Upload folder using huggingface_hub

README.md

@@ -1,3 +1,29 @@
----
-license: mit
----
+# DiT Model
+
+This repository contains the implementation of the DiT (Diffusion Transformer) model, which leverages NdLinear layers for efficient multi-dimensional linear transformations. The model is designed to be compact yet powerful, suitable for various tasks requiring high-dimensional data processing.
+
+## Overview
+
+The DiT model is built using several components:
+
+- **NdLinear**: A custom PyTorch layer for projecting tensors into multi-space representations, capturing multivariate structures.
+- **NdMlp**: A multi-layer perceptron using NdLinear layers for enhanced feature extraction.
+- **NdTimestepEmbedder**: Embeds scalar timesteps into vector representations using NdLinear transformations.
+
+## Files
+
+- **mlp.py**: Contains the implementation of various MLP architectures, including NdMlp and GluMlp.
+- **models_hf.py**: Defines the DiT model architecture, including the DiTBlock and FinalLayer.
+- **ndlinear.py**: Implements the NdLinear layer, which is central to the model's ability to handle multi-dimensional data efficiently.
+
+## Installation
+
+To use the DiT model, ensure you have the required dependencies installed:
+
+```bash
+pip install torch transformers==4.52.4
+```
+
+## License
+
+This project is licensed under the MIT License. See the LICENSE file for more details.

config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "DiT"
+  ],
+  "class_dropout_prob": 0.1,
+  "depth": 28,
+  "hidden_size": 1152,
+  "in_channels": 4,
+  "input_size": 32,
+  "learn_sigma": true,
+  "mlp_ratio": 4.0,
+  "model_type": "ndlinear_dit",
+  "num_classes": 1000,
+  "num_heads": 16,
+  "out_channels": 8,
+  "patch_size": 2,
+  "torch_dtype": "float32",
+  "transformers_version": "4.52.4",
+  "tse_scale_factor": 8.0,
+  "use_ndmlp": true,
+  "use_ndtse": true,
+  "use_num_transforms": 20,
+  "use_variant": 4
+}

mlp.py

@@ -0,0 +1,269 @@
+""" MLP module w/ dropout and configurable activation layer
+
+Hacked together by / Copyright 2020 Ross Wightman
+
+Modified by Ensemble AI to use NdLinear instead of Linear. Copyright 2025
+
+"""
+from functools import partial
+
+from torch import nn as nn
+from ndlinear import NdLinear
+from timm.layers.grn import GlobalResponseNorm
+from timm.layers.helpers import to_2tuple
+
+
+class NdMlp(nn.Module):
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.GELU,
+            norm_layer=None,
+            bias=True,
+            drop=0.,
+            use_variant=4
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
+        self.use_variant = use_variant
+        drop_probs = to_2tuple(drop)
+        self.fc1 = NdLinear((in_features, 1), (hidden_features // 4, 1))  # (384, 1), (384, 1)
+        self.fc2 = NdLinear((in_features, 1), (hidden_features // 4, 1))
+
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x_dim0, x_dim1, x_dim2 = x.shape
+        # print(f"x.shape: {x.shape}")
+        x = x.reshape(x_dim0 * x_dim1, x_dim2, 1) if self.use_variant != 9 else x
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        # x = self.norm(x) #
+        x = self.fc2(x)
+        x = x.reshape(x_dim0, x_dim1, x_dim2) if self.use_variant != 9 else x
+        x = self.drop2(x)
+        return x
+
+class GluMlp(nn.Module):
+    """ MLP w/ GLU style gating
+    See: https://arxiv.org/abs/1612.08083, https://arxiv.org/abs/2002.05202
+
+    NOTE: When use_conv=True, expects 2D NCHW tensors, otherwise N*C expected.
