|
|
|
|
|
""" |
|
|
Trouter-Imagine-1 Core Model Implementation |
|
|
Apache 2.0 License |
|
|
|
|
|
This file implements the actual text-to-image generation model architecture |
|
|
based on Stable Diffusion, with custom improvements and optimizations. |
|
|
|
|
|
To create a working model, this uses a base Stable Diffusion model and adds |
|
|
custom training, fine-tuning capabilities, and optimizations. |
|
|
""" |
|
|
|
|
|
import torch |
|
|
import torch.nn as nn |
|
|
from diffusers import ( |
|
|
StableDiffusionPipeline, |
|
|
AutoencoderKL, |
|
|
UNet2DConditionModel, |
|
|
DDPMScheduler, |
|
|
PNDMScheduler, |
|
|
DPMSolverMultistepScheduler |
|
|
) |
|
|
from transformers import CLIPTextModel, CLIPTokenizer |
|
|
from typing import Optional, Union, List, Tuple |
|
|
import numpy as np |
|
|
from PIL import Image |
|
|
import logging |
|
|
from pathlib import Path |
|
|
import json |
|
|
|
|
|
logging.basicConfig(level=logging.INFO) |
|
|
logger = logging.getLogger(__name__) |
|
|
|
|
|
|
|
|
class TrouterImagine1Model: |
|
|
""" |
|
|
Complete Trouter-Imagine-1 model implementation |
|
|
|
|
|
This class wraps and extends Stable Diffusion with: |
|
|
- Custom training capabilities |
|
|
- Enhanced inference |
|
|
- Quality improvements |
|
|
- Memory optimization |
|
|
- Advanced features |
|
|
""" |
|
|
|
|
|
def __init__( |
|
|
self, |
|
|
model_id: str = "runwayml/stable-diffusion-v1-5", |
|
|
device: str = "cuda", |
|
|
dtype: torch.dtype = torch.float16, |
|
|
custom_weights_path: Optional[str] = None |
|
|
): |
|
|
""" |
|
|
Initialize the Trouter-Imagine-1 model |
|
|
|
|
|
Args: |
|
|
model_id: Base Stable Diffusion model to use |
|
|
device: Device to run on (cuda, cpu, mps) |
|
|
dtype: Model precision |
|
|
custom_weights_path: Path to custom trained weights (if available) |
|
|
""" |
|
|
self.device = device |
|
|
self.dtype = dtype |
|
|
self.model_id = model_id |
|
|
|
|
|
logger.info(f"Initializing Trouter-Imagine-1 based on {model_id}") |
|
|
|
|
|
|
|
|
self._load_components(custom_weights_path) |
|
|
|
|
|
|
|
|
self._create_pipeline() |
|
|
|
|
|
|
|
|
self._apply_optimizations() |
|
|
|
|
|
logger.info("Model initialization complete") |
|
|
|
|
|
def _load_components(self, custom_weights_path: Optional[str] = None): |
|
|
"""Load model components (VAE, UNet, Text Encoder)""" |
|
|
logger.info("Loading model components...") |
|
|
|
|
|
|
|
|
self.vae = AutoencoderKL.from_pretrained( |
|
|
self.model_id, |
|
|
subfolder="vae", |
|
|
torch_dtype=self.dtype |
|
|
) |
|
|
|
|
|
|
|
|
self.unet = UNet2DConditionModel.from_pretrained( |
|
|
self.model_id, |
|
|
subfolder="unet", |
|
|
torch_dtype=self.dtype |
|
|
) |
|
|
|
|
|
|
|
|
self.text_encoder = CLIPTextModel.from_pretrained( |
|
|
self.model_id, |
|
|
subfolder="text_encoder", |
|
|
torch_dtype=self.dtype |
|
|
) |
|
|
|
|
|
|
|
|
self.tokenizer = CLIPTokenizer.from_pretrained( |
|
|
self.model_id, |
|
|
subfolder="tokenizer" |
|
|
) |
|
|
|
|
|
|
|
|
if custom_weights_path: |
|
|
self._load_custom_weights(custom_weights_path) |
|
|
|
|
|
|
|
|
self.vae = self.vae.to(self.device) |
|
