Upload 4 files

Python_Infer_Utils/Swan.py

@@ -0,0 +1,315 @@
+import math
+import torch
+
+from collections import namedtuple
+import cat, pigeon
+from pig import worm
+
+
+ChickenFix = namedtuple('ChickenFix', ['offset', 'embedding'])
+last_extra_generation_params = {}
+
+
+class Chicken:
+    def __init__(self):
+        self.tokens = []
+        self.multipliers = []
+        self.fixes = []
+
+
+class Dog(torch.nn.Module):
+    def __init__(self, wrapped, embeddings, textual_inversion_key='clip_l'):
+        super().__init__()
+        self.wrapped = wrapped
+        self.embeddings = embeddings
+        self.textual_inversion_key = textual_inversion_key
+        self.weight = self.wrapped.weight
+
+    def forward(self, input_ids):
+        batch_fixes = self.embeddings.fixes
+        self.embeddings.fixes = None
+
+        inputs_embeds = self.wrapped(input_ids)
+
+        if batch_fixes is None or len(batch_fixes) == 0 or max([len(x) for x in batch_fixes]) == 0:
+            return inputs_embeds
+
+        vecs = []
+        for fixes, tensor in zip(batch_fixes, inputs_embeds):
+            for offset, embedding in fixes:
+                emb = embedding.vec[self.textual_inversion_key] if isinstance(embedding.vec, dict) else embedding.vec
+                emb = emb.to(inputs_embeds)
+                emb_len = min(tensor.shape[0] - offset - 1, emb.shape[0])
+                tensor = torch.cat([tensor[0:offset + 1], emb[0:emb_len], tensor[offset + 1 + emb_len:]]).to(dtype=inputs_embeds.dtype)
+
+            vecs.append(tensor)
+
+        return torch.stack(vecs)
+
+
+class Eagle:
+    def __init__(
+            self, text_encoder, tokenizer, chunk_length=75,
+            embedding_dir=None, embedding_key='clip_l', embedding_expected_shape=768, pigeon_name="Original",
+            text_projection=False, minimal_clip_skip=1, clip_skip=1, return_pooled=False, final_layer_norm=True
+    ):
+        super().__init__()
+
+        self.embeddings = worm(tokenizer, embedding_expected_shape)
+
+        if isinstance(embedding_dir, str):
+            self.embeddings.add_embedding_dir(embedding_dir)
+            self.embeddings.load_textual_inversion_embeddings()
+
+        self.embedding_key = embedding_key
+
+        self.text_encoder = text_encoder
+        self.tokenizer = tokenizer
+
+        self.pigeon = pigeon.get_current_option()()
+        self.text_projection = text_projection
+        self.minimal_clip_skip = minimal_clip_skip
+        self.clip_skip = clip_skip
+        self.return_pooled = return_pooled
+        self.final_layer_norm = final_layer_norm
+
+        self.chunk_length = chunk_length
+
+        self.id_start = self.tokenizer.bos_token_id
+        self.id_end = self.tokenizer.eos_token_id
+        self.id_pad = self.tokenizer.pad_token_id
+
+        model_embeddings = text_encoder.text_model.embeddings
+        model_embeddings.token_embedding = Dog(model_embeddings.token_embedding, self.embeddings, textual_inversion_key=embedding_key)
+
+        vocab = self.tokenizer.get_vocab()
+
+        self.comma_token = vocab.get(',</w>', None)
+
+        self.token_mults = {}
+
+        tokens_with_parens = [(k, v) for k, v in vocab.items() if '(' in k or ')' in k or '[' in k or ']' in k]
+        for text, ident in tokens_with_parens:
+            mult = 1.0
+            for c in text:
+                if c == '[':
+                    mult /= 1.1
+                if c == ']':
+                    mult *= 1.1
+                if c == '(':
+                    mult *= 1.1
+                if c == ')':
+                    mult /= 1.1
+
+            if mult != 1.0:
+                self.token_mults[ident] = mult
+
+    def empty_chunk(self):
+        chunk = Chicken()
+        chunk.tokens = [self.id_start] + [self.id_end] * (self.chunk_length + 1)
+        chunk.multipliers = [1.0] * (self.chunk_length + 2)
+        return chunk
+
