|
|
import torch |
|
|
import torch.nn as nn |
|
|
from transformers import PreTrainedModel, T5ForConditionalGeneration, T5Config, AutoTokenizer |
|
|
from .configuration_imuru import ImuruConfig |
|
|
from diffusers import AutoencoderKL |
|
|
from einops.layers.torch import Rearrange |
|
|
from einops import repeat |
|
|
from torchvision.transforms import functional as F |
|
|
from typing import Optional, Tuple, List, Any |
|
|
from PIL import Image |
|
|
|
|
|
class Imuru(PreTrainedModel): |
|
|
""" |
|
|
Imuru is a conditional generative model that integrates a T5-based decoder with a VAE |
|
|
for image generation conditioned on text and style images. |
|
|
Attributes: |
|
|
config_class (Type): Configuration class for the model. |
|
|
tokenizer (AutoTokenizer): Tokenizer loaded from the provided tokenizer configuration. |
|
|
T5 (T5ForConditionalGeneration): T5 model adapted for conditional generation. |
|
|
sos (nn.Embedding): Start-of-sequence embedding. |
|
|
vae_to_t5 (nn.Linear): Linear projection from VAE latent space to T5 hidden space. |
|
|
t5_to_vae (nn.Linear): Linear projection from T5 hidden space back to VAE latent space. |
|
|
padding_token (nn.Parameter): Non-trainable parameter for padding tokens. |
|
|
padding_token_threshold (nn.Parameter): Non-trainable parameter for padding token threshold. |
|
|
vae (AutoencoderKL): Pre-trained Variational Autoencoder. |
|
|
query_rearrange (Rearrange): Layer to rearrange VAE latent representations for queries. |
|
|
z_rearrange (Rearrange): Layer to rearrange T5 outputs back to VAE latent dimensions. |
|
|
mse_criterion (nn.MSELoss): Mean squared error loss function. |
|
|
""" |
|
|
config_class = ImuruConfig |
|
|
|
|
|
def __init__(self, config: ImuruConfig) -> None: |
|
|
""" |
|
|
Initialize the Imuru model. |
|
|
Args: |
|
|
config (ImuruConfig): Configuration object containing model hyperparameters and paths. |
|
|
""" |
|
|
super().__init__(config) |
|
|
|
|
|
self.tokenizer = AutoTokenizer.from_pretrained(config.tokenizer_name_or_path) |
|
|
|
|
|
t5_config = T5Config.from_pretrained(config.t5_name_or_path) |
|
|
t5_config.vocab_size = len(self.tokenizer) |
|
|
self.T5 = T5ForConditionalGeneration(t5_config) |
|
|
self.T5.lm_head = nn.Identity() |
|
|
self.sos = nn.Embedding(1, t5_config.d_model) |
|
|
|
|
|
vae_latent_size = 8 * config.vae_channels * config.slices_per_query |
|
|
self.vae_to_t5 = nn.Linear(vae_latent_size, t5_config.d_model) |
|
|
self.t5_to_vae = nn.Linear(t5_config.d_model, vae_latent_size, bias=False) |
|
|
|
|
|
self.padding_token = nn.Parameter( torch.tensor([[-0.4951, 0.8021, 0.3429, 0.5622, 0.5271, 0.5756, 0.7194, 0.6150]]), requires_grad=False) |
|
|
self.padding_token_threshold = nn.Parameter(torch.tensor(0.484982096850872), requires_grad=False) |
|
|
|
|
|
self.query_rearrange = Rearrange('b c h (w q) -> b w (q c h)', q=config.slices_per_query) |
|
|
self.z_rearrange = Rearrange('b w (q c h) -> b c h (w q)', c=config.vae_channels, q=config.slices_per_query) |
|
|
|
|
|
self.style_enc = config.style_enc if hasattr(config, 'style_enc') else "mean" |
|
|
print(f"Using style encoder: {self.style_enc}") |
|
|
if self.style_enc == "MLP": |
|
|
self.style_encoder = nn.Linear(vae_latent_size, 1) |
|
|
elif self.style_enc == "MLP2": |
|
|
self.style_encoder = nn.Sequential( |
|
|
nn.Linear(vae_latent_size, vae_latent_size), |
|
|
nn.SiLU(), |
|
|
nn.Linear(vae_latent_size, 1) |
|
|
) |
|
|
|
|
|
self.mse_criterion = nn.MSELoss() |
|
|
self.init_weights() |