+    """
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.Sigmoid,
+            norm_layer=None,
+            bias=True,
+            drop=0.,
+            use_conv=False,
+            gate_last=True,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        assert hidden_features % 2 == 0
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+        self.chunk_dim = 1 if use_conv else -1
+        self.gate_last = gate_last  # use second half of width for gate
+
+        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.norm = norm_layer(hidden_features // 2) if norm_layer is not None else nn.Identity()
+        self.fc2 = linear_layer(hidden_features // 2, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def init_weights(self):
+        # override init of fc1 w/ gate portion set to weight near zero, bias=1
+        if self.fc1.bias is not None:
+            nn.init.ones_(self.fc1.bias[self.fc1.bias.shape[0] // 2:])
+        nn.init.normal_(self.fc1.weight[self.fc1.weight.shape[0] // 2:], std=1e-6)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x1, x2 = x.chunk(2, dim=self.chunk_dim)
+        x = x1 * self.act(x2) if self.gate_last else self.act(x1) * x2
+        x = self.drop1(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+SwiGLUPacked = partial(GluMlp, act_layer=nn.SiLU, gate_last=False)
+
+
+class SwiGLU(nn.Module):
+    """ SwiGLU
+    NOTE: GluMLP above can implement SwiGLU, but this impl has split fc1 and
+    better matches some other common impl which makes mapping checkpoints simpler.
+    """
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.SiLU,
+            norm_layer=None,
+            bias=True,
+            drop=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_g = nn.Linear(in_features, hidden_features, bias=bias[0])
+        self.fc1_x = nn.Linear(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def init_weights(self):
+        # override init of fc1 w/ gate portion set to weight near zero, bias=1
+        if self.fc1_g.bias is not None:
+            nn.init.ones_(self.fc1_g.bias)
+        nn.init.normal_(self.fc1_g.weight, std=1e-6)
+
+    def forward(self, x):
+        x_gate = self.fc1_g(x)
+        x = self.fc1_x(x)
+        x = self.act(x_gate) * x
+        x = self.drop1(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+class GatedMlp(nn.Module):
+    """ MLP as used in gMLP
+    """
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.GELU,
+            norm_layer=None,
+            gate_layer=None,
+            bias=True,
+            drop=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])
+        if gate_layer is not None:
+            assert hidden_features % 2 == 0
+            self.gate = gate_layer(hidden_features)
+            hidden_features = hidden_features // 2  # FIXME base reduction on gate property?
+        else:
+            self.gate = nn.Identity()
+        self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
+        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.gate(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+class ConvMlp(nn.Module):
+    """ MLP using 1x1 convs that keeps spatial dims (for 2D NCHW tensors)
+    """
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.ReLU,
+            norm_layer=None,
+            bias=True,
+            drop=0.,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
+
+        self.fc1 = nn.Conv2d(in_features, hidden_features, kernel_size=1, bias=bias[0])
+        self.norm = norm_layer(hidden_features) if norm_layer else nn.Identity()
+        self.act = act_layer()
+        self.drop = nn.Dropout(drop)
+        self.fc2 = nn.Conv2d(hidden_features, out_features, kernel_size=1, bias=bias[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.norm(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        return x
+
+
+class GlobalResponseNormMlp(nn.Module):
+    """ MLP w/ Global Response Norm (see grn.py), nn.Linear or 1x1 Conv2d
+
+    NOTE: Intended for '2D' NCHW (use_conv=True) or NHWC (use_conv=False, channels-last) tensor layouts
+    """
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.GELU,
+            bias=True,
+            drop=0.,
+            use_conv=False,
+    ):
+        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)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+
+        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.grn = GlobalResponseNorm(hidden_features, channels_last=not use_conv)
+        self.fc2 = linear_layer(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.grn(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8d3d3e218068ea8163743a1d4049dc3e225ac3bec897cda9816598f7ce07a19
+size 915766976

models_hf.py

@@ -0,0 +1,431 @@
+# 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.
+# --------------------------------------------------------
+# References:
+# GLIDE: https://github.com/openai/glide-text2im
+# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py
+#
+# Modifications Copyright (c) Ensemble AI, 2025.
+# Description of modifications: Using NdLinear in the model to
+# make the model more compact yet with similar performance.
