single_script.py

import torch
import torch.nn as nn
import numpy as np
import math
from timm.models.vision_transformer import PatchEmbed, Attention, Mlp




# diffusion policy import
from typing import Tuple, Sequence, Dict, Union, Optional, Callable
import numpy as np
import math
import torch
import torch.nn as nn
import torchvision
import collections

from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
from diffusers.training_utils import EMAModel
from diffusers.optimization import get_scheduler
from tqdm.auto import tqdm
import random

def modulate(x, shift, scale):
    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)


#################################################################################
#               Embedding Layers for Timesteps and Class Labels                 #
#################################################################################

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.
        :param t: a 1-D Tensor of N indices, one per batch element.
                          These may be fractional.
        :param dim: the dimension of the output.
        :param max_period: controls the minimum frequency of the embeddings.
        :return: 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(
            -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 LabelEmbedder(nn.Module):
    """
    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
    """
    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):
        """
        Drops labels to enable classifier-free guidance.
        """
        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




#@markdown ### **Vision Encoder**
#@markdown
#@markdown Defines helper functions:
#@markdown - `get_resnet` to initialize standard ResNet vision encoder
#@markdown - `replace_bn_with_gn` to replace all BatchNorm layers with GroupNorm

def get_resnet(name:str, weights=None, **kwargs) -> nn.Module:
    """
    name: resnet18, resnet34, resnet50
    weights: "IMAGENET1K_V1", None
    """
    # Use standard ResNet implementation from torchvision
    func = getattr(torchvision.models, name)
    resnet = func(weights=weights, **kwargs)

    # remove the final fully connected layer
    # for resnet18, the output dim should be 512
    resnet.fc = torch.nn.Identity()
    return resnet


def replace_submodules(
        root_module: nn.Module,
        predicate: Callable[[nn.Module], bool],
        func: Callable[[nn.Module], nn.Module]) -> nn.Module:
    """
    Replace all submodules selected by the predicate with
    the output of func.

    predicate: Return true if the module is to be replaced.
    func: Return new module to use.
    """
    if predicate(root_module):
        return func(root_module)

    bn_list = [k.split('.') for k, m
        in root_module.named_modules(remove_duplicate=True)
        if predicate(m)]
    for *parent, k in bn_list:
        parent_module = root_module
        if len(parent) > 0:
            parent_module = root_module.get_submodule('.'.join(parent))
        if isinstance(parent_module, nn.Sequential):
            src_module = parent_module[int(k)]
        else:
            src_module = getattr(parent_module, k)
        tgt_module = func(src_module)
        if isinstance(parent_module, nn.Sequential):
            parent_module[int(k)] = tgt_module
        else:
            setattr(parent_module, k, tgt_module)
    # verify that all modules are replaced
    bn_list = [k.split('.') for k, m
        in root_module.named_modules(remove_duplicate=True)
        if predicate(m)]
    assert len(bn_list) == 0
    return root_module

def replace_bn_with_gn(
    root_module: nn.Module,
    features_per_group: int=16) -> nn.Module:
    """
    Relace all BatchNorm layers with GroupNorm.
    """
    replace_submodules(
        root_module=root_module,
        predicate=lambda x: isinstance(x, nn.BatchNorm2d),
        func=lambda x: nn.GroupNorm(
            num_groups=x.num_features//features_per_group,
            num_channels=x.num_features)
    )
    return root_module





#################################################################################
#                                 Core DiT Model                                #
#################################################################################

class DiTBlock(nn.Module):
    """
    A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
    """
    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **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")
        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )

    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))
        x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
        return x


class FinalLayer(nn.Module):
    """
    The final layer of DiT.
    """
    def __init__(self, hidden_size, patch_size, out_channels):
        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.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 DiT(nn.Module):
    """
    Diffusion model with a Transformer backbone.
    """
    def __init__(
        self,
        horizon=8,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
        cond_dim = 512, ## dim of image encodings # ResNet18 has output dim of 512
        num_points = 25, ## remove if pos emb is not used
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = 2*horizon ## (x,y) for each point
        self.out_channels = self.in_channels * 2 if learn_sigma else self.in_channels
        self.patch_size = 1
        self.num_heads = num_heads
        self.num_points = num_points
        
        self.x_embedder = nn.Linear(self.in_channels, hidden_size)        
        self.t_embedder = TimestepEmbedder(hidden_size)