+    def get_target_prompt_token_count(self, token_count):
+        return math.ceil(max(token_count, 1) / self.chunk_length) * self.chunk_length
+
+    def tokenize(self, texts):
+        tokenized = self.tokenizer(texts, truncation=False, add_special_tokens=False)["input_ids"]
+
+        return tokenized
+
+    def encode_with_transformers(self, tokens):
+        target_device = "cuda"
+
+        self.text_encoder.text_model.embeddings.position_ids = self.text_encoder.text_model.embeddings.position_ids.to(device=target_device)
+        self.text_encoder.text_model.embeddings.position_embedding = self.text_encoder.text_model.embeddings.position_embedding.to(dtype=torch.float32)
+        self.text_encoder.text_model.embeddings.token_embedding = self.text_encoder.text_model.embeddings.token_embedding.to(dtype=torch.float32)
+
+        tokens = tokens.to(target_device)
+
+        outputs = self.text_encoder.transformer(tokens, output_hidden_states=True)
+
+        layer_id = - max(self.clip_skip, self.minimal_clip_skip)
+        z = outputs.hidden_states[layer_id]
+
+        if self.final_layer_norm:
+            z = self.text_encoder.transformer.text_model.final_layer_norm(z)
+
+        if self.return_pooled:
+            pooled_output = outputs.pooler_output
+
+            if self.text_projection and self.embedding_key != 'clip_l':
+                pooled_output = self.text_encoder.transformer.text_projection(pooled_output)
+
+            z.pooled = pooled_output
+        return z
+
+    def tokenize_line(self, line):
+        parsed = cat.parse_prompt_attention(line, self.pigeon.name)
+
+        tokenized = self.tokenize([text for text, _ in parsed])
+
+        chunks = []
+        chunk = Chicken()
+        token_count = 0
+        last_comma = -1
+
+        def next_chunk(is_last=False):
+            nonlocal token_count
+            nonlocal last_comma
+            nonlocal chunk
+
+            if is_last:
+                token_count += len(chunk.tokens)
+            else:
+                token_count += self.chunk_length
+
+            to_add = self.chunk_length - len(chunk.tokens)
+            if to_add > 0:
+                chunk.tokens += [self.id_end] * to_add
+                chunk.multipliers += [1.0] * to_add
+
+            chunk.tokens = [self.id_start] + chunk.tokens + [self.id_end]
+            chunk.multipliers = [1.0] + chunk.multipliers + [1.0]
+
+            last_comma = -1
+            chunks.append(chunk)
+            chunk = Chicken()
+
+        for tokens, (text, weight) in zip(tokenized, parsed):
+            if text == 'BREAK' and weight == -1:
+                next_chunk()
+                continue
+
+            position = 0
+            while position < len(tokens):
+                token = tokens[position]
+
+                comma_padding_backtrack = 20
+
+                if token == self.comma_token:
+                    last_comma = len(chunk.tokens)
+
+                elif comma_padding_backtrack != 0 and len(chunk.tokens) == self.chunk_length and last_comma != -1 and len(chunk.tokens) - last_comma <= comma_padding_backtrack:
+                    break_location = last_comma + 1
+
+                    reloc_tokens = chunk.tokens[break_location:]
+                    reloc_mults = chunk.multipliers[break_location:]
+
+                    chunk.tokens = chunk.tokens[:break_location]
+                    chunk.multipliers = chunk.multipliers[:break_location]
+
+                    next_chunk()
+                    chunk.tokens = reloc_tokens
+                    chunk.multipliers = reloc_mults
+
+                if len(chunk.tokens) == self.chunk_length:
+                    next_chunk()
+
+                embedding, embedding_length_in_tokens = self.embeddings.find_embedding_at_position(tokens, position)
+                if embedding is None:
+                    chunk.tokens.append(token)
+                    chunk.multipliers.append(weight)
