path: root/modules/textual_inversion/textual_inversion.py

import os
import sys
import traceback

import torch
import tqdm
import html
import datetime
import csv

from PIL import Image, PngImagePlugin

from modules import shared, devices, sd_hijack, processing, sd_models, images
import modules.textual_inversion.dataset
from modules.textual_inversion.learn_schedule import LearnRateScheduler

from modules.textual_inversion.image_embedding import (embedding_to_b64, embedding_from_b64,
                                                       insert_image_data_embed, extract_image_data_embed,
                                                       caption_image_overlay)

class Embedding:
    def __init__(self, vec, name, step=None):
        self.vec = vec
        self.name = name
        self.step = step
        self.cached_checksum = None
        self.sd_checkpoint = None
        self.sd_checkpoint_name = None

    def save(self, filename):
        embedding_data = {
            "string_to_token": {"*": 265},
            "string_to_param": {"*": self.vec},
            "name": self.name,
            "step": self.step,
            "sd_checkpoint": self.sd_checkpoint,
            "sd_checkpoint_name": self.sd_checkpoint_name,
        }

        torch.save(embedding_data, filename)

    def checksum(self):
        if self.cached_checksum is not None:
            return self.cached_checksum

        def const_hash(a):
            r = 0
            for v in a:
                r = (r * 281 ^ int(v) * 997) & 0xFFFFFFFF
            return r

        self.cached_checksum = f'{const_hash(self.vec.reshape(-1) * 100) & 0xffff:04x}'
        return self.cached_checksum


class EmbeddingDatabase:
    def __init__(self, embeddings_dir):
        self.ids_lookup = {}
        self.word_embeddings = {}
        self.dir_mtime = None
        self.embeddings_dir = embeddings_dir

    def register_embedding(self, embedding, model):

        self.word_embeddings[embedding.name] = embedding

        ids = model.cond_stage_model.tokenizer([embedding.name], add_special_tokens=False)['input_ids'][0]

        first_id = ids[0]
        if first_id not in self.ids_lookup:
            self.ids_lookup[first_id] = []

        self.ids_lookup[first_id] = sorted(self.ids_lookup[first_id] + [(ids, embedding)], key=lambda x: len(x[0]), reverse=True)

        return embedding

    def load_textual_inversion_embeddings(self):
        mt = os.path.getmtime(self.embeddings_dir)
        if self.dir_mtime is not None and mt <= self.dir_mtime:
            return

        self.dir_mtime = mt
        self.ids_lookup.clear()
        self.word_embeddings.clear()

        def process_file(path, filename):
            name = os.path.splitext(filename)[0]

            data = []

            if os.path.splitext(filename.upper())[-1] in ['.PNG', '.WEBP', '.JXL', '.AVIF']:
                embed_image = Image.open(path)
                if hasattr(embed_image, 'text') and 'sd-ti-embedding' in embed_image.text:
                    data = embedding_from_b64(embed_image.text['sd-ti-embedding'])
                    name = data.get('name', name)
                else:
                    data = extract_image_data_embed(embed_image)
                    name = data.get('name', name)
            else:
                data = torch.load(path, map_location="cpu")

            # textual inversion embeddings
            if 'string_to_param' in data:
                param_dict = data['string_to_param']
                if hasattr(param_dict, '_parameters'):
                    param_dict = getattr(param_dict, '_parameters')  # fix for torch 1.12.1 loading saved file from torch 1.11
                assert len(param_dict) == 1, 'embedding file has multiple terms in it'
                emb = next(iter(param_dict.items()))[1]
            # diffuser concepts
            elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor:
                assert len(data.keys()) == 1, 'embedding file has multiple terms in it'

                emb = next(iter(data.values()))
                if len(emb.shape) == 1:
                    emb = emb.unsqueeze(0)
            else:
                raise Exception(f"Couldn't identify {filename} as neither textual inversion embedding nor diffuser concept.")

            vec = emb.detach().to(devices.device, dtype=torch.float32)
            embedding = Embedding(vec, name)
            embedding.step = data.get('step', None)
            embedding.sd_checkpoint = data.get('hash', None)
            embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None)
            self.register_embedding(embedding, shared.sd_model)

        for fn in os.listdir(self.embeddings_dir):
            try:
                fullfn = os.path.join(self.embeddings_dir, fn)

                if os.stat(fullfn).st_size == 0:
                    continue

                process_file(fullfn, fn)
            except Exception:
                print(f"Error loading emedding {fn}:", file=sys.stderr)
                print(traceback.format_exc(), file=sys.stderr)
                continue

        print(f"Loaded a total of {len(self.word_embeddings)} textual inversion embeddings.")
        print("Embeddings:", ', '.join(self.word_embeddings.keys()))

    def find_embedding_at_position(self, tokens, offset):
        token = tokens[offset]
        possible_matches = self.ids_lookup.get(token, None)

        if possible_matches is None:
            return None, None

        for ids, embedding in possible_matches:
            if tokens[offset:offset + len(ids)] == ids:
                return embedding, len(ids)

        return None, None


def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'):
    cond_model = shared.sd_model.cond_stage_model
    embedding_layer = cond_model.wrapped.transformer.text_model.embeddings

    with devices.autocast():
        cond_model([""])  # will send cond model to GPU if lowvram/medvram is active

    ids = cond_model.tokenizer(init_text, max_length=num_vectors_per_token, return_tensors="pt", add_special_tokens=False)["input_ids"]
    embedded = embedding_layer.token_embedding.wrapped(ids.to(devices.device)).squeeze(0)
    vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device)

    for i in range(num_vectors_per_token):
        vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token]

