1

trellis/models/sparse_elastic_mixin.py

@@ -0,0 +1,24 @@
+from contextlib import contextmanager
+from typing import *
+import math
+from ..modules import sparse as sp
+from ..utils.elastic_utils import ElasticModuleMixin
+
+
+class SparseTransformerElasticMixin(ElasticModuleMixin):
+    def _get_input_size(self, x: sp.SparseTensor, *args, **kwargs):
+        return x.feats.shape[0]
+    
+    @contextmanager
+    def with_mem_ratio(self, mem_ratio=1.0):
+        if mem_ratio == 1.0:
+            yield 1.0
+            return
+        num_blocks = len(self.blocks)
+        num_checkpoint_blocks = min(math.ceil((1 - mem_ratio) * num_blocks) + 1, num_blocks)
+        exact_mem_ratio = 1 - (num_checkpoint_blocks - 1) / num_blocks
+        for i in range(num_blocks):
+            self.blocks[i].use_checkpoint = i < num_checkpoint_blocks
+        yield exact_mem_ratio
+        for i in range(num_blocks):
+            self.blocks[i].use_checkpoint = False

trellis/models/sparse_structure_flow.py

@@ -0,0 +1,200 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ..modules.utils import convert_module_to_f16, convert_module_to_f32
+from ..modules.transformer import AbsolutePositionEmbedder, ModulatedTransformerCrossBlock
+from ..modules.spatial import patchify, unpatchify
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        """
+        Create sinusoidal timestep embeddings.
+
+        Args:
+            t: a 1-D Tensor of N indices, one per batch element.
+                These may be fractional.
+            dim: the dimension of the output.
+            max_period: controls the minimum frequency of the embeddings.
+
+        Returns:
+            an (N, D) Tensor of positional embeddings.
+        """
+        # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+        half = dim // 2
+        freqs = torch.exp(
+            -np.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=t.device)
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class SparseStructureFlowModel(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        model_channels: int,
+        cond_channels: int,
+        out_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        patch_size: int = 2,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        share_mod: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+    ):
+        super().__init__()
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.cond_channels = cond_channels
+        self.out_channels = out_channels
+        self.num_blocks = num_blocks
+        self.num_heads = num_heads or model_channels // num_head_channels
+        self.mlp_ratio = mlp_ratio
+        self.patch_size = patch_size
+        self.pe_mode = pe_mode
+        self.use_fp16 = use_fp16
+        self.use_checkpoint = use_checkpoint
+        self.share_mod = share_mod
+        self.qk_rms_norm = qk_rms_norm
+        self.qk_rms_norm_cross = qk_rms_norm_cross
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.t_embedder = TimestepEmbedder(model_channels)
+        if share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(model_channels, 6 * model_channels, bias=True)
+            )
+
+        if pe_mode == "ape":
+            pos_embedder = AbsolutePositionEmbedder(model_channels, 3)
+            coords = torch.meshgrid(*[torch.arange(res, device=self.device) for res in [resolution // patch_size] * 3], indexing='ij')
+            coords = torch.stack(coords, dim=-1).reshape(-1, 3)
+            pos_emb = pos_embedder(coords)
+            self.register_buffer("pos_emb", pos_emb)
+
+        self.input_layer = nn.Linear(in_channels * patch_size**3, model_channels)
+            
+        self.blocks = nn.ModuleList([
+            ModulatedTransformerCrossBlock(
+                model_channels,
+                cond_channels,
+                num_heads=self.num_heads,
+                mlp_ratio=self.mlp_ratio,
+                attn_mode='full',
+                use_checkpoint=self.use_checkpoint,
+                use_rope=(pe_mode == "rope"),
+                share_mod=share_mod,
+                qk_rms_norm=self.qk_rms_norm,
+                qk_rms_norm_cross=self.qk_rms_norm_cross,
+            )
+            for _ in range(num_blocks)
+        ])
+
+        self.out_layer = nn.Linear(model_channels, out_channels * patch_size**3)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.blocks.apply(convert_module_to_f32)
+
+    def initialize_weights(self) -> None:
+        # Initialize transformer layers:
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+        self.apply(_basic_init)
+
+        # Initialize timestep embedding MLP:
+        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+        # Zero-out adaLN modulation layers in DiT blocks:
+        if self.share_mod:
+            nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
+        else:
+            for block in self.blocks:
+                nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+                nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
+        assert [*x.shape] == [x.shape[0], self.in_channels, *[self.resolution] * 3], \
+                f"Input shape mismatch, got {x.shape}, expected {[x.shape[0], self.in_channels, *[self.resolution] * 3]}"
+
+        h = patchify(x, self.patch_size)
+        h = h.view(*h.shape[:2], -1).permute(0, 2, 1).contiguous()
+
+        h = self.input_layer(h)
+        h = h + self.pos_emb[None]
+        t_emb = self.t_embedder(t)
+        if self.share_mod:
+            t_emb = self.adaLN_modulation(t_emb)
+        t_emb = t_emb.type(self.dtype)
+        h = h.type(self.dtype)
+        cond = cond.type(self.dtype)
+        for block in self.blocks:
+            h = block(h, t_emb, cond)
+        h = h.type(x.dtype)
+        h = F.layer_norm(h, h.shape[-1:])
+        h = self.out_layer(h)
+
+        h = h.permute(0, 2, 1).view(h.shape[0], h.shape[2], *[self.resolution // self.patch_size] * 3)
+        h = unpatchify(h, self.patch_size).contiguous()
+
+        return h

