import argparse import itertools import math import os import datetime import numpy as np import torch import torch.nn.functional as F import torch.utils.checkpoint from accelerate import Accelerator from accelerate.logging import get_logger from accelerate.utils import LoggerType, set_seed from diffusers import AutoencoderKL, DDPMScheduler, LMSDiscreteScheduler, StableDiffusionPipeline, UNet2DConditionModel from diffusers.optimization import get_scheduler from pipelines.stable_diffusion.no_check import NoCheck from PIL import Image from tqdm.auto import tqdm from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer from slugify import slugify import json import os from data.textual_inversion.csv import CSVDataModule logger = get_logger(__name__) def parse_args(): parser = argparse.ArgumentParser( description="Simple example of a training script." ) parser.add_argument( "--pretrained_model_name_or_path", type=str, default=None, help="Path to pretrained model or model identifier from huggingface.co/models.", ) parser.add_argument( "--tokenizer_name", type=str, default=None, help="Pretrained tokenizer name or path if not the same as model_name", ) parser.add_argument( "--train_data_dir", type=str, default=None, help="A folder containing the training data." ) parser.add_argument( "--placeholder_token", type=str, default=None, help="A token to use as a placeholder for the concept.", ) parser.add_argument( "--initializer_token", type=str, default=None, help="A token to use as initializer word." ) parser.add_argument( "--vectors_per_token", type=int, default=1, help="Vectors per token." ) parser.add_argument( "--repeats", type=int, default=100, help="How many times to repeat the training data.") parser.add_argument( "--output_dir", type=str, default="text-inversion-model", help="The output directory where the model predictions and checkpoints will be written.", ) parser.add_argument( "--seed", type=int, default=None, help="A seed for reproducible training.") parser.add_argument( "--resolution", type=int, default=512, help=( "The resolution for input images, all the images in the train/validation dataset will be resized to this" " resolution" ), ) parser.add_argument( "--center_crop", action="store_true", help="Whether to center crop images before resizing to resolution" ) parser.add_argument( "--num_train_epochs", type=int, default=100) parser.add_argument( "--max_train_steps", type=int, default=5000, help="Total number of training steps to perform. If provided, overrides num_train_epochs.", ) parser.add_argument( "--gradient_accumulation_steps", type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.", ) parser.add_argument( "--gradient_checkpointing", action="store_true", help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", ) parser.add_argument( "--learning_rate", type=float, default=1e-4, help="Initial learning rate (after the potential warmup period) to use.", ) parser.add_argument( "--scale_lr", action="store_true", default=True, help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", ) parser.add_argument( "--lr_scheduler", type=str, default="constant", help=( 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' ' "constant", "constant_with_warmup"]' ), ) parser.add_argument( "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." ) parser.add_argument( "--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer." ) parser.add_argument( "--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer." ) parser.add_argument( "--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use." ) parser.add_argument( "--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer" ) parser.add_argument( "--mixed_precision", type=str, default="no", choices=["no", "fp16", "bf16"], help=( "Whether to use mixed precision. Choose" "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10." "and an Nvidia Ampere GPU." ), ) parser.add_argument( "--local_rank", type=int, default=-1, help="For distributed training: local_rank" ) parser.add_argument( "--checkpoint_frequency", type=int, default=500, help="How often to save a checkpoint and sample image", ) parser.add_argument( "--sample_image_size", type=int, default=512, help="Size of sample images", ) parser.add_argument( "--stable_sample_batches", type=int, default=1, help="Number of fixed seed sample batches to generate per checkpoint", ) parser.add_argument( "--random_sample_batches", type=int, default=1, help="Number of random seed sample batches to generate per checkpoint", ) parser.add_argument( "--sample_batch_size", type=int, default=1, help="Number of samples to generate per