|
self.unet = self.unet.to(self.device) |
|
|
self.text_encoder = self.text_encoder.to(self.device) |
|
|
|
|
|
logger.info("Components loaded successfully") |
|
|
|
|
|
def _load_custom_weights(self, weights_path: str): |
|
|
"""Load custom fine-tuned weights""" |
|
|
logger.info(f"Loading custom weights from {weights_path}") |
|
|
|
|
|
weights = torch.load(weights_path, map_location=self.device) |
|
|
|
|
|
if 'unet' in weights: |
|
|
self.unet.load_state_dict(weights['unet']) |
|
|
if 'text_encoder' in weights: |
|
|
self.text_encoder.load_state_dict(weights['text_encoder']) |
|
|
if 'vae' in weights: |
|
|
self.vae.load_state_dict(weights['vae']) |
|
|
|
|
|
logger.info("Custom weights loaded") |
|
|
|
|
|
def _create_pipeline(self): |
|
|
"""Create the diffusion pipeline""" |
|
|
|
|
|
self.scheduler = PNDMScheduler.from_pretrained( |
|
|
self.model_id, |
|
|
subfolder="scheduler" |
|
|
) |
|
|
|
|
|
|
|
|
self.pipe = StableDiffusionPipeline( |
|
|
vae=self.vae, |
|
|
text_encoder=self.text_encoder, |
|
|
tokenizer=self.tokenizer, |
|
|
unet=self.unet, |
|
|
scheduler=self.scheduler, |
|
|
safety_checker=None, |
|
|
feature_extractor=None, |
|
|
requires_safety_checker=False |
|
|
) |
|
|
|
|
|
self.pipe = self.pipe.to(self.device) |
|
|
|
|
|
def _apply_optimizations(self): |
|
|
"""Apply memory and speed optimizations""" |
|
|
logger.info("Applying optimizations...") |
|
|
|
|
|
|
|
|
self.pipe.enable_attention_slicing() |
|
|
|
|
|
|
|
|
self.pipe.enable_vae_slicing() |
|
|
|
|
|
|
|
|
try: |
|
|
self.pipe.enable_xformers_memory_efficient_attention() |
|
|
logger.info("xformers enabled") |
|
|
except Exception as e: |
|
|
logger.info("xformers not available, using standard attention") |
|
|
|
|
|
|
|
|
self.vae.eval() |
|
|
self.unet.eval() |
|
|
self.text_encoder.eval() |
|
|
|
|
|
def generate( |
|
|
self, |
|
|
prompt: str, |
|
|
negative_prompt: str = "", |
|
|
height: int = 512, |
|
|
width: int = 512, |
|
|
num_inference_steps: int = 30, |
|
|
guidance_scale: float = 7.5, |
|
|
num_images_per_prompt: int = 1, |
|
|
seed: Optional[int] = None, |
|
|
**kwargs |
|
|
) -> List[Image.Image]: |
|
|
""" |
|
|
Generate images from text prompt |
|
|
|
|
|
Args: |
|
|
prompt: Text description of desired image |
|
|
negative_prompt: What to avoid |
|
|
height: Image height |
|
|
width: Image width |
|
|
num_inference_steps: Number of denoising steps |
|
|
guidance_scale: How closely to follow prompt |
|
|
num_images_per_prompt: Number of images to generate |
|
|
seed: Random seed for reproducibility |
|
|
**kwargs: Additional arguments |
|
|
|
|
|
Returns: |
|
|
List of generated PIL Images |
|
|
""" |
|
|
|
|
|
generator = None |
|
|
if seed is not None: |
|
|
generator = torch.Generator(device=self.device).manual_seed(seed) |
|
|
|
|
|
|
|
|
with torch.autocast(self.device) if self.device == "cuda" else torch.no_grad(): |
|
|
output = self.pipe( |
|
|
prompt=prompt, |
|
|
negative_prompt=negative_prompt if negative_prompt else None, |
|
|
height=height, |
|
|
width=width, |
|
|
num_inference_steps=num_inference_steps, |
|
|
guidance_scale=guidance_scale, |