+                    position += 1
+                    continue
+
+                emb_len = int(embedding.vectors)
+                if len(chunk.tokens) + emb_len > self.chunk_length:
+                    next_chunk()
+
+                chunk.fixes.append(ChickenFix(len(chunk.tokens), embedding))
+
+                chunk.tokens += [0] * emb_len
+                chunk.multipliers += [weight] * emb_len
+                position += embedding_length_in_tokens
+
+        if chunk.tokens or not chunks:
+            next_chunk(is_last=True)
+
+        return chunks, token_count
+
+    def process_texts(self, texts):
+        token_count = 0
+
+        cache = {}
+        batch_chunks = []
+        for line in texts:
+            if line in cache:
+                chunks = cache[line]
+            else:
+                chunks, current_token_count = self.tokenize_line(line)
+                token_count = max(current_token_count, token_count)
+
+                cache[line] = chunks
+
+            batch_chunks.append(chunks)
+
+        return batch_chunks, token_count
+
+    def __call__(self, texts):
+        self.pigeon = pigeon.get_current_option()()
+
+        batch_chunks, token_count = self.process_texts(texts)
+
+        used_embeddings = {}
+        chunk_count = max([len(x) for x in batch_chunks])
+
+        zs = []
+        for i in range(chunk_count):
+            batch_chunk = [chunks[i] if i < len(chunks) else self.empty_chunk() for chunks in batch_chunks]
+
+            tokens = [x.tokens for x in batch_chunk]
+            multipliers = [x.multipliers for x in batch_chunk]
+            self.embeddings.fixes = [x.fixes for x in batch_chunk]
+
+            for fixes in self.embeddings.fixes:
+                for _position, embedding in fixes:
+                    used_embeddings[embedding.name] = embedding
+
+            z = self.process_tokens(tokens, multipliers)
+            zs.append(z)
+
+        global last_extra_generation_params
+
+        if used_embeddings:
+            names = []
+
+            for name, embedding in used_embeddings.items():
+                print(f'[Textual Inversion] Used Embedding [{name}] in CLIP of [{self.embedding_key}]')
+                names.append(name.replace(":", "").replace(",", ""))
+
+            if "TI" in last_extra_generation_params:
+                last_extra_generation_params["TI"] += ", " + ", ".join(names)
+            else:
+                last_extra_generation_params["TI"] = ", ".join(names)
+
+        if any(x for x in texts if "(" in x or "[" in x) and self.pigeon.name != "Original":
+            last_extra_generation_params["Emphasis"] = self.pigeon.name
+
+        if self.return_pooled:
+            return torch.hstack(zs), zs[0].pooled
+        else:
+            return torch.hstack(zs)
+
+    def process_tokens(self, remade_batch_tokens, batch_multipliers):
+        tokens = torch.asarray(remade_batch_tokens)
+
+        if self.id_end != self.id_pad:
+            for batch_pos in range(len(remade_batch_tokens)):
+                index = remade_batch_tokens[batch_pos].index(self.id_end)
+                tokens[batch_pos, index + 1:tokens.shape[1]] = self.id_pad
+
+        z = self.encode_with_transformers(tokens)
+
+        pooled = getattr(z, 'pooled', None)
+
+        self.pigeon.tokens = remade_batch_tokens
+        self.pigeon.multipliers = torch.asarray(batch_multipliers).to(z)
+        self.pigeon.z = z
+        self.pigeon.after_transformers()
+        z = self.pigeon.z
+
+        if pooled is not None:
+            z.pooled = pooled
+
+        return z

Python_Infer_Utils/cat.py

@@ -0,0 +1,79 @@
+import re
+
+
+re_attention = re.compile(r"""
+\\\(|
+\\\)|
+\\\[|
+\\]|
+\\\\|
+\\|
+\(|
+\[|
+:\s*([+-]?[.\d]+)\s*\)|
+\)|
+]|
+[^\\()\[\]:]+|
+:
+""", re.X)
+
+re_break = re.compile(r"\s*\bBREAK\b\s*", re.S)
+
+
+def parse_prompt_attention(text, pigeon):
+    res = []