+import torch
+import torch.nn as nn
+import numpy as np
+import math
+from timm.models.vision_transformer import PatchEmbed, Attention, Mlp
+from mlp import NdMlp
+from ndlinear import NdLinear
+from transformers import PreTrainedModel, PretrainedConfig
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+class TimestepEmbedder(nn.Module):
+    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):
+        half = dim // 2
+        freqs = torch.exp(
+            -math.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 NdTimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256, use_num_transforms=2, tse_scale_factor=1, knowledge_transfer=False, src_layers=None):
+        super().__init__()
+        self.activation = nn.SiLU()
+        self.frequency_embedding_size = frequency_embedding_size
+        self.use_num_transforms = use_num_transforms
+
+        if knowledge_transfer and not src_layers:
+            raise ValueError("Source layers must be provided for knowledge transfer.")
+
+        if use_num_transforms == 2:
+            self.ndlinear_1 = NdLinear((frequency_embedding_size // 16, 16), (int(hidden_size // tse_scale_factor // 2), 2))
+            self.ndlinear_2 = NdLinear((int(hidden_size // tse_scale_factor // 2), 2), (hidden_size, 1))
+
+        if use_num_transforms == 20:
+            self.ndlinear_1 = NdLinear((frequency_embedding_size, 1), (int(hidden_size // tse_scale_factor), 1))
+            self.ndlinear_2 = NdLinear((int(hidden_size // tse_scale_factor), 1), (hidden_size, 1))
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        half = dim // 2
+        freqs = torch.exp(
+            -math.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)
+        if self.use_num_transforms == 2:
+            t_freq = t_freq.reshape(*t_freq.shape, 1)
+        elif self.use_num_transforms == 21:
+            t_freq = t_freq.reshape(t_freq.shape[0], 16, 16)
+        elif self.use_num_transforms == 3:
+            t_freq = t_freq.reshape(t_freq.shape[0], t_freq.shape[1] // 16, 16, 1)
+        elif self.use_num_transforms == 4:
+            t_freq = t_freq.reshape(t_freq.shape[0], t_freq.shape[1] // 16, 4, 4, 1)
+        t_emb = self.ndlinear_1(t_freq)
+        t_emb = self.activation(t_emb)
+        t_emb = self.ndlinear_2(t_emb)
+        t_emb = t_emb.squeeze()
+        return t_emb
+
+class LabelEmbedder(nn.Module):
+    def __init__(self, num_classes, hidden_size, dropout_prob):
+        super().__init__()
+        use_cfg_embedding = dropout_prob > 0
+        self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
+        self.num_classes = num_classes
+        self.dropout_prob = dropout_prob
+
+    def token_drop(self, labels, force_drop_ids=None):
+        if force_drop_ids is None:
+            drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
+        else:
+            drop_ids = force_drop_ids == 1
+        labels = torch.where(drop_ids, self.num_classes, labels)
+        return labels
+
+    def forward(self, labels, train, force_drop_ids=None):
+        use_dropout = self.dropout_prob > 0
+        if (train and use_dropout) or (force_drop_ids is not None):
+            labels = self.token_drop(labels, force_drop_ids)
+        embeddings = self.embedding_table(labels)
+        return embeddings
+
+class DiTBlock(nn.Module):
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, use_ndmlp=False, use_variant=4, use_ndadaln=False, **block_kwargs):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        if use_ndmlp:
+            self.mlp = NdMlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0, use_variant=use_variant)
+        else:
+            self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
+
+    def forward(self, x, c):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
+        x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
+        modulated_x = modulate(self.norm2(x), shift_mlp, scale_mlp)
+        mlp_output = self.mlp(modulated_x)
+        gated_mlp_output = gate_mlp.unsqueeze(1) * mlp_output
+        x = x + gated_mlp_output
+        return x
+
+class FinalLayer(nn.Module):
+    def __init__(self, hidden_size, patch_size, out_channels, use_ndadaln=False):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.use_ndadaln = use_ndadaln
+        if self.use_ndadaln:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                NdLinear((hidden_size, 1), (2 * hidden_size, 1))
+            )
+        else:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(hidden_size, 2 * hidden_size, bias=True)
+            )
+
+    def forward(self, x, c):
+        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+class DiTConfig(PretrainedConfig):
+    model_type = "ndlinear_dit"
+