|
|
|
|
|
self.vae = AutoencoderKL.from_pretrained(config.vae_name_or_path) |
|
|
self.set_training(self.vae, False) |
|
|
|
|
|
def set_training(self, model: nn.Module, training: bool) -> None: |
|
|
""" |
|
|
Set the training mode for a given model and freeze/unfreeze parameters accordingly. |
|
|
Args: |
|
|
model (nn.Module): The model to set the training mode for. |
|
|
training (bool): If True, set the model to training mode; otherwise, evaluation mode. |
|
|
""" |
|
|
model.train() if training else model.eval() |
|
|
for param in model.parameters(): |
|
|
param.requires_grad = training |
|
|
|
|
|
def forward_nonAR( |
|
|
self, |
|
|
img: Optional[torch.Tensor] = None, |
|
|
input_ids: Optional[torch.Tensor] = None, |
|
|
attention_mask: Optional[torch.Tensor] = None, |
|
|
noise: float = 0, |
|
|
label_img: Optional[torch.Tensor] = None, |
|
|
**kwargs: Any |
|
|
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
|
|
""" |
|
|
Forward pass of the model. |
|
|
Args: |
|
|
img (Optional[torch.Tensor]): Input Style image tensor. |
|
|
input_ids (Optional[torch.Tensor]): Tokenized input IDs. |
|
|
attention_mask (Optional[torch.Tensor]): Attention mask for the inputs. |
|
|
noise (float): Amount of noise to add in image encoding. |
|
|
**kwargs: Additional arguments. |
|
|
Returns: |
|
|
Tuple containing: |
|
|
- mse_loss (torch.Tensor): Mean squared error loss. |
|
|
- pred_latent (torch.Tensor): Predicted latent representations. |
|
|
- z (torch.Tensor): Sampled latent vector from VAE. |
|
|
""" |
|
|
decoder_inputs_embeds, z_sequence, z = self._img_encode(img, noise) |
|
|
|
|
|
output = self.T5(input_ids, attention_mask=attention_mask, decoder_inputs_embeds=decoder_inputs_embeds) |
|
|
vae_latent = self.t5_to_vae(output.logits[:, :-1]) |
|
|
pred_latent = self.z_rearrange(vae_latent) |
|
|
|
|
|
if self.training: |
|
|
assert label_img is not None, 'label_img must be provided during training' |
|
|
posterior_label = self.vae.encode(label_img.float()) |
|
|
z_label = posterior_label.latent_dist.sample() |
|
|
z_label_sequence = self.query_rearrange(z_label) |
|
|
|
|
|
|
|
|
min_seq_len = min(vae_latent.size(1), z_label_sequence.size(1)) |
|
|
vae_latent_trimmed = vae_latent[:, :min_seq_len] |
|
|
z_label_sequence_trimmed = z_label_sequence[:, :min_seq_len] |
|
|
|
|
|
mse_loss = self.mse_criterion(vae_latent_trimmed, z_label_sequence_trimmed) |
|
|
else: |
|
|
mse_loss = torch.tensor(0.0, device=self.device) |
|
|
return mse_loss, pred_latent, z |
|
|
|
|
|
def forward( |
|
|
self, |
|
|
img: Optional[torch.Tensor] = None, |
|
|
label_img: Optional[torch.Tensor] = None, |
|
|
input_ids: Optional[torch.Tensor] = None, |
|
|
attention_mask: Optional[torch.Tensor] = None, |
|
|
style_noise: float = 0, |
|
|
label_noise: float = 0, |
|
|
**kwargs: Any |
|
|
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
|
|
""" |
|
|
Forward pass of the model. |
|
|
Args: |
|
|
img (Optional[torch.Tensor]): Input Style image tensor. |
|
|
input_ids (Optional[torch.Tensor]): Tokenized input IDs. |
|
|
attention_mask (Optional[torch.Tensor]): Attention mask for the inputs. |
|
|
noise (float): Amount of noise to add in image encoding. |
|
|
**kwargs: Additional arguments. |
|
|
Returns: |
|
|
Tuple containing: |
|
|
- mse_loss (torch.Tensor): Mean squared error loss. |
|
|
- pred_latent (torch.Tensor): Predicted latent representations. |
|
|
- z (torch.Tensor): Sampled latent vector from VAE. |
|
|
""" |
|
|