trellis/models/sparse_structure_vae.py

@@ -0,0 +1,306 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from ..modules.norm import GroupNorm32, ChannelLayerNorm32
+from ..modules.spatial import pixel_shuffle_3d
+from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
+
+
+def norm_layer(norm_type: str, *args, **kwargs) -> nn.Module:
+    """
+    Return a normalization layer.
+    """
+    if norm_type == "group":
+        return GroupNorm32(32, *args, **kwargs)
+    elif norm_type == "layer":
+        return ChannelLayerNorm32(*args, **kwargs)
+    else:
+        raise ValueError(f"Invalid norm type {norm_type}")
+
+
+class ResBlock3d(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        out_channels: Optional[int] = None,
+        norm_type: Literal["group", "layer"] = "layer",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.norm1 = norm_layer(norm_type, channels)
+        self.norm2 = norm_layer(norm_type, self.out_channels)
+        self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
+        self.conv2 = zero_module(nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1))
+        self.skip_connection = nn.Conv3d(channels, self.out_channels, 1) if channels != self.out_channels else nn.Identity()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = self.norm1(x)
+        h = F.silu(h)
+        h = self.conv1(h)
+        h = self.norm2(h)
+        h = F.silu(h)
+        h = self.conv2(h)
+        h = h + self.skip_connection(x)
+        return h
+
+
+class DownsampleBlock3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        mode: Literal["conv", "avgpool"] = "conv",
+    ):
+        assert mode in ["conv", "avgpool"], f"Invalid mode {mode}"
+
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        if mode == "conv":
+            self.conv = nn.Conv3d(in_channels, out_channels, 2, stride=2)
+        elif mode == "avgpool":
+            assert in_channels == out_channels, "Pooling mode requires in_channels to be equal to out_channels"
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if hasattr(self, "conv"):
+            return self.conv(x)
+        else:
+            return F.avg_pool3d(x, 2)
+
+
+class UpsampleBlock3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        mode: Literal["conv", "nearest"] = "conv",
+    ):
+        assert mode in ["conv", "nearest"], f"Invalid mode {mode}"
+
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        if mode == "conv":
+            self.conv = nn.Conv3d(in_channels, out_channels*8, 3, padding=1)
+        elif mode == "nearest":
+            assert in_channels == out_channels, "Nearest mode requires in_channels to be equal to out_channels"
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if hasattr(self, "conv"):
+            x = self.conv(x)
+            return pixel_shuffle_3d(x, 2)
+        else:
+            return F.interpolate(x, scale_factor=2, mode="nearest")
+        
+
+class SparseStructureEncoder(nn.Module):
+    """
+    Encoder for Sparse Structure (\mathcal{E}_S in the paper Sec. 3.3).
+    
+    Args:
+        in_channels (int): Channels of the input.
+        latent_channels (int): Channels of the latent representation.
+        num_res_blocks (int): Number of residual blocks at each resolution.
+        channels (List[int]): Channels of the encoder blocks.
+        num_res_blocks_middle (int): Number of residual blocks in the middle.
+        norm_type (Literal["group", "layer"]): Type of normalization layer.
+        use_fp16 (bool): Whether to use FP16.
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        latent_channels: int,
+        num_res_blocks: int,
+        channels: List[int],
+        num_res_blocks_middle: int = 2,
+        norm_type: Literal["group", "layer"] = "layer",
+        use_fp16: bool = False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.latent_channels = latent_channels
+        self.num_res_blocks = num_res_blocks
+        self.channels = channels
+        self.num_res_blocks_middle = num_res_blocks_middle
+        self.norm_type = norm_type
+        self.use_fp16 = use_fp16
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.input_layer = nn.Conv3d(in_channels, channels[0], 3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+        for i, ch in enumerate(channels):
