4 files changed, 961 insertions, 12 deletions

diff --git a/dreambooth.py b/dreambooth.py
index 775aea2..699313e 100644
--- a/dreambooth.py
+++ b/dreambooth.py
@@ -129,7 +129,7 @@ def parse_args():
     parser.add_argument(
         "--learning_rate",
         type=float,
-        default=1e-6,
+        default=5e-6,
         help="Initial learning rate (after the potential warmup period) to use.",
     )
     parser.add_argument(
@@ -167,7 +167,7 @@ def parse_args():
     parser.add_argument(
         "--ema_power",
         type=float,
-        default=7 / 8
+        default=6 / 7
     )
     parser.add_argument(
         "--ema_max_decay",
@@ -270,6 +270,11 @@ def parse_args():
         help="Max gradient norm."
     )
     parser.add_argument(
+        "--noise_timesteps",
+        type=int,
+        default=1000,
+    )
+    parser.add_argument(
         "--config",
         type=str,
         default=None,
@@ -480,7 +485,8 @@ def main():
         unet,
         inv_gamma=args.ema_inv_gamma,
         power=args.ema_power,
-        max_value=args.ema_max_decay
+        max_value=args.ema_max_decay,
+        device=accelerator.device
     ) if args.use_ema else None
 
     if args.gradient_checkpointing:
@@ -523,7 +529,7 @@ def main():
         beta_start=0.00085,
         beta_end=0.012,
         beta_schedule="scaled_linear",
-        num_train_timesteps=1000
+        num_train_timesteps=args.noise_timesteps
     )
 
     def collate_fn(examples):
@@ -632,7 +638,6 @@ def main():
     # Move text_encoder and vae to device
     text_encoder.to(accelerator.device)
     vae.to(accelerator.device)
-    ema_unet.averaged_model.to(accelerator.device)
 
     # Keep text_encoder and vae in eval mode as we don't train these
     text_encoder.eval()
diff --git a/dreambooth_plus.py b/dreambooth_plus.py
new file mode 100644
index 0000000..9e482b3
--- /dev/null
+++ b/dreambooth_plus.py
@@ -0,0 +1,939 @@
+import argparse
+import itertools
+import math
+import os
+import datetime
+import logging
+import json
+from pathlib import Path
+
+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, PNDMScheduler, UNet2DConditionModel
+from diffusers.optimization import get_scheduler
+from diffusers.training_utils import EMAModel
+from PIL import Image
+from tqdm.auto import tqdm
+from transformers import CLIPTextModel, CLIPTokenizer
+from slugify import slugify
+
+from schedulers.scheduling_euler_a import EulerAScheduler
+from pipelines.stable_diffusion.vlpn_stable_diffusion import VlpnStableDiffusion
+from data.csv import CSVDataModule
+
+logger = get_logger(__name__)
+
+
+torch.backends.cuda.matmul.allow_tf32 = True
+
+
+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_file",
+        type=str,
+        default=None,
+        help="A CSV file 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(
+        "--num_class_images",
+        type=int,
+        default=400,
+        help="How many class images to generate per training image."
+    )
+    parser.add_argument(
+        "--repeats",
+        type=int,
+        default=1,
+        help="How many times to repeat the training data."
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="output/dreambooth-plus",
+        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=3000,
+        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_unet",
+        type=float,
+        default=1e-5,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--learning_rate_text",
+        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="cosine",
+        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(
+        "--use_ema",
+        action="store_true",
+        default=True,
+        help="Whether to use EMA model."
+    )
+    parser.add_argument(
+        "--ema_inv_gamma",
+        type=float,
+        default=1.0
+    )
+    parser.add_argument(
+        "--ema_power",
+        type=float,
+        default=6 / 7
+    )
+    parser.add_argument(
+        "--ema_max_decay",
+        type=float,
+        default=0.9999
+    )
+    parser.add_argument(
+        "--use_8bit_adam",
+        action="store_true",
+        default=True,
+        help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    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(
+        "--sample_frequency",
+        type=int,
+        default=100,
+        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(
+        "--sample_batches",
+        type=int,
+        default=1,
+        help="Number of 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=30,
+        help="Number of steps for sample generation. Higher values will result in more detailed samples, but longer runtimes.",
+    )
+    parser.add_argument(
+        "--prior_loss_weight",
+        type=float,
+        default=1.0,
+        help="The weight of prior preservation loss."
+    )
+    parser.add_argument(
+        "--max_grad_norm",
+        default=1.0,
+        type=float,
+        help="Max gradient norm."