assert label_img is not None, 'label_img must be provided during training' |
|
|
|
|
|
posterior_style = self.vae.encode(img.float()) |
|
|
z_style = posterior_style.latent_dist.sample() |
|
|
z_style_sequence = self.query_rearrange(z_style) |
|
|
if style_noise > 0: |
|
|
z_style_sequence = z_style_sequence + torch.randn_like(z_style_sequence) * style_noise |
|
|
|
|
|
if self.style_enc == "mean": |
|
|
style_global = z_style_sequence.mean(dim=1, keepdim=True) |
|
|
elif self.style_enc in ["MLP", "MLP2"]: |
|
|
style_scores = self.style_encoder(z_style_sequence) |
|
|
style_weights = torch.softmax(style_scores, dim=1) |
|
|
style_global = (z_style_sequence * style_weights).sum(dim=1, keepdim=True) |
|
|
elif self.style_enc == "full": |
|
|
style_global = z_style_sequence |
|
|
else: |
|
|
raise ValueError(f"Unknown style_enc type: {self.style_enc}") |
|
|
|
|
|
w_style = style_global.size(1) |
|
|
style_token_embed = self.vae_to_t5(style_global) |
|
|
|
|
|
posterior_label = self.vae.encode(label_img.float()) |
|
|
z_label = posterior_label.latent_dist.sample() |
|
|
z_label_sequence = self.query_rearrange(z_label) |
|
|
|
|
|
if label_noise > 0: |
|
|
z_label_sequence_noisy = z_label_sequence + torch.randn_like(z_label_sequence) * label_noise |
|
|
else: |
|
|
z_label_sequence_noisy = z_label_sequence |
|
|
|
|
|
sos = repeat(self.sos.weight, '1 d -> b 1 d', b=input_ids.size(0)) |
|
|
|
|
|
label_embeds = self.vae_to_t5(z_label_sequence_noisy) |
|
|
decoder_inputs_embeds = torch.cat( |
|
|
[style_token_embed, sos, label_embeds[:, :-1]], dim=1 |
|
|
) |
|
|
|
|
|
output = self.T5(input_ids, attention_mask=attention_mask, decoder_inputs_embeds=decoder_inputs_embeds) |
|
|
all_vae_latent = self.t5_to_vae(output.logits) |
|
|
|
|
|
vae_latent = all_vae_latent[:, w_style:, :] |
|
|
|
|
|
min_seq_len = min(vae_latent.size(1), z_label_sequence.size(1)) |
|
|
vae_latent_trimmed = vae_latent[:, :min_seq_len] |
|
|
z_label_sequence_trimmed = z_label_sequence[:, :min_seq_len] |
|
|
|
|
|
mse_loss = self.mse_criterion(vae_latent_trimmed, z_label_sequence_trimmed) |
|
|
pred_latent = self.z_rearrange(vae_latent_trimmed) |
|
|
|
|
|
return mse_loss, pred_latent, z_label |
|
|
|
|
|
@torch.inference_mode() |
|
|
def generate( |
|
|
self, |
|
|
gen_text: str, |
|
|
style_img: torch.Tensor, |
|
|
**kwargs: Any |
|
|
) -> Image.Image: |
|
|
""" |
|
|
Generate an image by combining style and generation texts with a style image. |
|
|
Args: |
|
|
style_text (str): Style-related text prompt. |
|
|
gen_text (str): Generation-related text prompt. |
|
|
style_img (torch.Tensor): Style image tensor. Expected shape is either 3D or 4D. |
|
|
**kwargs: Additional keyword arguments. |
|
|
Returns: |
|
|
Image.Image: Generated image as a PIL image. |
|
|
""" |
|
|
if style_img.ndim == 3: |
|
|
style_img = style_img.unsqueeze(0) |
|
|
elif style_img.ndim == 4: |
|
|
pass |
|
|
else: |
|
|
raise ValueError('style_img must be 3D or 4D') |
|
|
|
|
|
imgs, _ = self._generate(texts=[gen_text], imgs=style_img, **kwargs) |
|
|
imgs = (imgs + 1) / 2 |
|
|
return F.to_pil_image(imgs[0].detach().cpu()) |
|
|
|
|
|
@torch.inference_mode() |
|
|
def generate_batch( |
|
|
self, |
|
|
gen_texts: List[str], |
|
|
style_imgs: torch.Tensor, |
|
|
**kwargs: Any |
|
|
) -> List[Image.Image]: |
|
|
""" |
|
|
Generate a batch of images from lists of style texts, generation texts, and style images. |
|
|
Args: |
|
|
style_texts (List[str]): List of style-related text prompts. |
|
|
