import torch
import torch.nn as nn
import math
[docs]
class TimestepEmbedder(nn.Module):
"""
Embeds scalar timesteps into vector representations using a sinusoidal embedding
followed by a multilayer perceptron (MLP).
Parameters
----------
hidden_size : int
Output dimension of the MLP embedding.
frequency_embedding_size : int, optional
Size of the input frequency embedding, by default 256.
"""
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
[docs]
@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
[docs]
def forward(self, t):
"""
Forward pass for timestep embedding.
Parameters
----------
t : torch.Tensor
1D tensor of scalar timesteps.
Returns
-------
torch.Tensor
The final embedded representation of shape (N, hidden_size).
"""
t = t.view(-1)
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
t_emb = self.mlp(t_freq)
return t_emb
[docs]
class CategoricalEmbedder(nn.Module):
"""
Embeds categorical conditions (e.g., data source labels) into vector representations.
Supports label dropout for classifier-free guidance.
Parameters
----------
num_classes : int
Number of distinct label categories.
hidden_size : int
Size of the embedding vectors.
dropout_prob : float
Probability of label dropout.
"""
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
[docs]
def token_drop(self, labels, force_drop_ids=None):
"""
Drops labels to enable classifier-free guidance.
Parameters
----------
labels : torch.Tensor
Tensor of integer labels.
force_drop_ids : torch.Tensor or None, optional
Boolean mask to force specific labels to be dropped.
Returns
-------
torch.Tensor
Labels with some entries replaced by a dropout token.
"""
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
[docs]
def forward(self, labels, train, force_drop_ids=None):
"""
Forward pass for categorical embedding with optional label dropout.
Parameters
----------
labels : torch.Tensor
Tensor of categorical labels.
train : bool
Whether the model is in training mode.
force_drop_ids : torch.Tensor or None, optional
Explicit mask for which labels to drop.
Returns
-------
torch.Tensor
Embedded label representations, with optional noise added during training.
"""
labels = labels.long().view(-1)
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)
if True and train:
noise = torch.randn_like(embeddings)
embeddings = embeddings + noise
return embeddings
[docs]
class ClusterContinuousEmbedder(nn.Module):
"""
Embeds continuous input features into vector representations using a multilayer perceptron (MLP).
Supports optional embedding dropout for classifier-free guidance.
Parameters
----------
input_size : int
The size of the input features.
hidden_size : int
The size of the output embedding vectors.
dropout_prob : float
Probability of embedding dropout, used for classifier-free guidance.
"""
def __init__(self, input_size, hidden_size, dropout_prob):
super().__init__()
use_cfg_embedding = dropout_prob > 0
if use_cfg_embedding:
self.embedding_drop = nn.Embedding(1, hidden_size)
self.mlp = nn.Sequential(
nn.Linear(input_size, hidden_size, bias=True),
nn.Softmax(dim=1),
nn.Linear(hidden_size, hidden_size, bias=False)
)
self.hidden_size = hidden_size
self.dropout_prob = dropout_prob
[docs]
def forward(self, labels, train, force_drop_ids=None):
use_dropout = self.dropout_prob > 0
if force_drop_ids is not None:
drop_ids = force_drop_ids == 1
else:
drop_ids = None
if not use_dropout:
drop_ids = None
if (train and use_dropout):
drop_ids_rand = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
if force_drop_ids is not None:
drop_ids = torch.logical_or(drop_ids, drop_ids_rand)
else:
drop_ids = drop_ids_rand
if drop_ids is not None:
embeddings = torch.zeros((labels.shape[0], self.hidden_size), device=labels.device)
embeddings[~drop_ids] = self.mlp(labels[~drop_ids])
embeddings[drop_ids] += self.embedding_drop.weight[0]
else:
embeddings = self.mlp(labels)
if train:
noise = torch.randn_like(embeddings)
embeddings = embeddings + noise
return embeddings