    # Remove illegal characters from name.
    name = "".join( x for x in name if (x.isalnum() or x in "._- "))
    fn = os.path.join(shared.cmd_opts.embeddings_dir, f"{name}.pt")
    if not overwrite_old:
        assert not os.path.exists(fn), f"file {fn} already exists"

    embedding = Embedding(vec, name)
    embedding.step = 0
    embedding.save(fn)

    return fn


def write_loss(log_directory, filename, step, epoch_len, values):
    if shared.opts.training_write_csv_every == 0:
        return

    if (step + 1) % shared.opts.training_write_csv_every != 0:
        return
    write_csv_header = False if os.path.exists(os.path.join(log_directory, filename)) else True

    with open(os.path.join(log_directory, filename), "a+", newline='') as fout:
        csv_writer = csv.DictWriter(fout, fieldnames=["step", "epoch", "epoch_step", *(values.keys())])

        if write_csv_header:
            csv_writer.writeheader()

        epoch = step // epoch_len
        epoch_step = step % epoch_len 

        csv_writer.writerow({
            "step": step + 1,
            "epoch": epoch,
            "epoch_step": epoch_step + 1,
            **values,
        })

def validate_train_inputs(model_name, learn_rate, batch_size, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
    assert model_name, f"{name} not selected"
    assert learn_rate, "Learning rate is empty or 0"
    assert isinstance(batch_size, int), "Batch size must be integer"
    assert batch_size > 0, "Batch size must be positive"
    assert data_root, "Dataset directory is empty"
    assert os.path.isdir(data_root), "Dataset directory doesn't exist"
    assert os.listdir(data_root), "Dataset directory is empty"
    assert template_file, "Prompt template file is empty"
    assert os.path.isfile(template_file), "Prompt template file doesn't exist"
    assert steps, "Max steps is empty or 0"
    assert isinstance(steps, int), "Max steps must be integer"
    assert steps > 0 , "Max steps must be positive"
    assert isinstance(save_model_every, int), "Save {name} must be integer"
    assert save_model_every >= 0 , "Save {name} must be positive or 0"
    assert isinstance(create_image_every, int), "Create image must be integer"
    assert create_image_every >= 0 , "Create image must be positive or 0"
    if save_model_every or create_image_every:
        assert log_directory, "Log directory is empty"

def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
    save_embedding_every = save_embedding_every or 0
    create_image_every = create_image_every or 0
    validate_train_inputs(embedding_name, learn_rate, batch_size, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")

    shared.state.textinfo = "Initializing textual inversion training..."
    shared.state.job_count = steps

    filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')

    log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), embedding_name)

    if save_embedding_every > 0:
        embedding_dir = os.path.join(log_directory, "embeddings")
        os.makedirs(embedding_dir, exist_ok=True)
    else:
        embedding_dir = None

    if create_image_every > 0:
        images_dir = os.path.join(log_directory, "images")
        os.makedirs(images_dir, exist_ok=True)
    else:
        images_dir = None

    if create_image_every > 0 and save_image_with_stored_embedding:
        images_embeds_dir = os.path.join(log_directory, "image_embeddings")
        os.makedirs(images_embeds_dir, exist_ok=True)
    else:
        images_embeds_dir = None

    cond_model = shared.sd_model.cond_stage_model

    hijack = sd_hijack.model_hijack

    embedding = hijack.embedding_db.word_embeddings[embedding_name]

    ititial_step = embedding.step or 0
    if ititial_step > steps:
        shared.state.textinfo = f"Model has already been trained beyond specified max steps"
        return embedding, filename

    scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)

    # dataset loading may take a while, so input validations and early returns should be done before this
    shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
    with torch.autocast("cuda"):
        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size)

    embedding.vec.requires_grad = True
    optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)

    losses = torch.zeros((32,))

    last_saved_file = "<none>"
    last_saved_image = "<none>"
    forced_filename = "<none>"
    embedding_yet_to_be_embedded = False

    pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
    for i, entries in pbar:
        embedding.step = i + ititial_step

        scheduler.apply(optimizer, embedding.step)
        if scheduler.finished:
            break

        if shared.state.interrupted:
            break

        with torch.autocast("cuda"):
            c = cond_model([entry.cond_text for entry in entries])
            x = torch.stack([entry.latent for entry in entries]).to(devices.device)
            loss = shared.sd_model(x, c)[0]
            del x

            losses[embedding.step % losses.shape[0]] = loss.item()

            optimizer.zero_grad()
            loss.backward()