+    def __init__(self, input_size=32, patch_size=2, in_channels=4, hidden_size=1152, depth=28, num_heads=16, mlp_ratio=4.0, class_dropout_prob=0.1, num_classes=1000, learn_sigma=True, use_ndmlp=False, use_ndtse=False, use_variant=4, tse_scale_factor=2, use_num_transforms=2, **kwargs):
+        super().__init__(**kwargs)
+        self.input_size = input_size
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if learn_sigma else in_channels
+        self.hidden_size = hidden_size
+        self.depth = depth
+        self.num_heads = num_heads
+        self.mlp_ratio = mlp_ratio
+        self.class_dropout_prob = class_dropout_prob
+        self.num_classes = num_classes
+        self.learn_sigma = learn_sigma
+        self.use_ndmlp = use_ndmlp
+        self.use_ndtse = use_ndtse
+        self.use_variant = use_variant
+        self.tse_scale_factor = tse_scale_factor
+        self.use_num_transforms = use_num_transforms
+
+class DiT(PreTrainedModel):
+    config_class = DiTConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.input_size = config.input_size
+        self.patch_size = config.patch_size
+        self.in_channels = config.in_channels
+        self.hidden_size = config.hidden_size
+        self.depth = config.depth
+        self.num_heads = config.num_heads
+        self.mlp_ratio = config.mlp_ratio
+        self.class_dropout_prob = config.class_dropout_prob
+        self.num_classes = config.num_classes
+        self.learn_sigma = config.learn_sigma
+        self.use_ndmlp = config.use_ndmlp
+        self.use_ndtse = config.use_ndtse
+        self.use_variant = config.use_variant
+        self.tse_scale_factor = config.tse_scale_factor
+        self.use_num_transforms = config.use_num_transforms
+        self.out_channels = config.out_channels
+        self.ndadaln = getattr(config, "ndadaln", False)
+
+        self.x_embedder = PatchEmbed(self.input_size, self.patch_size, self.in_channels, self.hidden_size, bias=True)
+
+        if self.use_ndtse:
+            self.t_embedder = NdTimestepEmbedder(
+                hidden_size=self.hidden_size,
+                frequency_embedding_size=256,
+                use_num_transforms=self.use_num_transforms,
+                tse_scale_factor=1,
+                knowledge_transfer=False,
+                src_layers=None
+            )
+        else:
+            self.t_embedder = TimestepEmbedder(self.hidden_size)
+
+        self.y_embedder = LabelEmbedder(self.num_classes, self.hidden_size, self.class_dropout_prob)
+        num_patches = self.x_embedder.num_patches
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, self.hidden_size), requires_grad=False)
+
+        self.blocks = nn.ModuleList([
+            DiTBlock(self.hidden_size, self.num_heads, mlp_ratio=self.mlp_ratio,
+                     use_ndmlp=self.use_ndmlp, use_variant=self.use_variant)
+            for _ in range(self.depth)
+        ])
+        if self.use_ndmlp:
+            approx_gelu = lambda: nn.GELU(approximate="tanh")
+            for idx, layer in enumerate(self.blocks):
+                if idx % 2 == 0:
+                    layer.mlp = NdMlp(
+                        in_features=self.hidden_size,
+                        hidden_features=self.hidden_size * 4,
+                        act_layer=approx_gelu,
+                        drop=0,
+                        use_variant=self.use_variant
+                    )
+        self.final_layer = FinalLayer(self.hidden_size, self.patch_size, self.out_channels, use_ndadaln=self.ndadaln)
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        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)
+
+        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5))
+        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
+
+        w = self.x_embedder.proj.weight.data
+        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+        nn.init.constant_(self.x_embedder.proj.bias, 0)
+
+        nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)
+
+        if not self.use_ndtse:
+            nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+            nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
+        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
+        nn.init.constant_(self.final_layer.linear.weight, 0)
+        nn.init.constant_(self.final_layer.linear.bias, 0)
+
+    def unpatchify(self, x):
+        c = self.out_channels
+        p = self.x_embedder.patch_size[0]
+        h = w = int(x.shape[1] ** 0.5)
+        assert h * w == x.shape[1]
+
+        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
+        x = torch.einsum('nhwpqc->nchpwq', x)
+        imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
+        return imgs
+
+    def ckpt_wrapper(self, module):