        # construct ResNet18 encoder
        # if you have multiple camera views, use seperate encoder weights for each view.
        vision_encoder = get_resnet('resnet18')

        # IMPORTANT!
        # replace all BatchNorm with GroupNorm to work with EMA
        # performance will tank if you forget to do this!
        vision_encoder = replace_bn_with_gn(vision_encoder)
        self.vision_encoder = vision_encoder

        self.y_embedder = nn.Linear(2*cond_dim, hidden_size)


        # Will use fixed sin-cos embedding:
        self.pos_embed = nn.Parameter(torch.zeros(1, self.num_points, hidden_size), requires_grad=False)

        self.blocks = nn.ModuleList([
            DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(depth)
        ])

        patch_size = 1 ### because patches are points
        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)


        ## resnet is initialized by get_resnet() above
        ## x_embedder, y_embedder has default initilization
        self.initialize_weights()

    def initialize_weights(self):
        # 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 (and freeze) pos_embed by sin-cos embedding:
        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.num_points ** 0.5))
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))

        # 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:
        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.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):
        """
        x: (N, T, patch_size**2 * C)
        imgs: (N, H, W, C)
        """
        c = self.out_channels
        p = self.num_points
        
        x = torch.einsum('npc->ncp', x)
        #print("x after einsum",x.shape)
        # x = x.reshape(shape=(x.shape[0],p,int(c/2),2))
        return x


    def forward(self, x, t, y):
        """
        Forward pass of DiT.
        x: (N, C, P) tensor of point tracks, C = 2*T where T=8 horizon 
        t: (N,) tensor of diffusion timesteps
        y: (N,2,3,96,96) tensor of initial and goal images
        """
        x = torch.einsum('ncp->npc', x)
        
        x = self.x_embedder(x) + self.pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
        
        
        t = self.t_embedder(t)                   # (N, D)
        
        y_features = self.vision_encoder(y.flatten(end_dim=1)) # (2,512)
        y_features = y_features.reshape(*y.shape[:2],-1) # (1,2,512)
        y_features = y_features.flatten(start_dim=1) # (1,2*512)

       

        y = self.y_embedder(y_features)    # (N, D)

      
        c = t + y                                # (N, D)
        for block in self.blocks:
            x = block(x, c)                      # (N, T, D)
        x = self.final_layer(x, c)                # (N, T, patch_size ** 2 * out_channels)
        x = self.unpatchify(x)                   # (N, out_channels, H, W)
        return x

    def forward_with_cfg(self, x, t, y, cfg_scale):
        """
        Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance.
        """
        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)
        model_out = self.forward(combined, t, y)
        # For exact reproducibility reasons, we apply classifier-free guidance on only
        # three channels by default. The standard approach to cfg applies it to all channels.
        # This can be done by uncommenting the following line and commenting-out the line following that.
        # eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:]
        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)



class DiT_noPosEmb_withLang(nn.Module):
    """
    Diffusion model with a Transformer backbone without position embeddings for points
    """
    def __init__(
        self,
        horizon=8,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
        cond_dim = 512, ## dim of image encodings # ResNet18 has output dim of 512
        num_points = 25, ## remove if pos emb is not used
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = 2*horizon ## (x,y) for each point
        self.out_channels = self.in_channels * 2 if learn_sigma else self.in_channels
        self.patch_size = 1
        self.num_heads = num_heads        
        self.x_embedder = nn.Linear(self.in_channels, hidden_size)        
        self.t_embedder = TimestepEmbedder(hidden_size)


        # construct ResNet18 encoder
        # if you have multiple camera views, use seperate encoder weights for each view.
        vision_encoder = get_resnet('resnet18')

        # IMPORTANT!
        # replace all BatchNorm with GroupNorm to work with EMA
        # performance will tank if you forget to do this!
        vision_encoder = replace_bn_with_gn(vision_encoder)
        self.vision_encoder = vision_encoder

        # self.y_embedder = nn.Linear(2*cond_dim, hidden_size) ## change it to be different for initial and goal images
        self.y_embedder_init = nn.Linear(cond_dim, hidden_size)
        self.y_embedder_goal = nn.Linear(cond_dim, hidden_size)
        self.lang_embedder = nn.Linear(cond_dim, hidden_size) ## language encoder has dim 512


        self.blocks = nn.ModuleList([
            DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(depth)
        ])

        patch_size = 1 ### because patches are points
        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)


