linhaotong
update
b9f87ab
# flake8: noqa E501
# Copyright (c) 2025 ByteDance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Dict as TyDict
from typing import List, Sequence, Tuple
import torch
import torch.nn as nn
from addict import Dict
from einops import rearrange
from depth_anything_3.model.utils.head_utils import (
Permute,
create_uv_grid,
custom_interpolate,
position_grid_to_embed,
)
class DPT(nn.Module):
"""
DPT for dense prediction (main head + optional sky head, sky always 1 channel).
Returns:
- Main head:
* If output_dim>1: { head_name, f"{head_name}_conf" }
* If output_dim==1: { head_name }
- Sky head (if use_sky_head=True): { sky_name } # [B, S, 1, H/down_ratio, W/down_ratio]
"""
def __init__(
self,
dim_in: int,
*,
patch_size: int = 14,
output_dim: int = 1,
activation: str = "exp",
conf_activation: str = "expp1",
features: int = 256,
out_channels: Sequence[int] = (256, 512, 1024, 1024),
pos_embed: bool = False,
down_ratio: int = 1,
head_name: str = "depth",
# ---- sky head (fixed 1 channel) ----
use_sky_head: bool = True,
sky_name: str = "sky",
sky_activation: str = "relu", # 'sigmoid' / 'relu' / 'linear'
use_ln_for_heads: bool = False, # If needed, apply LayerNorm on intermediate features of both heads
norm_type: str = "idt", # use to match legacy GS-DPT head, "idt" / "layer"
fusion_block_inplace: bool = False,
) -> None:
super().__init__()
# -------------------- configuration --------------------
self.patch_size = patch_size
self.activation = activation
self.conf_activation = conf_activation
self.pos_embed = pos_embed
self.down_ratio = down_ratio
# Names
self.head_main = head_name
self.sky_name = sky_name
# Main head: output dimension and confidence switch
self.out_dim = output_dim
self.has_conf = output_dim > 1
# Sky head parameters (always 1 channel)
self.use_sky_head = use_sky_head
self.sky_activation = sky_activation
# Fixed 4 intermediate outputs
self.intermediate_layer_idx: Tuple[int, int, int, int] = (0, 1, 2, 3)
# -------------------- token pre-norm + per-stage projection --------------------
if norm_type == "layer":
self.norm = nn.LayerNorm(dim_in)
elif norm_type == "idt":
self.norm = nn.Identity()
else:
raise Exception(f"Unknown norm_type {norm_type}, should be 'layer' or 'idt'.")
self.projects = nn.ModuleList(
[nn.Conv2d(dim_in, oc, kernel_size=1, stride=1, padding=0) for oc in out_channels]
)
# -------------------- Spatial re-size (align to common scale before fusion) --------------------
# Design consistent with original: relative to patch grid (x4, x2, x1, /2)
self.resize_layers = nn.ModuleList(
[
nn.ConvTranspose2d(
out_channels[0], out_channels[0], kernel_size=4, stride=4, padding=0
),
nn.ConvTranspose2d(
out_channels[1], out_channels[1], kernel_size=2, stride=2, padding=0
),
nn.Identity(),
nn.Conv2d(out_channels[3], out_channels[3], kernel_size=3, stride=2, padding=1),
]
)
# -------------------- scratch: stage adapters + main fusion chain --------------------
self.scratch = _make_scratch(list(out_channels), features, expand=False)
# Main fusion chain
self.scratch.refinenet1 = _make_fusion_block(features, inplace=fusion_block_inplace)
self.scratch.refinenet2 = _make_fusion_block(features, inplace=fusion_block_inplace)
self.scratch.refinenet3 = _make_fusion_block(features, inplace=fusion_block_inplace)
