1

trellis/models/structured_latent_vae/decoder_gs.py

@@ -0,0 +1,131 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from ...modules import sparse as sp
+from ...utils.random_utils import hammersley_sequence
+from .base import SparseTransformerBase
+from ...representations import Gaussian
+from ..sparse_elastic_mixin import SparseTransformerElasticMixin
+
+
+class SLatGaussianDecoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+        representation_config: dict = None,
+    ):
+        super().__init__(
+            in_channels=latent_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.rep_config = representation_config
+        self._calc_layout()
+        self.out_layer = sp.SparseLinear(model_channels, self.out_channels)
+        self._build_perturbation()
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def _build_perturbation(self) -> None:
+        perturbation = [hammersley_sequence(3, i, self.rep_config['num_gaussians']) for i in range(self.rep_config['num_gaussians'])]
+        perturbation = torch.tensor(perturbation).float() * 2 - 1
+        perturbation = perturbation / self.rep_config['voxel_size']
+        perturbation = torch.atanh(perturbation).to(self.device)
+        self.register_buffer('offset_perturbation', perturbation)
+
+    def _calc_layout(self) -> None:
+        self.layout = {
+            '_xyz' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
+            '_features_dc' : {'shape': (self.rep_config['num_gaussians'], 1, 3), 'size': self.rep_config['num_gaussians'] * 3},
+            '_scaling' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
+            '_rotation' : {'shape': (self.rep_config['num_gaussians'], 4), 'size': self.rep_config['num_gaussians'] * 4},
+            '_opacity' : {'shape': (self.rep_config['num_gaussians'], 1), 'size': self.rep_config['num_gaussians']},
+        }
+        start = 0
+        for k, v in self.layout.items():
+            v['range'] = (start, start + v['size'])
+            start += v['size']
+        self.out_channels = start
+    
+    def to_representation(self, x: sp.SparseTensor) -> List[Gaussian]:
+        """
+        Convert a batch of network outputs to 3D representations.
+
+        Args:
+            x: The [N x * x C] sparse tensor output by the network.
+
+        Returns:
+            list of representations
+        """
+        ret = []
+        for i in range(x.shape[0]):
+            representation = Gaussian(
+                sh_degree=0,
+                aabb=[-0.5, -0.5, -0.5, 1.0, 1.0, 1.0],
+                mininum_kernel_size = self.rep_config['3d_filter_kernel_size'],
+                scaling_bias = self.rep_config['scaling_bias'],
+                opacity_bias = self.rep_config['opacity_bias'],
+                scaling_activation = self.rep_config['scaling_activation']
+            )
+            xyz = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
+            for k, v in self.layout.items():
+                if k == '_xyz':
+                    offset = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape'])
+                    offset = offset * self.rep_config['lr'][k]
+                    if self.rep_config['perturb_offset']:
+                        offset = offset + self.offset_perturbation
+                    offset = torch.tanh(offset) / self.resolution * 0.5 * self.rep_config['voxel_size']
+                    _xyz = xyz.unsqueeze(1) + offset
+                    setattr(representation, k, _xyz.flatten(0, 1))
+                else:
+                    feats = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']).flatten(0, 1)
+                    feats = feats * self.rep_config['lr'][k]
+                    setattr(representation, k, feats)
+            ret.append(representation)
+        return ret
+
+    def forward(self, x: sp.SparseTensor) -> List[Gaussian]:
+        h = super().forward(x)
+        h = h.type(x.dtype)
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h)
+        return self.to_representation(h)
+    
+
+class ElasticSLatGaussianDecoder(SparseTransformerElasticMixin, SLatGaussianDecoder):
+    """
+    Slat VAE Gaussian decoder with elastic memory management.
+    Used for training with low VRAM.
+    """
+    pass

