projk = nn.Linear(dim, dim, bias=qkv_bias)
        self.projv = nn.Linear(dim, dim, bias=qkv_bias)
        self.dropout_p = attn_drop
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        self.rope = rope

        self.attn_mask = attn_mask
        self.is_causal = is_causal
        self.attn_implementation = attn_implementation

    def forward(self, query, key, value, qpos, kpos):
        B, Nq, C = query.shape
        Nk = key.shape[1]
        Nv = value.shape[1]

        q = (
            self.projq(query)
            .reshape(B, Nq, self.num_heads, C // self.num_heads)
            .permute(0, 2, 1, 3)
        )
        k = (
            self.projk(key)
            .reshape(B, Nk, self.num_heads, C // self.num_heads)
            .permute(0, 2, 1, 3)
        )
        v = (
            self.projv(value)
            .reshape(B, Nv, self.num_heads, C // self.num_heads)
            .permute(0, 2, 1, 3)
        )

        if self.rope is not None:
            with torch.autocast(device_type=next(self.parameters()).device.type, dtype=torch.float32):  # FIXME: for some reason Lightning didn't pick up torch.cuda.amp.custom_fwd when using bf16-true
                q = self.rope(q, qpos) if qpos is not None else q
                k = self.rope(k, kpos) if kpos is not None else k

        if self.attn_implementation == "pytorch_naive":
            assert self.attn_mask is None, "attn_mask not supported for pytorch_naive implementation of scaled dot product attention"
            assert self.is_causal is False, "is_causal not supported for pytorch_naive implementation of scaled dot product attention"
            attn = (q @ k.transpose(-2, -1)) * self.scale
            attn = attn.softmax(dim=-1)
            attn = self.attn_drop(attn)

            x = (attn @ v).transpose(1, 2).reshape(B, Nq, C)
            x = self.proj(x)
            x = self.proj_drop(x)
        elif self.attn_implementation == "flash_attention":
            with torch.nn.attention.sdpa_kernel(SDPBackend.FLASH_ATTENTION):
                # cast to BF16 to use flash_attention
                q = q.to(torch.bfloat16)
                k = k.to(torch.bfloat16)
                v = v.to(torch.bfloat16)
                x = scaled_dot_product_attention(q, k, v, attn_mask=self.attn_mask, dropout_p=self.dropout_p, is_causal=self.is_causal, scale=self.scale)
                # cast back to FP32
                x = x.to(torch.float32)
                x = x.transpose(1, 2).reshape(B, Nq, C)
                x = self.proj(x)
                x = self.proj_drop(x)
        else:
            raise ValueError(f"Unknown attn_implementation: {self.attn_implementation}")

        return x


class DecoderBlock(nn.Module):
    def __init__(
        self,
        dim,
        num_heads,
        mlp_ratio=4.0,
        qkv_bias=False,
        drop=0.0,
        attn_drop=0.0,
        drop_path=0.0,
        act_layer=nn.GELU,
        norm_layer=nn.LayerNorm,
        norm_mem=True,
        rope=None,
        attn_implementation="pytorch_naive",
    ):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim,
            rope=rope,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=drop,
            attn_implementation=attn_implementation,
        )
        self.cross_attn = CrossAttention(
            dim,
            rope=rope,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=drop,
            attn_implementation=attn_implementation,
        )
        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
        self.norm2 = norm_layer(dim)
        self.norm3 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
        )
        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()

    def forward(self, x, y, xpos, ypos):
        x = x + self.drop_path(self.attn(self.norm1(x), xpos))
        y_ = self.norm_y(y)
        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos))
        x = x + self.drop_path(self.mlp(self.norm3(x)))
        return x, y