        ## resnet is initialized by get_resnet() above
        ## x_embedder, y_embedder has default initilization
        self.initialize_weights()

    def initialize_weights(self):
        # 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:
        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.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):
        """
        x: (N, T, patch_size**2 * C)
        imgs: (N, H, W, C)
        """        
        x = torch.einsum('npc->ncp', x)
        #print("x after einsum",x.shape)
        # x = x.reshape(shape=(x.shape[0],p,int(c/2),2))
        return x


    def forward(self, x, t, y, lang, iflang):
        """
        Forward pass of DiT.
        x: (N, C, P) tensor of point tracks, C = 2*T where T=8 horizon 
        t: (N,) tensor of diffusion timesteps
        y: (N,2,3,96,96) tensor of initial and goal images
        lang: (N,1024) tensor of language embedding
        """
        x = torch.einsum('ncp->npc', x)
        
        x = self.x_embedder(x)  # (N, T, D), where T = H * W / patch_size ** 2
        
        
        t = self.t_embedder(t)                   # (N, D)
        
        y_features = self.vision_encoder(y.flatten(end_dim=1)) # (2,512)
        y_features = y_features.reshape(*y.shape[:2],-1) # (1,2,512)

        y_features_init = y_features[:,0,:].flatten(start_dim=1)
        y_features_goal = y_features[:,1,:].flatten(start_dim=1)


        y_init = self.y_embedder_init(y_features_init)    # (N, D)
        y_goal = self.y_embedder_goal(y_features_goal)    # (N, D)
        lang_emb = self.lang_embedder(lang)    # (N, D)

        ## choose either goal image or lang goal 
        c = t + y_init + y_goal * (1 - iflang) + lang_emb * iflang      # (N, D)
        for block in self.blocks:
            x = block(x, c)                      # (N, T, D)
        x = self.final_layer(x, c)                # (N, T, patch_size ** 2 * out_channels)
        x = self.unpatchify(x)                   # (N, out_channels, H, W)
        return x

    def forward_with_cfg(self, x, t, y, cfg_scale):
        """
        Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance.
        """
        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)
        model_out = self.forward(combined, t, y)
        # For exact reproducibility reasons, we apply classifier-free guidance on only
        # three channels by default. The standard approach to cfg applies it to all channels.
        # This can be done by uncommenting the following line and commenting-out the line following that.
        # eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:]
        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)




class DiT_noPosEmb(nn.Module):
    """
    Diffusion model with a Transformer backbone without position embeddings for points
    """
    def __init__(
        self,
        horizon=8,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
        cond_dim = 512, ## dim of image encodings # ResNet18 has output dim of 512
        num_points = 25, ## remove if pos emb is not used
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = 2*horizon ## (x,y) for each point
        self.out_channels = self.in_channels * 2 if learn_sigma else self.in_channels
        self.patch_size = 1
        self.num_heads = num_heads        
        self.x_embedder = nn.Linear(self.in_channels, hidden_size)        
        self.t_embedder = TimestepEmbedder(hidden_size)


        # construct ResNet18 encoder
        # if you have multiple camera views, use seperate encoder weights for each view.
        vision_encoder = get_resnet('resnet18')

        # IMPORTANT!
        # replace all BatchNorm with GroupNorm to work with EMA
        # performance will tank if you forget to do this!
        vision_encoder = replace_bn_with_gn(vision_encoder)
        self.vision_encoder = vision_encoder

        self.y_embedder = nn.Linear(2*cond_dim, hidden_size)


        self.blocks = nn.ModuleList([
            DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(depth)
        ])

        patch_size = 1 ### because patches are points
        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)


        ## resnet is initialized by get_resnet() above
        ## x_embedder, y_embedder has default initilization
        self.initialize_weights()

    def initialize_weights(self):
        # 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:
        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.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):
        """
        x: (N, T, patch_size**2 * C)
        imgs: (N, H, W, C)
        """        
        x = torch.einsum('npc->ncp', x)
        #print("x after einsum",x.shape)
        # x = x.reshape(shape=(x.shape[0],p,int(c/2),2))
        return x


    def forward(self, x, t, y):
        """
        Forward pass of DiT.
        x: (N, C, P) tensor of point tracks, C = 2*T where T=8 horizon 
        t: (N,) tensor of diffusion timesteps
        y: (N,2,3,96,96) tensor of initial and goal images
        """
        x = torch.einsum('ncp->npc', x)
        