self.scratch.refinenet4 = _make_fusion_block(
features, has_residual=False, inplace=fusion_block_inplace
)
# Heads (shared neck1; then split into two heads)
head_features_1 = features
head_features_2 = 32
self.scratch.output_conv1 = nn.Conv2d(
head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1
)
ln_seq = (
[Permute((0, 2, 3, 1)), nn.LayerNorm(head_features_2), Permute((0, 3, 1, 2))]
if use_ln_for_heads
else []
)
# Main head
self.scratch.output_conv2 = nn.Sequential(
nn.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1),
*ln_seq,
nn.ReLU(inplace=True),
nn.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0),
)
# Sky head (fixed 1 channel)
if self.use_sky_head:
self.scratch.sky_output_conv2 = nn.Sequential(
nn.Conv2d(
head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1
),
*ln_seq,
nn.ReLU(inplace=True),
nn.Conv2d(head_features_2, 1, kernel_size=1, stride=1, padding=0),
)
# -------------------------------------------------------------------------
# Public forward (supports frame chunking to save memory)
# -------------------------------------------------------------------------
def forward(
self,
feats: List[torch.Tensor],
H: int,
W: int,
patch_start_idx: int,
chunk_size: int = 8,
**kwargs,
) -> Dict:
"""
Args:
feats: List of 4 entries, each entry is a tensor like [B, S, T, C] (or the 0th element of tuple/list is that tensor).
H, W: Original image dimensions
patch_start_idx: Starting index of patch tokens in sequence (for cropping non-patch tokens)
chunk_size: Chunk size along time dimension S
Returns:
Dict[str, Tensor]
"""
B, S, N, C = feats[0][0].shape
feats = [feat[0].reshape(B * S, N, C) for feat in feats]
# update image info, used by the GS-DPT head
extra_kwargs = {}
if "images" in kwargs:
extra_kwargs.update({"images": rearrange(kwargs["images"], "B S ... -> (B S) ...")})
if chunk_size is None or chunk_size >= S:
out_dict = self._forward_impl(feats, H, W, patch_start_idx, **extra_kwargs)
out_dict = {k: v.view(B, S, *v.shape[1:]) for k, v in out_dict.items()}
return Dict(out_dict)
out_dicts: List[TyDict[str, torch.Tensor]] = []
for s0 in range(0, S, chunk_size):
s1 = min(s0 + chunk_size, S)
kw = {}
if "images" in extra_kwargs:
kw.update({"images": extra_kwargs["images"][s0:s1]})
out_dicts.append(
self._forward_impl([f[s0:s1] for f in feats], H, W, patch_start_idx, **kw)
)
out_dict = {k: torch.cat([od[k] for od in out_dicts], dim=0) for k in out_dicts[0].keys()}
out_dict = {k: v.view(B, S, *v.shape[1:]) for k, v in out_dict.items()}
return Dict(out_dict)
# -------------------------------------------------------------------------
# Internal forward (single chunk)
# -------------------------------------------------------------------------
def _forward_impl(
self,
feats: List[torch.Tensor],
H: int,
W: int,
patch_start_idx: int,
) -> TyDict[str, torch.Tensor]:
B, _, C = feats[0].shape
ph, pw = H // self.patch_size, W // self.patch_size
resized_feats = []
for stage_idx, take_idx in enumerate(self.intermediate_layer_idx):
x = feats[take_idx][:, patch_start_idx:] # [B*S, N_patch, C]
x = self.norm(x)
x = x.permute(0, 2, 1).reshape(B, C, ph, pw) # [B*S, C, ph, pw]
x = self.projects[stage_idx](x)
if self.pos_embed:
x = self._add_pos_embed(x, W, H)