gen_texts (List[str]): List of generation-related text prompts. |
|
|
style_imgs (torch.Tensor): Batch of style images (4D tensor). |
|
|
lengths (List[int]): List of lengths corresponding to each image. |
|
|
**kwargs: Additional keyword arguments. |
|
|
Returns: |
|
|
List[Image.Image]: List of generated images as PIL images. |
|
|
""" |
|
|
assert style_imgs.ndim == 4, 'style_imgs must be 4D' |
|
|
assert len(gen_texts) == len(style_imgs), 'gen_texts and style_imgs must have the same length' |
|
|
texts = [gen_text for gen_text in gen_texts] |
|
|
|
|
|
imgs, _ = self._generate(texts=texts, imgs=style_imgs, **kwargs) |
|
|
imgs = (imgs + 1) / 2 |
|
|
|
|
|
out_imgs = [] |
|
|
for i in range(imgs.size(0)): |
|
|
out_imgs.append(F.to_pil_image(imgs[i].detach().cpu())) |
|
|
return out_imgs |
|
|
|
|
|
def _generate( |
|
|
self, |
|
|
texts: Optional[List[str]] = None, |
|
|
imgs: Optional[torch.Tensor] = None, |
|
|
input_ids: Optional[torch.Tensor] = None, |
|
|
z_sequence: Optional[torch.Tensor] = None, |
|
|
max_new_tokens: int = 256, |
|
|
stopping_criteria: str = 'latent', |
|
|
stopping_after: int = 10, |
|
|
stopping_patience: int = 1 |
|
|
) -> Tuple[torch.Tensor, torch.Tensor]: |
|
|
""" |
|
|
Internal generation routine that combines textual and visual inputs to iteratively generate |
|
|
latent representations and decode them into images. |
|
|
Args: |
|
|
texts (Optional[List[str]]): List of text prompts. |
|
|
imgs (Optional[torch.Tensor]): Input image tensor. |
|
|
lengths (Optional[List[int]]): Desired lengths for each image in latent space. |
|
|
input_ids (Optional[torch.Tensor]): Tokenized input IDs. |
|
|
z_sequence (Optional[torch.Tensor]): Precomputed latent sequence. |
|
|
max_new_tokens (int): Maximum tokens to generate. |
|
|
stopping_criteria (str): Criteria for stopping ('latent' or 'none'). |
|
|
stopping_after (int): Number of tokens to check for stopping condition. |
|
|
stopping_patience (int): Patience parameter for stopping condition. |
|
|
Returns: |
|
|
Tuple containing: |
|
|
- imgs (torch.Tensor): Generated images. |
|
|
- canvas_sequence (torch.Tensor): Generated latent canvas sequence. |
|
|
- img_ends (torch.Tensor): End indices for each generated image. |
|
|
""" |
|
|
assert texts is not None or input_ids is not None, 'Either texts or input_ids must be provided' |
|
|
assert imgs is not None or z_sequence is not None, 'Either imgs or z_sequence must be provided' |
|
|
|
|
|
if input_ids is None: |
|
|
input_ids = self.tokenizer(texts, return_tensors='pt', padding=True).input_ids |
|
|
input_ids = input_ids.to(self.device) |
|
|
|
|
|
if z_sequence is None: |
|
|
posterior_style = self.vae.encode(imgs.float()) |
|
|
z_style = posterior_style.latent_dist.sample() |
|
|
z_style_sequence = self.query_rearrange(z_style) |
|
|
z_sequence = z_style_sequence |
|
|
|
|
|
if self.style_enc == "mean": |
|
|
style_global = z_sequence.mean(dim=1, keepdim=True) |
|
|
elif self.style_enc == "MLP" or self.style_enc == "MLP2": |
|
|
style_scores = self.style_encoder(z_sequence) |
|
|
style_weights = torch.softmax(style_scores, dim=1) |
|
|
style_global = (z_sequence * style_weights).sum(dim=1, keepdim=True) |
|
|
elif self.style_enc == "full": |
|
|
style_global = z_sequence |
|
|
else: |
|
|
raise ValueError(f"Unknown style_enc type: {self.style_enc}") |
|
|
|
|
|
w_style = style_global.size(1) |
|
|
style_token_embed = self.vae_to_t5(style_global) |
|
|
|
|
|
|
|
|
sos = repeat(self.sos.weight, '1 d -> b 1 d', b=input_ids.size(0)) |
|
|