+        def ckpt_forward(*inputs):
+            outputs = module(*inputs)
+            return outputs
+
+        return ckpt_forward
+
+    def forward(self, x, t, y):
+        x = self.x_embedder(x) + self.pos_embed
+        t = self.t_embedder(t)
+        y = self.y_embedder(y, self.training)
+        c = t + y
+
+        for block in self.blocks:
+            x = torch.utils.checkpoint.checkpoint(self.ckpt_wrapper(block), x, c)
+        x = self.final_layer(x, c)
+        x = self.unpatchify(x)
+        return x
+
+    def forward_with_cfg(self, x, t, y, cfg_scale):
+        half = x[: len(x) // 2]
+        combined = torch.cat([half, half], dim=0)
+        model_out = self.forward(combined, t, y)
+        eps, rest = model_out[:, :3], model_out[:, 3:]
+        cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
+        half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
+        eps = torch.cat([half_eps, half_eps], dim=0)
+        return torch.cat([eps, rest], dim=1)
+
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
+    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)
+    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 cls_token and extra_tokens > 0:
+        pos_embed = np.concatenate([np.zeros([extra_tokens, 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
+
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)
+    return emb
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float64)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega
+
+    pos = pos.reshape(-1)
+    out = np.einsum('m,d->md', pos, omega)
+
+    emb_sin = np.sin(out)
+    emb_cos = np.cos(out)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)
+    return emb
+
+def DiT_XL_2(**kwargs):
+    config = DiTConfig(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs)
+    return DiT(config)
+
+def DiT_XL_4(**kwargs):
+    config = DiTConfig(depth=28, hidden_size=1152, patch_size=4, num_heads=16, **kwargs)
+    return DiT(config)
+
+def DiT_XL_8(**kwargs):
+    config = DiTConfig(depth=28, hidden_size=1152, patch_size=8, num_heads=16, **kwargs)
+    return DiT(config)
+
+def DiT_L_2(**kwargs):
+    config = DiTConfig(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs)
+    return DiT(config)
+
+def DiT_L_4(**kwargs):
+    config = DiTConfig(depth=24, hidden_size=1024, patch_size=4, num_heads=16, **kwargs)
+    return DiT(config)
+
+def DiT_L_8(**kwargs):
+    config = DiTConfig(depth=24, hidden_size=1024, patch_size=8, num_heads=16, **kwargs)
+    return DiT(config)
+
+def DiT_B_2(**kwargs):
+    config = DiTConfig(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs)
+    return DiT(config)
+
+def DiT_B_4(**kwargs):
+    config = DiTConfig(depth=12, hidden_size=768, patch_size=4, num_heads=12, **kwargs)
+    return DiT(config)
+
+def DiT_B_8(**kwargs):
+    config = DiTConfig(depth=12, hidden_size=768, patch_size=8, num_heads=12, **kwargs)
+    return DiT(config)
+
+def DiT_S_2(**kwargs):
+    config = DiTConfig(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_S_4(**kwargs):
+    config = DiTConfig(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_S_8(**kwargs):
+    config = DiTConfig(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_MS_2(**kwargs):
+    config = DiTConfig(depth=6, hidden_size=384, patch_size=2, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_MS_4(**kwargs):
+    config = DiTConfig(depth=6, hidden_size=384, patch_size=4, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_MS_8(**kwargs):
+    config = DiTConfig(depth=6, hidden_size=384, patch_size=8, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_XS_2(**kwargs):
+    config = DiTConfig(depth=1, hidden_size=384, patch_size=2, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_XS_4(**kwargs):
+    config = DiTConfig(depth=1, hidden_size=384, patch_size=4, num_heads=6, **kwargs)
+    return DiT(config)
+
+def DiT_XS_8(**kwargs):
+    config = DiTConfig(depth=1, hidden_size=384, patch_size=8, num_heads=6, **kwargs)
+    return DiT(config)
+
+DiT_models = {
+    'DiT-XL/2': DiT_XL_2, 'DiT-XL/4': DiT_XL_4, 'DiT-XL/8': DiT_XL_8,
+    'DiT-L/2': DiT_L_2, 'DiT-L/4': DiT_L_4, 'DiT-L/8': DiT_L_8,
+    'DiT-B/2': DiT_B_2, 'DiT-B/4': DiT_B_4, 'DiT-B/8': DiT_B_8,
+    'DiT-S/2': DiT_S_2, 'DiT-S/4': DiT_S_4, 'DiT-S/8': DiT_S_8,
+    'DiT-XS/2': DiT_XS_2, 'DiT-XS/4': DiT_XS_4, 'DiT-XS/8': DiT_XS_8,
+    'DiT-MS/2': DiT_MS_2, 'DiT-MS/4': DiT_MS_4, 'DiT-MS/8': DiT_MS_8
+}

ndlinear.py

@@ -0,0 +1,91 @@
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+
+
+class NdLinear(nn.Module):
+    def __init__(self, input_dims: tuple, hidden_size: tuple, transform_outer=True, act_func=None, use_bias=True):
+        """
+        NdLinear: A PyTorch layer for projecting tensors into multi-space representations.