+            self.blocks.extend([
+                ResBlock3d(ch, ch)
+                for _ in range(num_res_blocks)
+            ])
+            if i < len(channels) - 1:
+                self.blocks.append(
+                    DownsampleBlock3d(ch, channels[i+1])
+                )
+        
+        self.middle_block = nn.Sequential(*[
+            ResBlock3d(channels[-1], channels[-1])
+            for _ in range(num_res_blocks_middle)
+        ])
+
+        self.out_layer = nn.Sequential(
+            norm_layer(norm_type, channels[-1]),
+            nn.SiLU(),
+            nn.Conv3d(channels[-1], latent_channels*2, 3, padding=1)
+        )
+
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.use_fp16 = True
+        self.dtype = torch.float16
+        self.blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.use_fp16 = False
+        self.dtype = torch.float32
+        self.blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x: torch.Tensor, sample_posterior: bool = False, return_raw: bool = False) -> torch.Tensor:
+        h = self.input_layer(x)
+        h = h.type(self.dtype)
+
+        for block in self.blocks:
+            h = block(h)
+        h = self.middle_block(h)
+
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+
+        mean, logvar = h.chunk(2, dim=1)
+
+        if sample_posterior:
+            std = torch.exp(0.5 * logvar)
+            z = mean + std * torch.randn_like(std)
+        else:
+            z = mean
+            
+        if return_raw:
+            return z, mean, logvar
+        return z
+        
+
+class SparseStructureDecoder(nn.Module):
+    """
+    Decoder for Sparse Structure (\mathcal{D}_S in the paper Sec. 3.3).
+    
+    Args:
+        out_channels (int): Channels of the output.
+        latent_channels (int): Channels of the latent representation.
+        num_res_blocks (int): Number of residual blocks at each resolution.
+        channels (List[int]): Channels of the decoder blocks.
+        num_res_blocks_middle (int): Number of residual blocks in the middle.
+        norm_type (Literal["group", "layer"]): Type of normalization layer.
+        use_fp16 (bool): Whether to use FP16.
+    """ 
+    def __init__(
+        self,
+        out_channels: int,
+        latent_channels: int,
+        num_res_blocks: int,
+        channels: List[int],
+        num_res_blocks_middle: int = 2,
+        norm_type: Literal["group", "layer"] = "layer",
+        use_fp16: bool = False,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.latent_channels = latent_channels
+        self.num_res_blocks = num_res_blocks
+        self.channels = channels
+        self.num_res_blocks_middle = num_res_blocks_middle
+        self.norm_type = norm_type
+        self.use_fp16 = use_fp16
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
+
+        self.middle_block = nn.Sequential(*[
+            ResBlock3d(channels[0], channels[0])
+            for _ in range(num_res_blocks_middle)
+        ])
+
+        self.blocks = nn.ModuleList([])
+        for i, ch in enumerate(channels):
+            self.blocks.extend([
+                ResBlock3d(ch, ch)
+                for _ in range(num_res_blocks)
+            ])
+            if i < len(channels) - 1:
+                self.blocks.append(
+                    UpsampleBlock3d(ch, channels[i+1])
+                )
+
+        self.out_layer = nn.Sequential(
+            norm_layer(norm_type, channels[-1]),
+            nn.SiLU(),
+            nn.Conv3d(channels[-1], out_channels, 3, padding=1)
+        )
+
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+    
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.use_fp16 = True
+        self.dtype = torch.float16
+        self.blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.use_fp16 = False
+        self.dtype = torch.float32
+        self.blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = self.input_layer(x)
+        
+        h = h.type(self.dtype)
+                
+        h = self.middle_block(h)
+        for block in self.blocks:
+            h = block(h)
+
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+        return h