pad_token = repeat(self.padding_token, '1 d -> b 1 d', b=input_ids.size(0)) |
|
|
|
|
|
generated_latents: List[torch.Tensor] = [] |
|
|
active = torch.ones(input_ids.size(0), dtype=torch.bool, device=self.device) |
|
|
|
|
|
for step in range(max_new_tokens): |
|
|
if len(generated_latents) == 0: |
|
|
decoder_inputs_embeds = torch.cat([style_token_embed, sos], dim=1) |
|
|
else: |
|
|
lat_seq = torch.stack(generated_latents, dim=1) |
|
|
lat_embeds = self.vae_to_t5(lat_seq) |
|
|
decoder_inputs_embeds = torch.cat([style_token_embed, sos, lat_embeds], dim=1) |
|
|
|
|
|
output = self.T5(input_ids, decoder_inputs_embeds=decoder_inputs_embeds) |
|
|
last_hidden = output.logits[:, -1:, :] |
|
|
vae_latent = self.t5_to_vae(last_hidden)[:, 0, :] |
|
|
|
|
|
if stopping_criteria == 'latent' and (~active).any(): |
|
|
vae_latent = torch.where( |
|
|
active.unsqueeze(-1), |
|
|
vae_latent, |
|
|
pad_token.squeeze(1) |
|
|
) |
|
|
|
|
|
generated_latents.append(vae_latent) |
|
|
canvas_sequence = torch.stack(generated_latents, dim=1) |
|
|
|
|
|
if stopping_criteria == 'latent': |
|
|
similarity = torch.nn.functional.cosine_similarity(canvas_sequence, pad_token, dim=-1) |
|
|
if similarity.size(1) >= stopping_after: |
|
|
window = similarity[:, -stopping_after:] |
|
|
cnt = (window > self.padding_token_threshold).to(torch.int).sum(dim=1) |
|
|
new = (cnt >= (stopping_after - stopping_patience)) & active |
|
|
active = active & (~new) |
|
|
|
|
|
if not active.any(): |
|
|
break |
|
|
|
|
|
elif stopping_criteria == 'none': |
|
|
pass |
|
|
|
|
|
canvas_sequence = torch.stack(generated_latents, dim=1) |
|
|
imgs = torch.clamp(self.vae.decode(self.z_rearrange(canvas_sequence)).sample, -1, 1) |
|
|
return imgs, canvas_sequence |
|
|
|
|
|
def _img_encode( |
|
|
self, |
|
|
img: torch.Tensor, |
|
|
noise: float = 0 |
|
|
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
|
|
""" |
|
|
Encode the input image into a latent representation using the VAE. |
|
|
Args: |
|
|
img (torch.Tensor): Input image tensor. |
|
|
noise (float): Standard deviation of noise to add to the latent sequence. |
|
|
Returns: |
|
|
Tuple containing: |
|
|
- decoder_inputs_embeds (torch.Tensor): Embeddings to be used as T5 decoder inputs. |
|
|
- z_sequence (torch.Tensor): Rearranged latent sequence from the VAE. |
|
|
- z (torch.Tensor): Sampled latent vector from the VAE. |
|
|
""" |
|
|
posterior = self.vae.encode(img.float()) |
|
|
z = posterior.latent_dist.sample() |
|
|
z_sequence = self.query_rearrange(z) |
|
|
|
|
|
noise_sequence = z_sequence |
|
|
if noise > 0: |
|
|
noise_sequence = z_sequence + torch.randn_like(z_sequence) * noise |
|
|
|
|
|
decoder_inputs_embeds = self.vae_to_t5(noise_sequence) |
|
|
sos = repeat(self.sos.weight, '1 d -> b 1 d', b=decoder_inputs_embeds.size(0)) |
|
|
decoder_inputs_embeds = torch.cat([sos, decoder_inputs_embeds], dim=1) |
|
|
return decoder_inputs_embeds, z_sequence, z |
|
|
|
|
|
def compute_padding_token(self) -> None: |
|
|
""" |
|
|
Compute and update the padding token. |
|
|
Raises: |
|
|
NotImplementedError: This method must be implemented. |
|
|
""" |
|
|
raise NotImplementedError("compute_padding_token not implemented") |
|
|
|
|
|
def compute_padding_token_threshold(self) -> None: |
|
|
""" |
|
|
Compute and update the padding token threshold. |
|
|
Raises: |
|
|
NotImplementedError: This method must be implemented. |
|
|
""" |
|
|
raise NotImplementedError("compute_padding_token_threshold not implemented") |
|
|
|