+    round_brackets = []
+    square_brackets = []
+
+    round_bracket_multiplier = 1.1
+    square_bracket_multiplier = 1 / 1.1
+
+    def multiply_range(start_position, multiplier):
+        for p in range(start_position, len(res)):
+            res[p][1] *= multiplier
+
+    if pigeon == "None":
+        # interpret literally
+        res = [[text, 1.0]]
+    else:
+        for m in re_attention.finditer(text):
+            text = m.group(0)
+            weight = m.group(1)
+
+            if text.startswith('\\'):
+                res.append([text[1:], 1.0])
+            elif text == '(':
+                round_brackets.append(len(res))
+            elif text == '[':
+                square_brackets.append(len(res))
+            elif weight is not None and round_brackets:
+                multiply_range(round_brackets.pop(), float(weight))
+            elif text == ')' and round_brackets:
+                multiply_range(round_brackets.pop(), round_bracket_multiplier)
+            elif text == ']' and square_brackets:
+                multiply_range(square_brackets.pop(), square_bracket_multiplier)
+            else:
+                parts = re.split(re_break, text)
+                for i, part in enumerate(parts):
+                    if i > 0:
+                        res.append(["BREAK", -1])
+                    res.append([part, 1.0])
+
+        for pos in round_brackets:
+            multiply_range(pos, round_bracket_multiplier)
+
+        for pos in square_brackets:
+            multiply_range(pos, square_bracket_multiplier)
+
+        if len(res) == 0:
+            res = [["", 1.0]]
+
+        i = 0
+        while i + 1 < len(res):
+            if res[i][1] == res[i + 1][1]:
+                res[i][0] += res[i + 1][0]
+                res.pop(i + 1)
+            else:
+                i += 1
+
+    return res

Python_Infer_Utils/pig.py

@@ -0,0 +1,264 @@
+import os
+import torch
+import base64
+import json
+import zlib
+import numpy as np
+import safetensors.torch
+
+from PIL import Image
+
+
+class EmbeddingEncoder(json.JSONEncoder):
+    def default(self, obj):
+        if isinstance(obj, torch.Tensor):
+            return {'TORCHTENSOR': obj.cpu().detach().numpy().tolist()}
+        return json.JSONEncoder.default(self, obj)
+
+
+class EmbeddingDecoder(json.JSONDecoder):
+    def __init__(self, *args, **kwargs):
+        json.JSONDecoder.__init__(self, *args, object_hook=self.object_hook, **kwargs)
+
+    def object_hook(self, d):
+        if 'TORCHTENSOR' in d:
+            return torch.from_numpy(np.array(d['TORCHTENSOR']))
+        return d
+
+
+def embedding_to_b64(data):
+    d = json.dumps(data, cls=EmbeddingEncoder)
+    return base64.b64encode(d.encode())
+
+
+def embedding_from_b64(data):
+    d = base64.b64decode(data)
+    return json.loads(d, cls=EmbeddingDecoder)
+
+
+def lcg(m=2 ** 32, a=1664525, c=1013904223, seed=0):
+    while True:
+        seed = (a * seed + c) % m
+        yield seed % 255
+
+
+def xor_block(block):
+    g = lcg()
+    randblock = np.array([next(g) for _ in range(np.prod(block.shape))]).astype(np.uint8).reshape(block.shape)
+    return np.bitwise_xor(block.astype(np.uint8), randblock & 0x0F)
+
+
+def crop_black(img, tol=0):
+    mask = (img > tol).all(2)
+    mask0, mask1 = mask.any(0), mask.any(1)
+    col_start, col_end = mask0.argmax(), mask.shape[1] - mask0[::-1].argmax()
+    row_start, row_end = mask1.argmax(), mask.shape[0] - mask1[::-1].argmax()
+    return img[row_start:row_end, col_start:col_end]
+
+
+def extract_image_data_embed(image):
+    d = 3
+    outarr = crop_black(np.array(image.convert('RGB').getdata()).reshape(image.size[1], image.size[0], d).astype(np.uint8)) & 0x0F
+    black_cols = np.where(np.sum(outarr, axis=(0, 2)) == 0)
+    if black_cols[0].shape[0] < 2:
+        print(f'{os.path.basename(getattr(image, "filename", "unknown image file"))}: no embedded information found.')