+    )
+    parser.add_argument(
+        "--noise_timesteps",
+        type=int,
+        default=1000,
+    )
+    parser.add_argument(
+        "--config",
+        type=str,
+        default=None,
+        help="Path to a JSON configuration file containing arguments for invoking this script."
+    )
+
+    args = parser.parse_args()
+    if 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_file is None:
+        raise ValueError("You must specify --train_data_file")
+
+    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 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,
+        ema_unet,
+        tokenizer,
+        text_encoder,
+        placeholder_token,
+        placeholder_token_id,
+        output_dir: Path,
+        sample_image_size,
+        sample_batches,
+        sample_batch_size,
+        seed
+    ):
+        self.datamodule = datamodule
+        self.accelerator = accelerator
+        self.vae = vae
+        self.unet = unet
+        self.ema_unet = ema_unet
+        self.tokenizer = tokenizer
+        self.text_encoder = text_encoder
+        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 or torch.random.seed()
+        self.sample_batches = sample_batches
+        self.sample_batch_size = sample_batch_size
+
+    @torch.no_grad()
+    def checkpoint(self):
+        print("Saving model...")
+
+        unwrapped_unet = self.accelerator.unwrap_model(
+            self.ema_unet.averaged_model if self.ema_unet is not None else self.unet)
+        unwrapped_text_encoder = self.accelerator.unwrap_model(self.text_encoder)
+        pipeline = VlpnStableDiffusion(
+            text_encoder=unwrapped_text_encoder,
+            vae=self.vae,
+            unet=unwrapped_unet,
+            tokenizer=self.tokenizer,
+            scheduler=PNDMScheduler(
+                beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", skip_prk_steps=True
+            ),
+        )
+        pipeline.save_pretrained(self.output_dir.joinpath("model"))
+
+        del unwrapped_unet
+        del unwrapped_text_encoder
+        del pipeline
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+
+    @torch.no_grad()
+    def save_samples(self, step, height, width, guidance_scale, eta, num_inference_steps):
+        samples_path = Path(self.output_dir).joinpath("samples")
+
+        unwrapped_unet = self.accelerator.unwrap_model(
+            self.ema_unet.averaged_model if self.ema_unet is not None else self.unet)
+        unwrapped_text_encoder = self.accelerator.unwrap_model(self.text_encoder)
+        scheduler = EulerAScheduler(
+            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
+        )
+
+        # Save a sample image
+        pipeline = VlpnStableDiffusion(
+            text_encoder=unwrapped_text_encoder,
+            vae=self.vae,
+            unet=unwrapped_unet,
+            tokenizer=self.tokenizer,
+            scheduler=scheduler,
+        ).to(self.accelerator.device)
+        pipeline.set_progress_bar_config(dynamic_ncols=True)
+
+        train_data = self.datamodule.train_dataloader()
+        val_data = self.datamodule.val_dataloader()
+
+        generator = torch.Generator(device=pipeline.device).manual_seed(self.seed)
+        stable_latents = torch.randn(
+            (self.sample_batch_size, pipeline.unet.in_channels, height // 8, width // 8),
+            device=pipeline.device,
+            generator=generator,
+        )
+
+        with torch.inference_mode():
+            for pool, data, latents in [("stable", val_data, stable_latents), ("val", val_data, None), ("train", train_data, None)]:
+                all_samples = []
+                file_path = samples_path.joinpath(pool, f"step_{step}.png")
+                file_path.parent.mkdir(parents=True, exist_ok=True)
+
+                data_enum = enumerate(data)
+
+                for i in range(self.sample_batches):
+                    batches = [batch for j, batch in data_enum if j * data.batch_size < self.sample_batch_size]
+                    prompt = [prompt.format(self.placeholder_token)
+                              for batch in batches for prompt in batch["prompts"]][:self.sample_batch_size]
+                    nprompt = [prompt for batch in batches for prompt in batch["nprompts"]][:self.sample_batch_size]
+
+                    samples = pipeline(
+                        prompt=prompt,
+                        negative_prompt=nprompt,
+                        height=self.sample_image_size,
+                        width=self.sample_image_size,
+                        latents=latents[:len(prompt)] if latents is not None else None,
+                        generator=generator if latents is not None else None,
+                        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.sample_batches, self.sample_batch_size)
+                image_grid.save(file_path)
+