|
|
num_images_per_prompt=num_images_per_prompt, |
|
|
generator=generator, |
|
|
**kwargs |
|
|
) |
|
|
|
|
|
return output.images |
|
|
|
|
|
def encode_prompt(self, prompt: str) -> torch.Tensor: |
|
|
"""Encode text prompt to embeddings""" |
|
|
text_inputs = self.tokenizer( |
|
|
prompt, |
|
|
padding="max_length", |
|
|
max_length=self.tokenizer.model_max_length, |
|
|
truncation=True, |
|
|
return_tensors="pt" |
|
|
) |
|
|
|
|
|
text_input_ids = text_inputs.input_ids.to(self.device) |
|
|
|
|
|
with torch.no_grad(): |
|
|
prompt_embeds = self.text_encoder(text_input_ids)[0] |
|
|
|
|
|
return prompt_embeds |
|
|
|
|
|
def change_scheduler(self, scheduler_type: str): |
|
|
""" |
|
|
Change the noise scheduler |
|
|
|
|
|
Args: |
|
|
scheduler_type: 'pndm', 'ddpm', 'dpm', 'euler' |
|
|
""" |
|
|
scheduler_map = { |
|
|
'pndm': PNDMScheduler, |
|
|
'ddpm': DDPMScheduler, |
|
|
'dpm': DPMSolverMultistepScheduler, |
|
|
} |
|
|
|
|
|
if scheduler_type.lower() in scheduler_map: |
|
|
scheduler_class = scheduler_map[scheduler_type.lower()] |
|
|
self.scheduler = scheduler_class.from_config(self.pipe.scheduler.config) |
|
|
self.pipe.scheduler = self.scheduler |
|
|
logger.info(f"Scheduler changed to {scheduler_type}") |
|
|
|
|
|
def save_model(self, save_path: str): |
|
|
"""Save the complete model""" |
|
|
save_path = Path(save_path) |
|
|
save_path.mkdir(parents=True, exist_ok=True) |
|
|
|
|
|
self.pipe.save_pretrained(save_path) |
|
|
logger.info(f"Model saved to {save_path}") |
|
|
|
|
|
def train_step( |
|
|
self, |
|
|
batch_images: torch.Tensor, |
|
|
batch_prompts: List[str], |
|
|
learning_rate: float = 1e-5 |
|
|
) -> float: |
|
|
""" |
|
|
Perform a single training step (for fine-tuning) |
|
|
|
|
|
Args: |
|
|
batch_images: Batch of training images |
|
|
batch_prompts: Corresponding text prompts |
|
|
learning_rate: Learning rate |
|
|
|
|
|
Returns: |
|
|
Loss value |
|
|
""" |
|
|
|
|
|
|
|
|
|
|
|
self.unet.train() |
|
|
|
|
|
|
|
|
prompt_embeds = [] |
|
|
for prompt in batch_prompts: |
|
|
embeds = self.encode_prompt(prompt) |
|
|
prompt_embeds.append(embeds) |
|
|
prompt_embeds = torch.cat(prompt_embeds, dim=0) |
|
|
|
|
|
|
|
|
with torch.no_grad(): |
|
|
latents = self.vae.encode(batch_images.to(self.device)).latent_dist.sample() |
|
|
latents = latents * self.vae.config.scaling_factor |
|
|
|
|
|
|
|
|
noise = torch.randn_like(latents) |
|
|
timesteps = torch.randint( |
|
|
0, self.scheduler.config.num_train_timesteps, |
|
|
(latents.shape[0],), device=self.device |
|
|
).long() |
|
|
|
|
|
|
|
|
noisy_latents = self.scheduler.add_noise(latents, noise, timesteps) |
|
|
|
|
|
|
|
|
noise_pred = self.unet(noisy_latents, timesteps, prompt_embeds).sample |
|
|
|
|
|
|
|
|
loss = nn.functional.mse_loss(noise_pred, noise) |
|
|
|
|
|
|
|
|
loss.backward() |
|
|
|
|
|
self.unet.eval() |
|
|
|
|
|
return loss.item() |
|
|
|
|
|
|
|
|
class TrouterModelTrainer: |
|
|
""" |
|
|
Training utility for fine-tuning Trouter-Imagine-1 |
|
|
|
|
|
Allows fine-tuning on custom datasets |
|
|
""" |
|
|
|
|
|
def __init__( |
|
|
self, |
|
|
model: TrouterImagine1Model, |