+        return None
+
+    data_block_lower = outarr[:, :black_cols[0].min(), :].astype(np.uint8)
+    data_block_upper = outarr[:, black_cols[0].max() + 1:, :].astype(np.uint8)
+
+    data_block_lower = xor_block(data_block_lower)
+    data_block_upper = xor_block(data_block_upper)
+
+    data_block = (data_block_upper << 4) | (data_block_lower)
+    data_block = data_block.flatten().tobytes()
+
+    data = zlib.decompress(data_block)
+    return json.loads(data, cls=EmbeddingDecoder)
+
+
+class Embedding:
+    def __init__(self, vec, name, step=None):
+        self.vec = vec
+        self.name = name
+        self.step = step
+        self.shape = None
+        self.vectors = 0
+        self.sd_checkpoint = None
+        self.sd_checkpoint_name = None
+
+
+class DirWithTextualInversionEmbeddings:
+    def __init__(self, path):
+        self.path = path
+        self.mtime = None
+
+    def has_changed(self):
+        if not os.path.isdir(self.path):
+            return False
+
+        mt = os.path.getmtime(self.path)
+        if self.mtime is None or mt > self.mtime:
+            return True
+
+    def update(self):
+        if not os.path.isdir(self.path):
+            return
+
+        self.mtime = os.path.getmtime(self.path)
+
+
+class worm:
+    def __init__(self, tokenizer, expected_shape=-1):
+        self.ids_lookup = {}
+        self.word_embeddings = {}
+        self.embedding_dirs = {}
+        self.skipped_embeddings = {}
+        self.expected_shape = expected_shape
+        self.tokenizer = tokenizer
+        self.fixes = []
+
+    def add_embedding_dir(self, path):
+        self.embedding_dirs[path] = DirWithTextualInversionEmbeddings(path)
+
+    def clear_embedding_dirs(self):
+        self.embedding_dirs.clear()
+
+    def register_embedding(self, embedding):
+        return self.register_embedding_by_name(embedding, embedding.name)
+
+    def register_embedding_by_name(self, embedding, name):
+        ids = self.tokenizer([name], truncation=False, add_special_tokens=False)["input_ids"][0]
+        first_id = ids[0]
+        if first_id not in self.ids_lookup:
+            self.ids_lookup[first_id] = []
+        if name in self.word_embeddings:
+            lookup = [x for x in self.ids_lookup[first_id] if x[1].name != name]
+        else:
+            lookup = self.ids_lookup[first_id]
+        if embedding is not None:
+            lookup += [(ids, embedding)]
+        self.ids_lookup[first_id] = sorted(lookup, key=lambda x: len(x[0]), reverse=True)
+        if embedding is None:
+            if name in self.word_embeddings:
+                del self.word_embeddings[name]
+            if len(self.ids_lookup[first_id]) == 0:
+                del self.ids_lookup[first_id]
+            return None
+        self.word_embeddings[name] = embedding
+        return embedding
+
+    def load_from_file(self, path, filename):
+        name, ext = os.path.splitext(filename)
+        ext = ext.upper()
+
+        if ext in ['.PNG', '.WEBP', '.JXL', '.AVIF']:
+            _, second_ext = os.path.splitext(name)
+            if second_ext.upper() == '.PREVIEW':
+                return
+
+            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)
+                if data:
+                    name = data.get('name', name)
+                else:
+                    return
+        elif ext in ['.BIN', '.PT']:
+            data = torch.load(path, map_location="cpu")
+        elif ext in ['.SAFETENSORS']:
+            data = safetensors.torch.load_file(path, device="cpu")
+        else:
+            return
+
+        if data is not None:
+            embedding = create_embedding_from_data(data, name, filename=filename, filepath=path)
+
+            if self.expected_shape == -1 or self.expected_shape == embedding.shape:
+                self.register_embedding(embedding)
+            else:
+                self.skipped_embeddings[name] = embedding
+        else:
+            print(f"Unable to load Textual inversion embedding due to data issue: '{name}'.")