batch", ) parser.add_argument( "--train_batch_size", type=int, default=1, help="Batch size (per device) for the training dataloader." ) parser.add_argument( "--sample_steps", type=int, default=50, help="Number of steps for sample generation. Higher values will result in more detailed samples, but longer runtimes.", ) parser.add_argument( "--resume_from", type=str, default=None, help="Path to a directory to resume training from (ie, logs/token_name/2022-09-22T23-36-27)" ) parser.add_argument( "--resume_checkpoint", type=str, default=None, help="Path to a specific checkpoint to resume training from (ie, logs/token_name/2022-09-22T23-36-27/checkpoints/something.bin)." ) parser.add_argument( "--config", type=str, default=None, help="Path to a JSON configuration file containing arguments for invoking this script. If resume_from is given, its resume.json takes priority over this." ) args = parser.parse_args() if args.resume_from is not None: with open(f"{args.resume_from}/resume.json", 'rt') as f: args = parser.parse_args( namespace=argparse.Namespace(**json.load(f)["args"])) elif args.config is not None: with open(args.config, 'rt') as f: args = parser.parse_args( namespace=argparse.Namespace(**json.load(f)["args"])) env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) if env_local_rank != -1 and env_local_rank != args.local_rank: args.local_rank = env_local_rank if args.train_data_dir is None: raise ValueError("You must specify --train_data_dir") if args.pretrained_model_name_or_path is None: raise ValueError("You must specify --pretrained_model_name_or_path") if args.placeholder_token is None: raise ValueError("You must specify --placeholder_token") if args.initializer_token is None: raise ValueError("You must specify --initializer_token") if args.output_dir is None: raise ValueError("You must specify --output_dir") return args def freeze_params(params): for param in params: param.requires_grad = False def save_resume_file(basepath, args, extra={}): info = {"args": vars(args)} info["args"].update(extra) with open(f"{basepath}/resume.json", "w") as f: json.dump(info, f, indent=4) def make_grid(images, rows, cols): w, h = images[0].size grid = Image.new('RGB', size=(cols*w, rows*h)) for i, image in enumerate(images): grid.paste(image, box=(i % cols*w, i//cols*h)) return grid class Checkpointer: def __init__( self, datamodule, accelerator, vae, unet, tokenizer, placeholder_token, placeholder_token_id, output_dir, sample_image_size, random_sample_batches, sample_batch_size, stable_sample_batches, seed ): self.datamodule = datamodule self.accelerator = accelerator self.vae = vae self.unet = unet self.tokenizer = tokenizer self.placeholder_token = placeholder_token self.placeholder_token_id = placeholder_token_id self.output_dir = output_dir self.sample_image_size = sample_image_size self.seed = seed self.random_sample_batches = random_sample_batches self.sample_batch_size = sample_batch_size self.stable_sample_batches = stable_sample_batches @torch.no_grad() def checkpoint(self, step, postfix, text_encoder, save_samples=True, path=None): print("Saving checkpoint for step %d..." % step) with self.accelerator.autocast(): if path is None: checkpoints_path = f"{self.output_dir}/checkpoints" os.makedirs(checkpoints_path, exist_ok=True) unwrapped = self.accelerator.unwrap_model(text_encoder) # Save a checkpoint learned_embeds = unwrapped.get_input_embeddings().weight[self.placeholder_token_id] learned_embeds_dict = {self.placeholder_token: learned_embeds.detach().cpu()} filename = f"%s_%d_%s.bin" % (slugify(self.placeholder_token), step, postfix) if path is not None: torch.save(learned_embeds_dict, path) else: torch.save(learned_embeds_dict, f"{checkpoints_path}/{filename}") torch.save(learned_embeds_dict, f"{checkpoints_path}/last.bin") del unwrapped del learned_embeds @torch.no_grad() def save_samples(self, mode, step, text_encoder, height, width, guidance_scale, eta, num_inference_steps): samples_path = f"{self.output_dir}/samples/{mode}" os.makedirs(samples_path, exist_ok=True) checker = NoCheck() unwrapped = self.accelerator.unwrap_model(text_encoder) # Save a sample image pipeline = StableDiffusionPipeline( text_encoder=unwrapped, vae=self.vae, unet=self.unet, tokenizer=self.tokenizer, scheduler=LMSDiscreteScheduler( beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear" ), safety_checker=NoCheck(), feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"), ).to(self.accelerator.device) pipeline.enable_attention_slicing() data = { "training": self.datamodule.train_dataloader(), "validation": self.datamodule.val_dataloader(), }[mode] if mode == "validation" and self.stable_sample_batches > 0 and step > 0: stable_latents = torch.randn( (self.sample_batch_size, pipeline.unet.in_channels, height // 8, width // 8), device=pipeline.device, generator=torch.Generator(device=pipeline.device).manual_seed(self.seed), ) all_samples = [] filename = f"stable_step_%d.png" % (step) data_enum = enumerate(data) # Generate and save stable samples for i in range(0, self.stable_sample_batches): prompt = [prompt for i, batch in data_enum for j, prompt in enumerate( batch["prompt"]) if i * data.batch_size + j < self.sample_batch_size] with self.accelerator.autocast(): samples = pipeline( prompt=prompt, height=self.sample_image_size, latents=stable_latents[:len(prompt)], width=self.sample_image_size, guidance_scale=guidance_scale, eta=eta, num_inference_steps=num_inference_steps, output_type='pil' )["sample"] all_samples += samples del samples image_grid = make_grid(all_samples, self.stable_sample_batches, self.sample_batch_size) image_grid.save(f"{samples_path}/{filename}") del all_samples del image_grid del stable_latents all_samples = [] filename = f"step_%d.png" % (step) data_enum = enumerate(data) # Generate and save random samples for i in range(0, self.random_sample_batches): prompt = [prompt for i, batch in data_enum for j, prompt in enumerate( batch["prompt"]) if i * data.batch_size + j < self.sample_batch_size] with self.accelerator.autocast(): samples = pipeline( prompt=prompt, height=self.sample_image_size, width=self.sample_image_size, guidance_scale=guidance_scale, eta=eta, num_inference_steps=num_inference_steps, output_type='pil' )["sample"] all_samples += samples del samples image_grid = make_grid(all_samples, self.random_sample_batches, self.sample_batch_size) image_grid.save(f"{samples_path}/{filename}") del all_samples del image_grid del checker del unwrapped del pipeline torch.cuda.empty_cache() class ImageToLatents(): def __init__(self, vae): self.vae = vae self.encoded_pixel_values_cache = {} @torch.no_grad() def __call__(self, batch): key = "|".join(batch["key"]) if self.encoded_pixel_values_cache.get(key, None) is None: self.encoded_pixel_values_cache[key] = self.vae.encode(batch["pixel_values"]).latent_dist latents = self.encoded_pixel_values_cache[key].sample().detach().half() * 0.18215 return latents def main(): args = parse_args() global_step_offset = 0 if args.resume_from is not None: basepath = f"{args.resume_from}" print("Resuming state from %s" % args.resume_from) with open(f"{basepath}/resume.json", 'r') as f: state = json.load(f) global_step_offset = state["args"].get("global_step", 0) print("We've trained %d steps so far" % global_step_offset) else: now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S") basepath = f"{args.output_dir}/{slugify(args.placeholder_token)}/{now}" os.makedirs(basepath, exist_ok=True) accelerator = Accelerator( log_with=LoggerType.TENSORBOARD, logging_dir=f"{basepath}", gradient_accumulation_steps=args.gradient_accumulation_steps, mixed_precision=args.mixed_precision ) # If passed along, set the training seed now. if args.seed is not None: set_seed(args.seed) # Load the tokenizer and add the placeholder token as a additional special token if args.tokenizer_name: tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name) elif args.pretrained_model_name_or_path: tokenizer = CLIPTokenizer.from_pretrained( args.pretrained_model_name_or_path + '/tokenizer' ) # Add the placeholder token in tokenizer num_added_tokens = tokenizer.add_tokens(args.placeholder_token) if num_added_tokens == 0: raise ValueError( f"The tokenizer already contains the token {args.placeholder_token}. Please pass a different" " `placeholder_token` that is not already in the tokenizer." ) # Convert the initializer_token, placeholder_token to ids initializer_token_ids = tokenizer.encode(args.initializer_token, add_special_tokens=False) # Check if initializer_token is a single token or a sequence of tokens if args.vectors_per_token % len(initializer_token_ids) != 0: raise ValueError( f"vectors_per_token ({args.vectors_per_token}) must be divisible by initializer token ({len(initializer_token_ids)}).") initializer_token_ids = torch.tensor(initializer_token_ids) placeholder_token_id = tokenizer.convert_tokens_to_ids(args.placeholder_token) # Load models and create wrapper for stable diffusion text_encoder = CLIPTextModel.from_pretrained( args.pretrained_model_name_or_path + '/text_encoder', ) vae = AutoencoderKL.from_pretrained( args.pretrained_model_name_or_path + '/vae', ) unet = UNet2DConditionModel.from_pretrained( args.pretrained_model_name_or_path + '/unet', ) if