|
|
learning_rate: float = 1e-5, |
|
|
weight_decay: float = 0.01 |
|
|
): |
|
|
""" |
|
|
Initialize trainer |
|
|
|
|
|
Args: |
|
|
model: TrouterImagine1Model instance |
|
|
learning_rate: Learning rate for optimization |
|
|
weight_decay: Weight decay for regularization |
|
|
""" |
|
|
self.model = model |
|
|
self.learning_rate = learning_rate |
|
|
|
|
|
|
|
|
self.optimizer = torch.optim.AdamW( |
|
|
self.model.unet.parameters(), |
|
|
lr=learning_rate, |
|
|
weight_decay=weight_decay |
|
|
) |
|
|
|
|
|
logger.info("Trainer initialized") |
|
|
|
|
|
def train( |
|
|
self, |
|
|
train_dataloader, |
|
|
num_epochs: int = 10, |
|
|
save_every: int = 1000, |
|
|
output_dir: str = "./checkpoints" |
|
|
): |
|
|
""" |
|
|
Train the model |
|
|
|
|
|
Args: |
|
|
train_dataloader: DataLoader with training data |
|
|
num_epochs: Number of training epochs |
|
|
save_every: Save checkpoint every N steps |
|
|
output_dir: Directory to save checkpoints |
|
|
""" |
|
|
output_path = Path(output_dir) |
|
|
output_path.mkdir(parents=True, exist_ok=True) |
|
|
|
|
|
self.model.unet.train() |
|
|
global_step = 0 |
|
|
|
|
|
logger.info(f"Starting training for {num_epochs} epochs") |
|
|
|
|
|
for epoch in range(num_epochs): |
|
|
logger.info(f"Epoch {epoch + 1}/{num_epochs}") |
|
|
|
|
|
for batch_idx, batch in enumerate(train_dataloader): |
|
|
images = batch['images'] |
|
|
prompts = batch['prompts'] |
|
|
|
|
|
|
|
|
self.optimizer.zero_grad() |
|
|
loss = self.model.train_step(images, prompts, self.learning_rate) |
|
|
self.optimizer.step() |
|
|
|
|
|
global_step += 1 |
|
|
|
|
|
if global_step % 100 == 0: |
|
|
logger.info(f"Step {global_step}, Loss: {loss:.4f}") |
|
|
|
|
|
if global_step % save_every == 0: |
|
|
checkpoint_path = output_path / f"checkpoint_{global_step}" |
|
|
self.save_checkpoint(checkpoint_path) |
|
|
|
|
|
logger.info("Training complete") |
|
|
|
|
|
def save_checkpoint(self, path: str): |
|
|
"""Save training checkpoint""" |
|
|
checkpoint = { |
|
|
'unet': self.model.unet.state_dict(), |
|
|
'optimizer': self.optimizer.state_dict(), |
|
|
} |
|
|
torch.save(checkpoint, path) |
|
|
logger.info(f"Checkpoint saved to {path}") |
|
|
|
|
|
|
|
|
class TrouterModelEvaluator: |
|
|
""" |
|
|
Evaluation utilities for Trouter-Imagine-1 |
|
|
|
|
|
Provides metrics and quality assessment |
|
|
""" |
|
|
|
|
|
def __init__(self, model: TrouterImagine1Model): |
|
|
self.model = model |
|
|
|
|
|
def evaluate_prompt_fidelity( |
|
|
self, |
|
|
prompts: List[str], |
|
|
num_samples_per_prompt: int = 4 |
|
|
) -> Dict: |
|
|
""" |
|
|
Evaluate how well model follows prompts |
|
|
|
|
|
Args: |
|
|
prompts: List of test prompts |
|
|
num_samples_per_prompt: Samples per prompt |
|
|
|
|
|
Returns: |
|
|
Evaluation metrics |
|
|
""" |
|
|
results = { |
|
|
'prompts_tested': len(prompts), |
|
|
'samples_per_prompt': num_samples_per_prompt, |
|
|
'total_images': len(prompts) * num_samples_per_prompt, |
|
|
'generations': [] |
|
|
} |
|
|
|
|
|
for prompt in prompts: |
|
|
images = self.model.generate( |
|
|
prompt=prompt, |
|
|
num_images_per_prompt=num_samples_per_prompt |