+        
+        Unlike conventional embedding layers that map into a single vector space, NdLinear 
+        transforms tensors across a collection of vector spaces, capturing multivariate structure 
+        and topical information that standard deep learning architectures typically lose.
+
+        Args:
+            input_dims (tuple): Shape of input tensor (excluding batch dimension).
+            hidden_size (tuple): Target hidden dimensions after transformation.
+        """
+        super(NdLinear, self).__init__()
+
+        if len(input_dims) != len(hidden_size):
+            raise Exception("Input shape and hidden shape do not match.")
+
+        self.input_dims = input_dims
+        self.hidden_size = hidden_size
+        self.num_layers = len(input_dims)  # Must match since dims are equal
+        # Custom activation function. Default to Identity -> Do nothing.
+        self.act_func = act_func if act_func is not None else nn.Identity()
+        self.transform_outer = transform_outer
+
+        # Define transformation layers per dimension
+        self.align_layers = nn.ModuleList([
+            nn.Linear(input_dims[i], hidden_size[i], bias=use_bias) for i in range(self.num_layers)
+        ])
+        self.initialize_weights()
+
+
+    def initialize_weights(self, mean=0.0, std=0.02):
+        for layer in self.align_layers:
+            nn.init.normal_(layer.weight, mean=mean, std=std)
+            if layer.bias is not None:
+                nn.init.constant_(layer.bias, 0)
+
+
+    def forward(self, X):
+        """
+        Forward pass to project input tensor into a new multi-space representation.
+        - Incrementally transposes, flattens, applies linear layers, and restores shape.
+
+        Expected Input Shape: [batch_size, *input_dims]
+        Output Shape: [batch_size, *hidden_size]
+
+        Args:
+            X (torch.Tensor): Input tensor with shape [batch_size, *input_dims]
+
+        Returns:
+            torch.Tensor: Output tensor with shape [batch_size, *hidden_size]
+        """
+        num_transforms = self.num_layers  # Number of transformations
+        
+        # Define iteration order
+        # transform_indices = range(num_transforms) if transform_outer else reversed(range(num_transforms))
+
+        for i in range(num_transforms):
+            if self.transform_outer:
+                layer = self.align_layers[i]
+                transpose_dim = i + 1
+            else:
+                layer = self.align_layers[num_transforms - (i+1)]
+                transpose_dim = num_transforms - i
+
+            # Transpose the selected dimension to the last position
+            X = torch.transpose(X, transpose_dim, num_transforms).contiguous()
+
+            # Store original shape before transformation
+            X_size = X.shape[:-1]
+
+            # Flatten everything except the last dimension
+            X = X.view(-1, X.shape[-1])
+
+            # Apply transformation
+            X = self.act_func(layer(X))
+            
+            # Reshape back to the original spatial structure (with new embedding dim)
+            X = X.view(*X_size, X.shape[-1])
+
+            # Transpose the dimension back to its original position
+            X = torch.transpose(X, transpose_dim, num_transforms).contiguous()
+
+        return X