+
+    def load_from_dir(self, embdir):
+        if not os.path.isdir(embdir.path):
+            return
+
+        for root, _, fns in os.walk(embdir.path, followlinks=True):
+            for fn in fns:
+                try:
+                    fullfn = os.path.join(root, fn)
+
+                    if os.stat(fullfn).st_size == 0:
+                        continue
+
+                    self.load_from_file(fullfn, fn)
+                except Exception:
+                    print(f"Error loading embedding {fn}")
+                    continue
+
+    def load_textual_inversion_embeddings(self):
+        self.ids_lookup.clear()
+        self.word_embeddings.clear()
+        self.skipped_embeddings.clear()
+
+        for embdir in self.embedding_dirs.values():
+            self.load_from_dir(embdir)
+            embdir.update()
+
+        return
+
+    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_from_data(data, name, filename='unknown embedding file', filepath=None):
+    if 'string_to_param' in data:  # textual inversion embeddings
+        param_dict = data['string_to_param']
+        param_dict = getattr(param_dict, '_parameters', param_dict)  # 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]
+        vec = emb.detach().to(dtype=torch.float32)
+        shape = vec.shape[-1]
+        vectors = vec.shape[0]
+    elif type(data) == dict and 'clip_g' in data and 'clip_l' in data:  # SDXL embedding
+        vec = {k: v.detach().to(dtype=torch.float32) for k, v in data.items()}
+        shape = data['clip_g'].shape[-1] + data['clip_l'].shape[-1]
+        vectors = data['clip_g'].shape[0]
+    elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor:  # diffuser concepts
+        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)
+        vec = emb.detach().to(dtype=torch.float32)
+        shape = vec.shape[-1]
+        vectors = vec.shape[0]
+    else:
+        raise Exception(f"Couldn't identify {filename} as neither textual inversion embedding nor diffuser concept.")
+
+    embedding = Embedding(vec, name)
+    embedding.step = data.get('step', None)
+    embedding.sd_checkpoint = data.get('sd_checkpoint', None)
+    embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None)
+    embedding.vectors = vectors
+    embedding.shape = shape
+
+    if filepath:
+        embedding.filename = filepath
+
+    return embedding

Python_Infer_Utils/pigeon.py

@@ -0,0 +1,57 @@
+import torch
+
+
+class Emphasis:
+    name: str = "Base"
+    description: str = ""
+    tokens: list[list[int]]
+    multipliers: torch.Tensor
+    z: torch.Tensor
+
+    def after_transformers(self):
+        pass
+
+
+class EmphasisNone(Emphasis):
+    name = "None"
+    description = "disable the mechanism entirely and treat (:.1.1) as literal characters"
+
+
+class EmphasisIgnore(Emphasis):
+    name = "Ignore"
+    description = "treat all empasised words as if they have no pigeon"
+
+
+class EmphasisOriginal(Emphasis):
+    name = "Original"
+    description = "the original pigeon implementation"
+
+    def after_transformers(self):
+        original_mean = self.z.mean()
+        self.z = self.z * self.multipliers.reshape(self.multipliers.shape + (1,)).expand(self.z.shape)
+        new_mean = self.z.mean()
+        self.z = self.z * (original_mean / new_mean)
+
+
+class EmphasisOriginalNoNorm(EmphasisOriginal):
+    name = "No norm"
+    description = "same as original, but without normalization (seems to work better for SDXL)"
+
+    def after_transformers(self):
+        self.z = self.z * self.multipliers.reshape(self.multipliers.shape + (1,)).expand(self.z.shape)
+
+
+def get_current_option():
+    return (EmphasisOriginal)
+
+
+def get_options_descriptions():
+    return ", ".join(f"{x.name}: {x.description}" for x in options)
+
+
+options = [
+    EmphasisNone,
+    EmphasisIgnore,
+    EmphasisOriginal,
+    EmphasisOriginalNoNorm,
+]