File size: 11,238 Bytes
87608ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
"""Pixel-Space Depth Predictor that preserves the dense image lattice.

Normalized RGB is embedded with a 1x1 projection and processed by shared
CM-PiT trunk blocks before branching into depth and finite-mask predictors.
Linear patch compression is used only within transformer blocks, after which
features are expanded back to per-pixel tokens for dense output heads.
"""

from typing import Iterable, Optional, Tuple

import torch
import torch.nn as nn
from torch.utils.checkpoint import checkpoint

from ..registry import PREDICTORS
from .CM_PiT import CMPiTBlock
from .RoPE import PositionGetter, RotaryPositionEmbedding2D
from .precision import full_precision, reduced_precision


@PREDICTORS.register()
class PixelSpaceDepthPredictor(nn.Module):
    """Pixel-Space Depth Predictor built from cascaded CM-PiT blocks.

    A 1x1 projection first embeds normalized RGB into dense pixel features.
    Shared trunk blocks refine those features, after which independent depth
    and validity branches predict normalized log-depth and finite-depth logits.
    No convolution larger than 1x1 is applied to the pixel representation.
    """

    def __init__(
        self,
        in_channels: int = 3,
        dim_ctx: int = 1024,
        attn_dim: int = 1536,
        ctx_patch_size: int = 14,
        dim_pix: int = 16,
        trunk_patch_size: int = 14,
        depth_patch_size: int = 7,
        mask_patch_size: int = 14,
        num_heads: int = 24,
        trunk_depth: int = 4,
        depth_depth: int = 4,
        mask_depth: int = 2,
        mlp_ratio: float = 4.0,
        qk_norm: bool = True,
        rope_frequency: float = 100.0,
        eps: float = 1e-6,
        gradient_checkpointing: bool = True,
    ) -> None:
        """Construct the shared trunk and two prediction branches.

        Args:
            in_channels: Number of image channels, equal to three for RGB.
            dim_ctx: Global context-token channel count ``C_ctx``.
            attn_dim: Channel count ``D`` after linear patch compression.
            ctx_patch_size: Encoder patch size ``P_ctx`` in image pixels.
            dim_pix: Channel count ``C_pix`` of the dense pixel feature map.
            trunk_patch_size: Attention patch size used by shared trunk blocks.
            depth_patch_size: Attention patch size used by depth blocks.
            mask_patch_size: Attention patch size used by validity-mask blocks.
            num_heads: Number of gated-attention heads.
            trunk_depth: Number of shared CM-PiT blocks.
            depth_depth: Number of depth-branch CM-PiT blocks.
            mask_depth: Number of validity-branch CM-PiT blocks.
            mlp_ratio: SwiGLU expansion ratio inside every block.
            qk_norm: Enable FP32 RMSNorm for attention queries and keys.
            rope_frequency: Base frequency of the shared 2D RoPE module.
            eps: Numerical epsilon for normalization layers.
            gradient_checkpointing: Recompute CM-PiT blocks during backward to
                reduce activation memory.

        Returns:
            ``None``. The complete pixel predictor is registered on the module.
        """
        super().__init__()
        if attn_dim % num_heads != 0:
            raise ValueError(f"attn_dim ({attn_dim}) must be divisible by num_heads ({num_heads})")
        for name, patch_size in {
            "trunk_patch_size": trunk_patch_size,
            "depth_patch_size": depth_patch_size,
            "mask_patch_size": mask_patch_size,
        }.items():
            if patch_size <= 0 or ctx_patch_size % patch_size != 0:
                raise ValueError(
                    f"{name} ({patch_size}) must be positive and divide ctx_patch_size ({ctx_patch_size})"
                )

        self.dim_ctx = dim_ctx
        self.dim_pix = dim_pix
        self.attn_dim = attn_dim
        self.ctx_patch_size = ctx_patch_size
        self.trunk_patch_size = trunk_patch_size
        self.depth_patch_size = depth_patch_size
        self.mask_patch_size = mask_patch_size
        self.gradient_checkpointing = gradient_checkpointing

        self.pos = PositionGetter()
        self.rope = RotaryPositionEmbedding2D(frequency=rope_frequency)
        self.input_proj = nn.Conv2d(in_channels, dim_pix, kernel_size=1, bias=True)

        block_args = dict(
            dim_ctx=dim_ctx,
            ctx_patch_size=ctx_patch_size,
            dim_pix=dim_pix,
            attn_dim=attn_dim,
            num_heads=num_heads,
            mlp_ratio=mlp_ratio,
            qk_norm=qk_norm,
            rope=self.rope,
            eps=eps,
        )
        self.trunk_blocks = nn.ModuleList(
            CMPiTBlock(patch_size=trunk_patch_size, **block_args) for _ in range(trunk_depth)
        )
        self.depth_blocks = nn.ModuleList(
            CMPiTBlock(patch_size=depth_patch_size, **block_args) for _ in range(depth_depth)
        )
        self.mask_blocks = nn.ModuleList(
            CMPiTBlock(patch_size=mask_patch_size, **block_args) for _ in range(mask_depth)
        )
        self.depth_head = nn.Conv2d(dim_pix, 1, kernel_size=1, bias=True)
        self.mask_head = nn.Conv2d(dim_pix, 1, kernel_size=1, bias=True)
        self.reset_parameters()

    def _blocks(self) -> Iterable[CMPiTBlock]:
        """Iterate over every CM-PiT block in execution-independent order.