+                del all_samples
+                del image_grid
+
+        del unwrapped_unet
+        del unwrapped_text_encoder
+        del scheduler
+        del pipeline
+        del generator
+        del stable_latents
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+
+
+def main():
+    args = parse_args()
+
+    global_step_offset = 0
+    now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
+    basepath = Path(args.output_dir).joinpath(slugify(args.placeholder_token), now)
+    basepath.mkdir(parents=True, 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
+    )
+
+    logging.basicConfig(filename=basepath.joinpath("log.txt"), level=logging.DEBUG)
+
+    # 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, subfolder='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 len(initializer_token_ids) > 1:
+        raise ValueError(
+            f"initializer_token_ids must not have more than 1 vector, but it's {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, subfolder='text_encoder')
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder='vae')
+    unet = UNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder='unet')
+
+    ema_unet = EMAModel(
+        unet,
+        inv_gamma=args.ema_inv_gamma,
+        power=args.ema_power,
+        max_value=args.ema_max_decay,
+        device=accelerator.device
+    ) if args.use_ema else None
+
+    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)
+    token_embeds[placeholder_token_id] = initializer_token_embeddings
+
+    # Freeze vae and unet
+    freeze_params(vae.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_unet = (
+            args.learning_rate_unet * args.gradient_accumulation_steps *
+            args.train_batch_size * accelerator.num_processes
+        )
+        args.learning_rate_text = (
+            args.learning_rate_text * args.gradient_accumulation_steps *
+            args.train_batch_size * accelerator.num_processes
+        )
+
+    # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
+    if args.use_8bit_adam:
+        try:
+            import bitsandbytes as bnb
+        except ImportError:
+            raise ImportError("To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`.")
+
+        optimizer_class = bnb.optim.AdamW8bit
+    else:
+        optimizer_class = torch.optim.AdamW
+
+    # Initialize the optimizer
+    optimizer = optimizer_class(
+        [
+            {
+                'params': unet.parameters(),
+                'lr': args.learning_rate_unet,
+            },
+            {
+                'params': text_encoder.get_input_embeddings().parameters(),
+                'lr': args.learning_rate_text,
+            }
+        ],
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    noise_scheduler = DDPMScheduler(
+        beta_start=0.00085,
+        beta_end=0.012,
+        beta_schedule="scaled_linear",
+        num_train_timesteps=args.noise_timesteps
+    )
+
+    def collate_fn(examples):
+        prompts = [example["prompts"] for example in examples]
+        nprompts = [example["nprompts"] for example in examples]
+        input_ids = [example["instance_prompt_ids"] for example in examples]
+        pixel_values = [example["instance_images"] for example in examples]
+
+        # concat class and instance examples for prior preservation
+        if args.num_class_images != 0 and "class_prompt_ids" in examples[0]:
+            input_ids += [example["class_prompt_ids"] for example in examples]
+            pixel_values += [example["class_images"] for example in examples]
+
+        pixel_values = torch.stack(pixel_values)
+        pixel_values = pixel_values.to(dtype=torch.float32, memory_format=torch.contiguous_format)
+
+        input_ids = tokenizer.pad({"input_ids": input_ids}, padding=True, return_tensors="pt").input_ids
+
+        batch = {
+            "prompts": prompts,
+            "nprompts": nprompts,
+            "input_ids": input_ids,
+            "pixel_values": pixel_values,
+        }
+        return batch
+
+    datamodule = CSVDataModule(
+        data_file=args.train_data_file,
+        batch_size=args.train_batch_size,
+        tokenizer=tokenizer,
+        instance_identifier=args.placeholder_token,
+        class_identifier=args.initializer_token,
+        class_subdir="cls",
+        num_class_images=args.num_class_images,
+        size=args.resolution,
+        repeats=args.repeats,
+        center_crop=args.center_crop,
+        valid_set_size=args.sample_batch_size*args.sample_batches,
+        collate_fn=collate_fn
+    )
+
+    datamodule.prepare_data()
+    datamodule.setup()
+
+    if args.num_class_images != 0:
+        missing_data = [item for item in datamodule.data_train if not item.class_image_path.exists()]
+
+        if len(missing_data) != 0:
+            batched_data = [missing_data[i:i+args.sample_batch_size]