args.gradient_checkpointing: unet.enable_gradient_checkpointing() slice_size = unet.config.attention_head_dim // 2 unet.set_attention_slice(slice_size) # Resize the token embeddings as we are adding new special tokens to the tokenizer text_encoder.resize_token_embeddings(len(tokenizer)) # Initialise the newly added placeholder token with the embeddings of the initializer token token_embeds = text_encoder.get_input_embeddings().weight.data initializer_token_embeddings = text_encoder.get_input_embeddings()(initializer_token_ids) if args.resume_checkpoint is not None: token_embeds[placeholder_token_id] = torch.load(args.resume_checkpoint)[ args.placeholder_token] else: token_embeds[placeholder_token_id] = initializer_token_embeddings # Freeze vae and unet freeze_params(vae.parameters()) freeze_params(unet.parameters()) # Freeze all parameters except for the token embeddings in text encoder params_to_freeze = itertools.chain( text_encoder.text_model.encoder.parameters(), text_encoder.text_model.final_layer_norm.parameters(), text_encoder.text_model.embeddings.position_embedding.parameters(), ) freeze_params(params_to_freeze) if args.scale_lr: args.learning_rate = ( args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes ) # Initialize the optimizer optimizer = torch.optim.AdamW( text_encoder.get_input_embeddings().parameters(), # only optimize the embeddings lr=args.learning_rate, betas=(args.adam_beta1, args.adam_beta2), weight_decay=args.adam_weight_decay, eps=args.adam_epsilon, ) # TODO (patil-suraj): laod scheduler using args noise_scheduler = DDPMScheduler( beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000, tensor_format="pt" ) datamodule = CSVDataModule( data_root=args.train_data_dir, batch_size=args.train_batch_size, tokenizer=tokenizer, size=args.resolution, placeholder_token=args.placeholder_token, repeats=args.repeats, center_crop=args.center_crop) datamodule.prepare_data() datamodule.setup() train_dataloader = datamodule.train_dataloader() val_dataloader = datamodule.val_dataloader() checkpointer = Checkpointer( datamodule=datamodule, accelerator=accelerator, vae=vae, unet=unet, tokenizer=tokenizer, placeholder_token=args.placeholder_token, placeholder_token_id=placeholder_token_id, output_dir=basepath, sample_image_size=args.sample_image_size, sample_batch_size=args.sample_batch_size, random_sample_batches=args.random_sample_batches, stable_sample_batches=args.stable_sample_batches, seed=args.seed ) # Scheduler and math around the number of training steps. overrode_max_train_steps = False num_update_steps_per_epoch = math.ceil( (len(train_dataloader) + len(val_dataloader)) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch overrode_max_train_steps = True lr_scheduler = get_scheduler( args.lr_scheduler, optimizer=optimizer, num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, ) text_encoder, optimizer, train_dataloader, val_dataloader, lr_scheduler = accelerator.prepare( text_encoder, optimizer, train_dataloader, val_dataloader, lr_scheduler ) # Move vae and unet to device vae.to(accelerator.device) unet.to(accelerator.device) # Keep vae and unet in eval mode as we don't train these vae.eval() unet.eval() # We need to recalculate our total training steps as the size of the training dataloader may have changed. num_update_steps_per_epoch = math.ceil( (len(train_dataloader) + len(val_dataloader)) / args.gradient_accumulation_steps) if overrode_max_train_steps: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch # Afterwards we recalculate our number of training epochs args.num_train_epochs = math.ceil( args.max_train_steps / num_update_steps_per_epoch) # We need to initialize the trackers we use, and also store our configuration. # The trackers initializes automatically on the main process. if accelerator.is_main_process: accelerator.init_trackers("textual_inversion", config=vars(args)) # Train! total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") # Only show the progress bar once on each machine. global_step = 0 min_val_loss = np.inf imageToLatents = ImageToLatents(vae) checkpointer.save_samples( "validation", 0, text_encoder, args.resolution, args.resolution, 7.5, 0.0, args.sample_steps) progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process) progress_bar.set_description("Global steps") local_progress_bar = tqdm(range(num_update_steps_per_epoch), disable=not accelerator.is_local_main_process) local_progress_bar.set_description("Steps") try: for epoch in range(args.num_train_epochs): local_progress_bar.reset() text_encoder.train() train_loss = 0.0 for step, batch in enumerate(train_dataloader): with