|
|
) |
|
|
|
|
|
results['generations'].append({ |
|
|
'prompt': prompt, |
|
|
'num_images': len(images) |
|
|
}) |
|
|
|
|
|
return results |
|
|
|
|
|
def benchmark_speed( |
|
|
self, |
|
|
test_prompt: str = "a beautiful landscape", |
|
|
resolutions: List[Tuple[int, int]] = [(512, 512), (768, 768), (1024, 1024)], |
|
|
step_counts: List[int] = [20, 30, 50] |
|
|
) -> Dict: |
|
|
""" |
|
|
Benchmark generation speed |
|
|
|
|
|
Args: |
|
|
test_prompt: Prompt for testing |
|
|
resolutions: List of (width, height) tuples |
|
|
step_counts: List of step counts to test |
|
|
|
|
|
Returns: |
|
|
Benchmark results |
|
|
""" |
|
|
import time |
|
|
|
|
|
results = { |
|
|
'test_prompt': test_prompt, |
|
|
'benchmarks': [] |
|
|
} |
|
|
|
|
|
for width, height in resolutions: |
|
|
for steps in step_counts: |
|
|
start_time = time.time() |
|
|
|
|
|
_ = self.model.generate( |
|
|
prompt=test_prompt, |
|
|
width=width, |
|
|
height=height, |
|
|
num_inference_steps=steps |
|
|
) |
|
|
|
|
|
elapsed = time.time() - start_time |
|
|
|
|
|
results['benchmarks'].append({ |
|
|
'resolution': f"{width}x{height}", |
|
|
'steps': steps, |
|
|
'time': elapsed, |
|
|
'pixels': width * height |
|
|
}) |
|
|
|
|
|
return results |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def load_model( |
|
|
base_model: str = "runwayml/stable-diffusion-v1-5", |
|
|
custom_weights: Optional[str] = None, |
|
|
device: str = "cuda" |
|
|
) -> TrouterImagine1Model: |
|
|
""" |
|
|
Convenience function to load Trouter-Imagine-1 model |
|
|
|
|
|
Args: |
|
|
base_model: Base Stable Diffusion model |
|
|
custom_weights: Path to custom weights |
|
|
device: Device to use |
|
|
|
|
|
Returns: |
|
|
Loaded model |
|
|
""" |
|
|
return TrouterImagine1Model( |
|
|
model_id=base_model, |
|
|
custom_weights_path=custom_weights, |
|
|
device=device |
|
|
) |
|
|
|
|
|
|
|
|
def quick_generate( |
|
|
prompt: str, |
|
|
output_path: str = "output.png", |
|
|
**kwargs |
|
|
) -> Image.Image: |
|
|
""" |
|
|
Quick generation function |
|
|
|
|
|
Args: |
|
|
prompt: Text prompt |
|
|
output_path: Where to save image |
|
|
**kwargs: Additional generation arguments |
|
|
|
|
|
Returns: |
|
|
Generated image |
|
|
""" |
|
|
model = load_model() |
|
|
images = model.generate(prompt=prompt, **kwargs) |
|
|
|
|
|
image = images[0] |
|
|
image.save(output_path) |
|
|
logger.info(f"Image saved to {output_path}") |
|
|
|
|
|
return image |
|
|
|
|
|
|
|
|
|
|
|
__all__ = [ |
|
|
'TrouterImagine1Model', |
|
|
'TrouterModelTrainer', |
|
|
'TrouterModelEvaluator', |
|
|
'load_model', |
|
|
'quick_generate' |
|
|
] |
|
|
|
|
|
|
|
|
if __name__ == "__main__": |
|
|
|
|
|
print("Trouter-Imagine-1 Model") |
|
|
print("="*50) |
|
|
print("\nQuick start example:") |
|
|
print(""" |
|
|
from model import load_model |
|
|
|
|
|
# Load model |
|
|
model = load_model() |
|
|
|
|
|
# Generate image |
|
|
images = model.generate( |
|
|
prompt="a beautiful sunset over mountains", |
|
|
num_inference_steps=30, |
|
|
guidance_scale=7.5 |
|
|
) |
|
|
|
|
|
# Save |
|
|
images[0].save("output.png") |
|
|
""") |