x = self.resize_layers[stage_idx](x) # Align scale
resized_feats.append(x)
# 2) Fusion pyramid (main branch only)
fused = self._fuse(resized_feats)
# 3) Upsample to target resolution, optionally add position encoding again
h_out = int(ph * self.patch_size / self.down_ratio)
w_out = int(pw * self.patch_size / self.down_ratio)
fused = self.scratch.output_conv1(fused)
fused = custom_interpolate(fused, (h_out, w_out), mode="bilinear", align_corners=True)
if self.pos_embed:
fused = self._add_pos_embed(fused, W, H)
# 4) Shared neck1
feat = fused
# 5) Main head: logits -> activation
main_logits = self.scratch.output_conv2(feat)
outs: TyDict[str, torch.Tensor] = {}
if self.has_conf:
fmap = main_logits.permute(0, 2, 3, 1)
pred = self._apply_activation_single(fmap[..., :-1], self.activation)
conf = self._apply_activation_single(fmap[..., -1], self.conf_activation)
outs[self.head_main] = pred.squeeze(1)
outs[f"{self.head_main}_conf"] = conf.squeeze(1)
else:
outs[self.head_main] = self._apply_activation_single(
main_logits, self.activation
).squeeze(1)
# 6) Sky head (fixed 1 channel)
if self.use_sky_head:
sky_logits = self.scratch.sky_output_conv2(feat)
outs[self.sky_name] = self._apply_sky_activation(sky_logits).squeeze(1)
return outs
# -------------------------------------------------------------------------
# Subroutines
# -------------------------------------------------------------------------
def _fuse(self, feats: List[torch.Tensor]) -> torch.Tensor:
"""
4-layer top-down fusion, returns finest scale features (after fusion, before neck1).
"""
l1, l2, l3, l4 = feats
l1_rn = self.scratch.layer1_rn(l1)
l2_rn = self.scratch.layer2_rn(l2)
l3_rn = self.scratch.layer3_rn(l3)
l4_rn = self.scratch.layer4_rn(l4)
# 4 -> 3 -> 2 -> 1
out = self.scratch.refinenet4(l4_rn, size=l3_rn.shape[2:])
out = self.scratch.refinenet3(out, l3_rn, size=l2_rn.shape[2:])
out = self.scratch.refinenet2(out, l2_rn, size=l1_rn.shape[2:])
out = self.scratch.refinenet1(out, l1_rn)
return out
def _apply_activation_single(
self, x: torch.Tensor, activation: str = "linear"
) -> torch.Tensor:
"""
Apply activation to single channel output, maintaining semantic consistency with value branch in multi-channel case.
Supports: exp / relu / sigmoid / softplus / tanh / linear / expp1
"""
act = activation.lower() if isinstance(activation, str) else activation
if act == "exp":
return torch.exp(x)
if act == "expp1":
return torch.exp(x) + 1
if act == "expm1":
return torch.expm1(x)
if act == "relu":
return torch.relu(x)
if act == "sigmoid":
return torch.sigmoid(x)
if act == "softplus":
return torch.nn.functional.softplus(x)
if act == "tanh":
return torch.tanh(x)
# Default linear
return x
def _apply_sky_activation(self, x: torch.Tensor) -> torch.Tensor:
"""
Sky head activation (fixed 1 channel):
* 'sigmoid' -> Sigmoid probability map
* 'relu' -> ReLU positive domain output
* 'linear' -> Original value (logits)
"""
act = (
self.sky_activation.lower()
if isinstance(self.sky_activation, str)
else self.sky_activation
)
if act == "sigmoid":
return torch.sigmoid(x)
if act == "relu":
return torch.relu(x)
# 'linear'
return x
def _add_pos_embed(self, x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor:
"""Simple UV position encoding directly added to feature map."""