+                            for i in range(0, len(missing_data), args.sample_batch_size)]
+
+            scheduler = EulerAScheduler(
+                beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
+            )
+
+            pipeline = VlpnStableDiffusion(
+                text_encoder=text_encoder,
+                vae=vae,
+                unet=unet,
+                tokenizer=tokenizer,
+                scheduler=scheduler,
+            ).to(accelerator.device)
+            pipeline.set_progress_bar_config(dynamic_ncols=True)
+
+            with torch.inference_mode():
+                for batch in batched_data:
+                    image_name = [p.class_image_path for p in batch]
+                    prompt = [p.prompt.format(args.initializer_token) for p in batch]
+                    nprompt = [p.nprompt for p in batch]
+
+                    images = pipeline(
+                        prompt=prompt,
+                        negative_prompt=nprompt,
+                        num_inference_steps=args.sample_steps
+                    ).images
+
+                    for i, image in enumerate(images):
+                        image.save(image_name[i])
+
+            del pipeline
+
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+    train_dataloader = datamodule.train_dataloader()
+    val_dataloader = datamodule.val_dataloader()
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_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, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler = accelerator.prepare(
+        text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler
+    )
+
+    # Move vae and unet to device
+    vae.to(accelerator.device)
+
+    # Keep vae and unet in eval mode as we don't train these
+    vae.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) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+
+    num_val_steps_per_epoch = len(val_dataloader)
+    num_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+    val_steps = num_val_steps_per_epoch * num_epochs
+
+    # 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("dreambooth_plus", 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 = {num_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
+
+    checkpointer = Checkpointer(
+        datamodule=datamodule,
+        accelerator=accelerator,
+        vae=vae,
+        unet=unet,
+        ema_unet=ema_unet,
+        tokenizer=tokenizer,
+        text_encoder=text_encoder,
+        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,
+        sample_batches=args.sample_batches,
+        seed=args.seed
+    )
+
+    if accelerator.is_main_process:
+        checkpointer.save_samples(
+            0,
+            args.resolution, args.resolution, 7.5, 0.0, args.sample_steps)
+
+    local_progress_bar = tqdm(
+        range(num_update_steps_per_epoch + num_val_steps_per_epoch),
+        disable=not accelerator.is_local_main_process,
+        dynamic_ncols=True
+    )
+    local_progress_bar.set_description("Epoch X / Y")
+
+    global_progress_bar = tqdm(
+        range(args.max_train_steps + val_steps),
+        disable=not accelerator.is_local_main_process,
+        dynamic_ncols=True
+    )
+    global_progress_bar.set_description("Total progress")
+
+    try:
+        for epoch in range(num_epochs):
+            local_progress_bar.set_description(f"Epoch {epoch + 1} / {num_epochs}")
+            local_progress_bar.reset()
+
+            text_encoder.train()
+            train_loss = 0.0
+
+            sample_checkpoint = False
+
+            for step, batch in enumerate(train_dataloader):
+                with accelerator.accumulate(itertools.chain(unet, text_encoder)):
+                    # Convert images to latent space
+                    latents = vae.encode(batch["pixel_values"]).latent_dist.sample()
+                    latents = latents * 0.18215
+
+                    # 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.config.num_train_timesteps,
+                                              (bsz,), device=latents.device)
+                    timesteps = timesteps.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
+
+                    if args.num_class_images != 0:
+                        # Chunk the noise and noise_pred into two parts and compute the loss on each part separately.
+                        noise_pred, noise_pred_prior = torch.chunk(noise_pred, 2, dim=0)
+                        noise, noise_prior = torch.chunk(noise, 2, dim=0)
+
+                        # Compute instance loss
+                        loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
+
+                        # Compute prior loss
+                        prior_loss = F.mse_loss(noise_pred_prior, noise_prior, reduction="none").mean([1, 2, 3]).mean()
+
+                        # Add the prior loss to the instance loss.