# patch embedding
class PositionGetter(object):
    """return positions of patches"""

    def __init__(self):
        self.cache_positions = {}

    def __call__(self, b, h, w, device):
        if not (h, w) in self.cache_positions:
            x = torch.arange(w, device=device)
            y = torch.arange(h, device=device)
            self.cache_positions[h, w] = torch.cartesian_prod(y, x)  # (h, w, 2)
        pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone()
        return pos


class PatchEmbed(nn.Module):
    """just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed"""

    def __init__(
        self,
        img_size=224,
        patch_size=16,
        in_chans=3,
        embed_dim=768,
        norm_layer=None,
        flatten=True,
    ):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        self.img_size = img_size
        self.patch_size = patch_size
        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
        self.num_patches = self.grid_size[0] * self.grid_size[1]
        self.flatten = flatten

        self.proj = nn.Conv2d(
            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size
        )
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

        self.position_getter = PositionGetter()

    def forward(self, x):
        B, C, H, W = x.shape
        torch._assert(
            H == self.img_size[0],
            f"Input image height ({H}) doesn't match model ({self.img_size[0]}).",
        )
        torch._assert(
            W == self.img_size[1],
            f"Input image width ({W}) doesn't match model ({self.img_size[1]}).",
        )
        x = self.proj(x)
        pos = self.position_getter(B, x.size(2), x.size(3), x.device)
        if self.flatten:
            x = x.flatten(2).transpose(1, 2).contiguous()  # BCHW -> BNC
        x = self.norm(x)
        return x, pos

    def _init_weights(self):
        w = self.proj.weight.data
        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))


# ==============================================================================================================
# trace_anything/layers/pos_embed.py
# ==============================================================================================================
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# 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.

# 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: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [n_cls_token+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    grid_h = np.arange(grid_size, dtype=np.float32)
    grid_w = np.arange(grid_size, 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, grid_size])
    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 / 2.0
    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:
# 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):
    if "pos_embed" in checkpoint_model:
        pos_embed_checkpoint = checkpoint_model["pos_embed"]
        embedding_size = pos_embed_checkpoint.shape[-1]
        num_patches = model.patch_embed.num_patches
        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
        # height (== width) for the checkpoint position embedding
        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
        # height (== width) for the new position embedding
        new_size = int(num_patches**0.5)
        # class_token and dist_token are kept unchanged
        if orig_size != new_size:
            print(
                "Position interpolate from %dx%d to %dx%d"
                % (orig_size, orig_size, new_size, new_size)
            )
            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
            # only the position tokens are interpolated
            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, new_size),
                mode="bicubic",
                align_corners=False,
            )
            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
            checkpoint_model["pos_embed"] = new_pos_embed


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

try:
    from fast3r.croco.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


# ==============================================================================================================
# trace_anything/layers/patch_embed.py
# ==============================================================================================================
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# 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.

# 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


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
        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)
        pos = self.position_getter(B, x.size(2), x.size(3), x.device)
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
        x = self.norm(x)
        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

        # linear projection, transposed if necessary
        if is_landscape.any():
            new_landscape_content = self.proj(img[is_landscape])
            new_landscape_content = new_landscape_content.permute(0, 2, 3, 1).flatten(1, 2)
        if is_portrait.any():
            new_protrait_content = self.proj(img[is_portrait].swapaxes(-1, -2))
            new_protrait_content = new_protrait_content.permute(0, 2, 3, 1).flatten(1, 2)

        # allocate space for result and set the content
        x = img.new_empty((B, n_tokens, self.embed_dim), dtype=next(self.named_parameters())[1].dtype)  # dynamically set dtype based on the current precision
        if is_landscape.any():
            x[is_landscape] = new_landscape_content
        if is_portrait.any():
            x[is_portrait] = new_protrait_content

        # allocate space for result and set the content
        pos = img.new_empty((B, n_tokens, 2), dtype=torch.int64)
        if is_landscape.any():
            pos[is_landscape] = self.position_getter(1, H, W, pos.device).expand(is_landscape.sum(), -1, -1)
        if is_portrait.any():
            pos[is_portrait] = self.position_getter(1, W, H, pos.device).expand(is_portrait.sum(), -1, -1)

        x = self.norm(x)
        return x, pos