        x = self.x_embedder(x)  # (N, T, D), where T = H * W / patch_size ** 2
        
        
        t = self.t_embedder(t)                   # (N, D)
        
        y_features = self.vision_encoder(y.flatten(end_dim=1)) # (2,512)
        y_features = y_features.reshape(*y.shape[:2],-1) # (1,2,512)
        y_features = y_features.flatten(start_dim=1) # (1,2*512)

        y = self.y_embedder(y_features)    # (N, D)

      
        c = t + y                                # (N, D)
        for block in self.blocks:
            x = block(x, c)                      # (N, T, D)
        x = self.final_layer(x, c)                # (N, T, patch_size ** 2 * out_channels)
        x = self.unpatchify(x)                   # (N, out_channels, H, W)
        return x

    def forward_with_cfg(self, x, t, y, cfg_scale):
        """
        Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance.
        """
        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)
        model_out = self.forward(combined, t, y)
        # For exact reproducibility reasons, we apply classifier-free guidance on only
        # three channels by default. The standard approach to cfg applies it to all channels.
        # This can be done by uncommenting the following line and commenting-out the line following that.
        # eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:]
        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)


#################################################################################
#                   Sine/Cosine Positional Embedding Functions                  #
#################################################################################
# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py

def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
    """
    grid_size: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    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)  # here w goes first
    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

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
    return emb


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position
    pos: a list of positions to be encoded: size (M,)
    out: (M, D)
    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out) # (M, D/2)
    emb_cos = np.cos(out) # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb







# #################################################################################
# #                                   DiT Configs                                  #
# #################################################################################

def DiT_XL_NoPosEmb(**kwargs):
    return DiT_noPosEmb(depth=34, hidden_size=1440,  num_heads=18, **kwargs)


def DiT_XL_NoPosEmb_Lang_LARGE(**kwargs):
    return DiT_noPosEmb_withLang(depth=34, hidden_size=1440,  num_heads=18, **kwargs)

def DiT_XL_NoPosEmb_Lang(**kwargs):
    return DiT_noPosEmb_withLang(depth=24, hidden_size=1152,  num_heads=16, **kwargs)

def DiT_XL_2(**kwargs):
    return DiT(depth=28, hidden_size=1152,  num_heads=16, **kwargs)

def DiT_XL_2_NoPosEmb(**kwargs):
    return DiT_noPosEmb(depth=28, hidden_size=1152,  num_heads=16, **kwargs)

def DiT_L_2_NoPosEmb(**kwargs):
    return DiT_noPosEmb(depth=24, hidden_size=1024,  num_heads=16, **kwargs)

def DiT_XL_4(**kwargs):
    return DiT(depth=28, hidden_size=1152,  num_heads=16, **kwargs)

def DiT_XL_8(**kwargs):
    return DiT(depth=28, hidden_size=1152,  num_heads=16, **kwargs)

def DiT_L_2(**kwargs):
    return DiT(depth=24, hidden_size=1024,  num_heads=16, **kwargs)

def DiT_L_4(**kwargs):
    return DiT(depth=24, hidden_size=1024,  num_heads=16, **kwargs)

def DiT_L_8(**kwargs):
    return DiT(depth=24, hidden_size=1024,  num_heads=16, **kwargs)

def DiT_B_2(**kwargs):
    return DiT(depth=12, hidden_size=768,  num_heads=12, **kwargs)

def DiT_B_4(**kwargs):
    return DiT(depth=12, hidden_size=768, num_heads=12, **kwargs)

def DiT_B_8(**kwargs):
    return DiT(depth=12, hidden_size=768,  num_heads=12, **kwargs)

def DiT_S_2(**kwargs):
    return DiT(depth=12, hidden_size=384, num_heads=6, **kwargs)

def DiT_S_4(**kwargs):
    return DiT(depth=12, hidden_size=384, num_heads=6, **kwargs)

def DiT_S_8(**kwargs):
    return DiT(depth=12, hidden_size=384, num_heads=6, **kwargs)


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-XL/2-NoPosEmb': DiT_XL_2_NoPosEmb,
    'DiT-L/2-NoPosEmb': DiT_L_2_NoPosEmb,
    'DiT_XL_NoPosEmb' : DiT_XL_NoPosEmb,
    'DiT-XL-NoPosEmb-Lang' : DiT_XL_NoPosEmb_Lang,
    'DiT-XL-NoPosEmb-Lang-LARGE' : DiT_XL_NoPosEmb_Lang_LARGE,
}