+                        loss = loss + args.prior_loss_weight * prior_loss
+                    else:
+                        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)
+
+                    if accelerator.sync_gradients:
+                        accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
+
+                    optimizer.step()
+                    if not accelerator.optimizer_step_was_skipped:
+                        lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=True)
+
+                    loss = loss.detach().item()
+                    train_loss += loss
+
+                # Checks if the accelerator has performed an optimization step behind the scenes
+                if accelerator.sync_gradients:
+                    if args.use_ema:
+                        ema_unet.step(unet)
+
+                    local_progress_bar.update(1)
+                    global_progress_bar.update(1)
+
+                    global_step += 1
+
+                    if global_step % args.sample_frequency == 0:
+                        sample_checkpoint = True
+
+                logs = {
+                    "train/loss": loss,
+                    "lr/unet": lr_scheduler.get_last_lr()[0],
+                    "lr/text": lr_scheduler.get_last_lr()[1]
+                }
+                if args.use_ema:
+                    logs["ema_decay"] = ema_unet.decay
+
+                accelerator.log(logs, step=global_step)
+
+                local_progress_bar.set_postfix(**logs)
+
+                if global_step >= args.max_train_steps:
+                    break
+
+            train_loss /= len(train_dataloader)
+
+            accelerator.wait_for_everyone()
+
+            text_encoder.eval()
+            val_loss = 0.0
+
+            for step, batch in enumerate(val_dataloader):
+                with torch.no_grad():
+                    latents = vae.encode(batch["pixel_values"]).latent_dist.sample()
+                    latents = latents * 0.18215
+
+                    noise = torch.randn(latents.shape).to(latents.device)
+                    bsz = latents.shape[0]
+                    timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps,
+                                              (bsz,), device=latents.device)
+                    timesteps = timesteps.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:
+                    local_progress_bar.update(1)
+                    global_progress_bar.update(1)
+
+                logs = {"val/loss": loss}
+                local_progress_bar.set_postfix(**logs)
+
+            val_loss /= len(val_dataloader)
+
+            accelerator.log({"val/loss": val_loss}, step=global_step)
+
+            local_progress_bar.clear()
+            global_progress_bar.clear()
+
+            if min_val_loss > val_loss:
+                min_val_loss = val_loss
+
+            if sample_checkpoint and accelerator.is_main_process:
+                checkpointer.save_samples(
+                    global_step + global_step_offset,
+                    args.resolution, args.resolution, 7.5, 0.0, args.sample_steps)
+
+        # 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()
+
+            accelerator.end_training()
+
+    except KeyboardInterrupt:
+        if accelerator.is_main_process:
+            print("Interrupted, saving checkpoint and resume state...")
+            checkpointer.checkpoint()
+            accelerator.end_training()
+        quit()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/infer.py b/infer.py
index 5bd4abc..63b16d8 100644
--- a/infer.py
+++ b/infer.py
@@ -205,10 +205,10 @@ def load_embeddings(tokenizer, text_encoder, embeddings_dir):
 def create_pipeline(model, scheduler, embeddings_dir, dtype):
     print("Loading Stable Diffusion pipeline...")
 
-    tokenizer = CLIPTokenizer.from_pretrained(model, subfolder='/tokenizer', torch_dtype=dtype)
-    text_encoder = CLIPTextModel.from_pretrained(model, subfolder='/text_encoder', torch_dtype=dtype)
-    vae = AutoencoderKL.from_pretrained(model, subfolder='/vae', torch_dtype=dtype)
-    unet = UNet2DConditionModel.from_pretrained(model, subfolder='/unet', torch_dtype=dtype)
+    tokenizer = CLIPTokenizer.from_pretrained(model, subfolder='tokenizer', torch_dtype=dtype)
+    text_encoder = CLIPTextModel.from_pretrained(model, subfolder='text_encoder', torch_dtype=dtype)
+    vae = AutoencoderKL.from_pretrained(model, subfolder='vae', torch_dtype=dtype)
+    unet = UNet2DConditionModel.from_pretrained(model, subfolder='unet', torch_dtype=dtype)
 
     load_embeddings(tokenizer, text_encoder, embeddings_dir)
 
diff --git a/textual_inversion.py b/textual_inversion.py
index 3a3741d..181a318 100644
--- a/textual_inversion.py
+++ b/textual_inversion.py
@@ -111,7 +111,7 @@ def parse_args():
     parser.add_argument(
         "--max_train_steps",
         type=int,
-        default=3000,
+        default=10000,
         help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
     )
     parser.add_argument(
@@ -128,7 +128,7 @@ def parse_args():
     parser.add_argument(
         "--learning_rate",
         type=float,
-        default=5e-5,
+        default=1e-4,
         help="Initial learning rate (after the potential warmup period) to use.",
     )
     parser.add_argument(
@@ -247,6 +247,11 @@ def parse_args():
         help="The weight of prior preservation loss."
     )
     parser.add_argument(
+        "--noise_timesteps",
+        type=int,
+        default=1000,
+    )
+    parser.add_argument(
         "--resume_from",
         type=str,
         default=None,
@@ -568,7 +573,7 @@ def main():
         beta_start=0.00085,
         beta_end=0.012,
         beta_schedule="scaled_linear",
-        num_train_timesteps=1000
+        num_train_timesteps=args.noise_timesteps
     )
 
     def collate_fn(examples):