pw, ph = x.shape[-1], x.shape[-2]
pe = create_uv_grid(pw, ph, aspect_ratio=W / H, dtype=x.dtype, device=x.device)
pe = position_grid_to_embed(pe, x.shape[1]) * ratio
pe = pe.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1)
return x + pe
# -----------------------------------------------------------------------------
# Building blocks (preserved, consistent with original)
# -----------------------------------------------------------------------------
def _make_fusion_block(
features: int,
size: Tuple[int, int] = None,
has_residual: bool = True,
groups: int = 1,
inplace: bool = False,
) -> nn.Module:
return FeatureFusionBlock(
features=features,
activation=nn.ReLU(inplace=inplace),
deconv=False,
bn=False,
expand=False,
align_corners=True,
size=size,
has_residual=has_residual,
groups=groups,
)
def _make_scratch(
in_shape: List[int], out_shape: int, groups: int = 1, expand: bool = False
) -> nn.Module:
scratch = nn.Module()
# Optional expansion by stage
c1 = out_shape
c2 = out_shape * (2 if expand else 1)
c3 = out_shape * (4 if expand else 1)
c4 = out_shape * (8 if expand else 1)
scratch.layer1_rn = nn.Conv2d(in_shape[0], c1, 3, 1, 1, bias=False, groups=groups)
scratch.layer2_rn = nn.Conv2d(in_shape[1], c2, 3, 1, 1, bias=False, groups=groups)
scratch.layer3_rn = nn.Conv2d(in_shape[2], c3, 3, 1, 1, bias=False, groups=groups)
scratch.layer4_rn = nn.Conv2d(in_shape[3], c4, 3, 1, 1, bias=False, groups=groups)
return scratch
class ResidualConvUnit(nn.Module):
"""Lightweight residual convolution block for fusion"""
def __init__(self, features: int, activation: nn.Module, bn: bool, groups: int = 1) -> None:
super().__init__()
self.bn = bn
self.groups = groups
self.conv1 = nn.Conv2d(features, features, 3, 1, 1, bias=True, groups=groups)
self.conv2 = nn.Conv2d(features, features, 3, 1, 1, bias=True, groups=groups)
self.norm1 = None
self.norm2 = None
self.activation = activation
self.skip_add = nn.quantized.FloatFunctional()
def forward(self, x: torch.Tensor) -> torch.Tensor: # type: ignore[override]
out = self.activation(x)
out = self.conv1(out)
if self.norm1 is not None:
out = self.norm1(out)
out = self.activation(out)
out = self.conv2(out)
if self.norm2 is not None:
out = self.norm2(out)
return self.skip_add.add(out, x)
class FeatureFusionBlock(nn.Module):
"""Top-down fusion block: (optional) residual merge + upsampling + 1x1 contraction"""
def __init__(
self,
features: int,
activation: nn.Module,
deconv: bool = False,
bn: bool = False,
expand: bool = False,
align_corners: bool = True,
size: Tuple[int, int] = None,
has_residual: bool = True,
groups: int = 1,
) -> None:
super().__init__()
self.align_corners = align_corners
self.size = size
self.has_residual = has_residual
self.resConfUnit1 = (
ResidualConvUnit(features, activation, bn, groups=groups) if has_residual else None
)
self.resConfUnit2 = ResidualConvUnit(features, activation, bn, groups=groups)
out_features = (features // 2) if expand else features
self.out_conv = nn.Conv2d(features, out_features, 1, 1, 0, bias=True, groups=groups)
self.skip_add = nn.quantized.FloatFunctional()
def forward(self, *xs: torch.Tensor, size: Tuple[int, int] = None) -> torch.Tensor: # type: ignore[override]
"""
xs:
- xs[0]: Top branch input
- xs[1]: Lateral input (can do residual addition with top branch)
"""
y = xs[0]
if self.has_residual and len(xs) > 1 and self.resConfUnit1 is not None:
y = self.skip_add.add(y, self.resConfUnit1(xs[1]))
y = self.resConfUnit2(y)
# Upsampling
if (size is None) and (self.size is None):
up_kwargs = {"scale_factor": 2}
elif size is None:
up_kwargs = {"size": self.size}
else:
up_kwargs = {"size": size}
y = custom_interpolate(y, **up_kwargs, mode="bilinear", align_corners=self.align_corners)
y = self.out_conv(y)
return y