trellis/models/structured_latent_vae/decoder_mesh.py

@@ -0,0 +1,176 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ...modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
+from ...modules import sparse as sp
+from .base import SparseTransformerBase
+from ...representations import MeshExtractResult
+from ...representations.mesh import SparseFeatures2Mesh
+from ..sparse_elastic_mixin import SparseTransformerElasticMixin
+
+
+class SparseSubdivideBlock3d(nn.Module):
+    """
+    A 3D subdivide block that can subdivide the sparse tensor.
+
+    Args:
+        channels: channels in the inputs and outputs.
+        out_channels: if specified, the number of output channels.
+        num_groups: the number of groups for the group norm.
+    """
+    def __init__(
+        self,
+        channels: int,
+        resolution: int,
+        out_channels: Optional[int] = None,
+        num_groups: int = 32
+    ):
+        super().__init__()
+        self.channels = channels
+        self.resolution = resolution
+        self.out_resolution = resolution * 2
+        self.out_channels = out_channels or channels
+
+        self.act_layers = nn.Sequential(
+            sp.SparseGroupNorm32(num_groups, channels),
+            sp.SparseSiLU()
+        )
+        
+        self.sub = sp.SparseSubdivide()
+        
+        self.out_layers = nn.Sequential(
+            sp.SparseConv3d(channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}"),
+            sp.SparseGroupNorm32(num_groups, self.out_channels),
+            sp.SparseSiLU(),
+            zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}")),
+        )
+        
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        else:
+            self.skip_connection = sp.SparseConv3d(channels, self.out_channels, 1, indice_key=f"res_{self.out_resolution}")
+        
+    def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+
+        Args:
+            x: an [N x C x ...] Tensor of features.
+        Returns:
+            an [N x C x ...] Tensor of outputs.
+        """
+        h = self.act_layers(x)
+        h = self.sub(h)
+        x = self.sub(x)
+        h = self.out_layers(h)
+        h = h + self.skip_connection(x)
+        return h
+
+
+class SLatMeshDecoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+        representation_config: dict = None,
+    ):
+        super().__init__(
+            in_channels=latent_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.rep_config = representation_config
+        self.mesh_extractor = SparseFeatures2Mesh(res=self.resolution*4, use_color=self.rep_config.get('use_color', False))
+        self.out_channels = self.mesh_extractor.feats_channels
+        self.upsample = nn.ModuleList([
+            SparseSubdivideBlock3d(
+                channels=model_channels,
+                resolution=resolution,
+                out_channels=model_channels // 4
+            ),
+            SparseSubdivideBlock3d(
+                channels=model_channels // 4,
+                resolution=resolution * 2,
+                out_channels=model_channels // 8
+            )
+        ])
+        self.out_layer = sp.SparseLinear(model_channels // 8, self.out_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        super().convert_to_fp16()
+        self.upsample.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        super().convert_to_fp32()
+        self.upsample.apply(convert_module_to_f32)  
+    
+    def to_representation(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
+        """
+        Convert a batch of network outputs to 3D representations.
+
+        Args:
+            x: The [N x * x C] sparse tensor output by the network.
+
+        Returns:
+            list of representations
+        """
+        ret = []
+        for i in range(x.shape[0]):
+            mesh = self.mesh_extractor(x[i], training=self.training)
+            ret.append(mesh)
+        return ret
+
+    def forward(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
+        h = super().forward(x)
+        for block in self.upsample:
+            h = block(h)
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+        return self.to_representation(h)
+    
+
+class ElasticSLatMeshDecoder(SparseTransformerElasticMixin, SLatMeshDecoder):
+    """
+    Slat VAE Mesh decoder with elastic memory management.
+    Used for training with low VRAM.
+    """
+    pass

trellis/models/structured_latent_vae/decoder_rf.py

@@ -0,0 +1,113 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ...modules import sparse as sp
+from .base import SparseTransformerBase
+from ...representations import Strivec
+from ..sparse_elastic_mixin import SparseTransformerElasticMixin
+
+
+class SLatRadianceFieldDecoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+        representation_config: dict = None,
+    ):
+        super().__init__(
+            in_channels=latent_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.rep_config = representation_config
+        self._calc_layout()
+        self.out_layer = sp.SparseLinear(model_channels, self.out_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def _calc_layout(self) -> None:
+        self.layout = {
+            'trivec': {'shape': (self.rep_config['rank'], 3, self.rep_config['dim']), 'size': self.rep_config['rank'] * 3 * self.rep_config['dim']},
+            'density': {'shape': (self.rep_config['rank'],), 'size': self.rep_config['rank']},
+            'features_dc': {'shape': (self.rep_config['rank'], 1, 3), 'size': self.rep_config['rank'] * 3},
+        }
+        start = 0
+        for k, v in self.layout.items():
+            v['range'] = (start, start + v['size'])
+            start += v['size']
+        self.out_channels = start    
+    
+    def to_representation(self, x: sp.SparseTensor) -> List[Strivec]:
+        """
+        Convert a batch of network outputs to 3D representations.
+
+        Args:
+            x: The [N x * x C] sparse tensor output by the network.
+
+        Returns:
+            list of representations
+        """
+        ret = []
+        for i in range(x.shape[0]):
+            representation = Strivec(
+                sh_degree=0,
+                resolution=self.resolution,
+                aabb=[-0.5, -0.5, -0.5, 1, 1, 1],
+                rank=self.rep_config['rank'],
+                dim=self.rep_config['dim'],
+                device='cuda',
+            )
+            representation.density_shift = 0.0
+            representation.position = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
+            representation.depth = torch.full((representation.position.shape[0], 1), int(np.log2(self.resolution)), dtype=torch.uint8, device='cuda')
+            for k, v in self.layout.items():
+                setattr(representation, k, x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']))
+            representation.trivec = representation.trivec + 1
+            ret.append(representation)
+        return ret
+
+    def forward(self, x: sp.SparseTensor) -> List[Strivec]:
+        h = super().forward(x)
+        h = h.type(x.dtype)
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h)
+        return self.to_representation(h)
+
+
+class ElasticSLatRadianceFieldDecoder(SparseTransformerElasticMixin, SLatRadianceFieldDecoder):
+    """
+    Slat VAE Radiance Field Decoder with elastic memory management.
+    Used for training with low VRAM.
+    """
+    pass

trellis/models/structured_latent_vae/encoder.py

@@ -0,0 +1,80 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from ...modules import sparse as sp
+from .base import SparseTransformerBase
+from ..sparse_elastic_mixin import SparseTransformerElasticMixin
+
+
+class SLatEncoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.out_layer = sp.SparseLinear(model_channels, 2 * latent_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def forward(self, x: sp.SparseTensor, sample_posterior=True, return_raw=False):
+        h = super().forward(x)
+        h = h.type(x.dtype)
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h)
+        
+        # Sample from the posterior distribution
+        mean, logvar = h.feats.chunk(2, dim=-1)
+        if sample_posterior:
+            std = torch.exp(0.5 * logvar)
+            z = mean + std * torch.randn_like(std)
+        else:
+            z = mean
+        z = h.replace(z)
+            
+        if return_raw:
+            return z, mean, logvar
+        else:
+            return z
+        
+
+class ElasticSLatEncoder(SparseTransformerElasticMixin, SLatEncoder):
+    """
+    SLat VAE encoder with elastic memory management.
+    Used for training with low VRAM.
+    """