        Returns:
            Iterable containing shared trunk, depth, and mask blocks. The method
            takes no tensor inputs and is used for parameter initialization.
        """
        return (*self.trunk_blocks, *self.depth_blocks, *self.mask_blocks)

    def reset_parameters(self) -> None:
        """Initialize projections and start adaptive modulation at identity.

        Linear and 1x1 convolution weights use Xavier uniform initialization.
        Normalization scales start at one. The final adaptive-normalization
        projections are zeroed so every CM-PiT residual branch initially has
        zero modulation and zero gate.

        Returns:
            ``None``. Parameters are modified in place.
        """
        def init(module: nn.Module) -> None:
            """Initialize one child module visited by :meth:`nn.Module.apply`.

            Args:
                module: Child ``nn.Module`` to initialize in place.

            Returns:
                ``None``.
            """
            if isinstance(module, (nn.Linear, nn.Conv2d)):
                nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.zeros_(module.bias)
            elif isinstance(module, (nn.LayerNorm, nn.RMSNorm)):
                if module.weight is not None:
                    nn.init.ones_(module.weight)
                if getattr(module, "bias", None) is not None:
                    nn.init.zeros_(module.bias)

        self.apply(init)
        for block in self._blocks():
            nn.init.zeros_(block.ada_norm.proj[-1].weight)
            nn.init.zeros_(block.ada_norm.proj[-1].bias)

    def enable_gradient_checkpointing(self) -> None:
        """Enable activation recomputation for CM-PiT blocks.

        The flag is consulted only while the module is in training mode.

        Returns:
            ``None``. The runtime flag is changed in place.
        """
        self.gradient_checkpointing = True

    def disable_gradient_checkpointing(self) -> None:
        """Disable activation recomputation for CM-PiT blocks.

        Subsequent training forwards retain block activations for backward.

        Returns:
            ``None``. The runtime flag is changed in place.
        """
        self.gradient_checkpointing = False

    def _position(self, batch: int, height: int, width: int, patch_size: int, device: torch.device):
        """Create cached 2D coordinates for one decoder patch grid.

        Args:
            batch: Batch size ``B``.
            height: Dense image-feature height ``H``.
            width: Dense image-feature width ``W``.
            patch_size: Block patch side length ``P``.
            device: Device on which coordinates are allocated.

        Returns:
            Integer position tensor ``[B, (H/P)*(W/P), 2]``.
        """
        return self.pos(batch, height // patch_size, width // patch_size, device=device).to(device)

    def _run(
        self,
        x: torch.Tensor,
        blocks: nn.ModuleList,
        ctx: torch.Tensor,
        pos: torch.Tensor,
    ) -> torch.Tensor:
        """Run a sequence of CM-PiT blocks with optional checkpointing.

        Args:
            x: Dense pixel features ``[B, C_pix, H, W]``.
            blocks: Ordered CM-PiT block collection for one branch.
            ctx: Global context tokens ``[B, N_ctx, C_ctx]``.
            pos: 2D positions ``[B, N, 2]`` matching the blocks' patch grid.

        Returns:
            Refined dense features ``[B, C_pix, H, W]``.
        """
        for block in blocks:
            if self.training and self.gradient_checkpointing:
                x = checkpoint(block, x, ctx, pos, use_reentrant=False)
            else:
                x = block(x, ctx, pos)
        return x

    def forward(
        self,
        image: torch.Tensor,
        ctx: torch.Tensor,
        autocast_dtype: Optional[torch.dtype] = torch.bfloat16,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Predict normalized log-depth and finite-depth logits from RGB.

        Args:
            image: ImageNet-normalized RGB tensor ``[B, 3, H, W]``.
            ctx: Global context tokens ``[B, (H/P_ctx)*(W/P_ctx), C_ctx]``.
            autocast_dtype: Decoder attention dtype. Use ``None`` for FP32,
                ``torch.float16`` for FP16, or ``torch.bfloat16`` for BF16.

        Returns:
            depth: Raw normalized log-depth tensor ``[B, 1, H, W]``.
            mask: Raw finite-depth logit tensor ``[B, 1, H, W]``.
        """
        batch, _, height, width = image.shape
        p_ctx = self.ctx_patch_size
        if height % p_ctx != 0 or width % p_ctx != 0:
            raise ValueError(f"Input resolution ({height}, {width}) must be divisible by {p_ctx}")
        expected = (height // p_ctx) * (width // p_ctx)
        if tuple(ctx.shape) != (batch, expected, self.dim_ctx):
            raise ValueError(
                f"Context shape {tuple(ctx.shape)} does not match ({batch}, {expected}, {self.dim_ctx})"
            )

        with full_precision(image.device):
            pix = self.input_proj(image.float())

        with reduced_precision(image.device, autocast_dtype):
            trunk_pos = self._position(batch, height, width, self.trunk_patch_size, image.device)
            pix = self._run(pix, self.trunk_blocks, ctx, trunk_pos)

            depth_pos = self._position(batch, height, width, self.depth_patch_size, image.device)
            depth_feat = self._run(pix, self.depth_blocks, ctx, depth_pos)

            mask_pos = self._position(batch, height, width, self.mask_patch_size, image.device)
            mask_feat = self._run(pix, self.mask_blocks, ctx, mask_pos)

        with full_precision(image.device):
            depth = self.depth_head(depth_feat.float())
            mask = self.mask_head(mask_feat.float())
        return depth, mask