accelerator.accumulate(text_encoder): with accelerator.autocast(): # Convert images to latent space latents = imageToLatents(batch) # Sample noise that we'll add to the latents noise = torch.randn(latents.shape).to(latents.device) bsz = latents.shape[0] # Sample a random timestep for each image timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device).long() # Add noise to the latents according to the noise magnitude at each timestep # (this is the forward diffusion process) noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) # Get the text embedding for conditioning encoder_hidden_states = text_encoder(batch["input_ids"])[0] # Predict the noise residual noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean() accelerator.backward(loss) # Zero out the gradients for all token embeddings except the newly added # embeddings for the concept, as we only want to optimize the concept embeddings if accelerator.num_processes > 1: grads = text_encoder.module.get_input_embeddings().weight.grad else: grads = text_encoder.get_input_embeddings().weight.grad # Get the index for tokens that we want to zero the grads for index_grads_to_zero = torch.arange(len(tokenizer)) != placeholder_token_id grads.data[index_grads_to_zero, :] = grads.data[index_grads_to_zero, :].fill_(0) optimizer.step() if not accelerator.optimizer_step_was_skipped: lr_scheduler.step() optimizer.zero_grad() loss = loss.detach().item() train_loss += loss # Checks if the accelerator has performed an optimization step behind the scenes if accelerator.sync_gradients: progress_bar.update(1) local_progress_bar.update(1) global_step += 1 if global_step % args.checkpoint_frequency == 0 and global_step > 0 and accelerator.is_main_process: progress_bar.clear() local_progress_bar.clear() checkpointer.checkpoint(global_step + global_step_offset, "training", text_encoder) save_resume_file(basepath, args, { "global_step": global_step + global_step_offset, "resume_checkpoint": f"{basepath}/checkpoints/last.bin" }) checkpointer.save_samples( "training", global_step + global_step_offset, text_encoder, args.resolution, args.resolution, 7.5, 0.0, args.sample_steps) logs = {"mode": "training", "loss": loss, "lr": lr_scheduler.get_last_lr()[0]} local_progress_bar.set_postfix(**logs) if global_step >= args.max_train_steps: break train_loss /= len(train_dataloader) text_encoder.eval() val_loss = 0.0 for step, batch in enumerate(val_dataloader): with torch.no_grad(), accelerator.autocast(): latents = imageToLatents(batch) noise = torch.randn(latents.shape).to(latents.device) bsz = latents.shape[0] timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device).long() noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) encoder_hidden_states = text_encoder(batch["input_ids"])[0] noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample noise_pred, noise = accelerator.gather_for_metrics((noise_pred, noise)) loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean() loss = loss.detach().item() val_loss += loss if accelerator.sync_gradients: progress_bar.update(1) local_progress_bar.update(1) logs = {"mode": "validation", "loss": loss} local_progress_bar.set_postfix(**logs) val_loss /= len(val_dataloader) accelerator.log({"train/loss": train_loss, "val/loss": val_loss}, step=global_step) progress_bar.clear() local_progress_bar.clear() if min_val_loss > val_loss: accelerator.print(f"Validation loss reached new minimum: {min_val_loss:.2e} -> {val_loss:.2e}") checkpointer.checkpoint(global_step + global_step_offset, "milestone", text_encoder) min_val_loss = val_loss checkpointer.save_samples( "validation", global_step + global_step_offset, text_encoder, args.resolution, args.resolution, 7.5, 0.0, args.sample_steps) accelerator.wait_for_everyone() # Create the pipeline using using the trained modules and save it. if accelerator.is_main_process: print("Finished! Saving final checkpoint and resume state.") checkpointer.checkpoint( global_step + global_step_offset, "end", text_encoder, path=f"{basepath}/learned_embeds.bin" ) save_resume_file(basepath, args, { "global_step": global_step + global_step_offset, "resume_checkpoint": f"{basepath}/checkpoints/last.bin" }) accelerator.end_training() except KeyboardInterrupt: if accelerator.is_main_process: print("Interrupted, saving checkpoint and resume state...") checkpointer.checkpoint(global_step + global_step_offset, "end", text_encoder) save_resume_file(basepath, args, { "global_step": global_step + global_step_offset, "resume_checkpoint": f"{basepath}/checkpoints/last.bin" }) accelerator.end_training() quit() if __name__ == "__main__": main()