2022-09-10 17:06:19 +08:00
|
|
|
import math
|
2022-09-03 17:08:45 +08:00
|
|
|
import os
|
|
|
|
import sys
|
|
|
|
import traceback
|
|
|
|
import torch
|
|
|
|
import numpy as np
|
2022-09-05 06:41:20 +08:00
|
|
|
from torch import einsum
|
2022-10-03 05:31:19 +08:00
|
|
|
from torch.nn.functional import silu
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
import modules.textual_inversion.textual_inversion
|
|
|
|
from modules import prompt_parser, devices, sd_hijack_optimizations, shared
|
2022-09-05 06:41:20 +08:00
|
|
|
from modules.shared import opts, device, cmd_opts
|
|
|
|
|
|
|
|
import ldm.modules.attention
|
2022-09-13 19:29:56 +08:00
|
|
|
import ldm.modules.diffusionmodules.model
|
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward
|
|
|
|
diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity
|
|
|
|
diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward
|
2022-09-13 19:29:56 +08:00
|
|
|
|
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
def apply_optimizations():
|
2022-10-07 10:31:53 +08:00
|
|
|
ldm.modules.diffusionmodules.model.nonlinearity = silu
|
2022-10-02 20:03:39 +08:00
|
|
|
if cmd_opts.opt_split_attention_v1:
|
|
|
|
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
|
2022-10-07 11:01:14 +08:00
|
|
|
elif cmd_opts.opt_split_attention:
|
2022-10-07 10:22:28 +08:00
|
|
|
ldm.modules.attention_CrossAttention_forward = sd_hijack_optimizations.split_cross_attention_forward
|
|
|
|
ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
|
2022-10-07 11:38:01 +08:00
|
|
|
elif not cmd_opts.disable_opt_xformers_attention and not (cmd_opts.opt_split_attention or torch.version.hip):
|
2022-10-07 10:22:28 +08:00
|
|
|
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
|
|
|
|
ldm.modules.attention.CrossAttention._maybe_init = sd_hijack_optimizations._maybe_init
|
|
|
|
ldm.modules.attention.CrossAttention.attention_op = None
|
2022-10-02 20:03:39 +08:00
|
|
|
ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
|
2022-09-13 19:29:56 +08:00
|
|
|
|
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
def undo_optimizations():
|
|
|
|
ldm.modules.attention.CrossAttention.forward = attention_CrossAttention_forward
|
|
|
|
ldm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
|
|
|
|
ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
|
2022-09-13 19:29:56 +08:00
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
class StableDiffusionModelHijack:
|
|
|
|
fixes = None
|
|
|
|
comments = []
|
2022-09-05 08:25:37 +08:00
|
|
|
layers = None
|
|
|
|
circular_enabled = False
|
2022-09-28 03:56:18 +08:00
|
|
|
clip = None
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
embedding_db = modules.textual_inversion.textual_inversion.EmbeddingDatabase(cmd_opts.embeddings_dir)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
def hijack(self, m):
|
|
|
|
model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
|
|
|
|
|
|
|
|
model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
|
|
|
|
m.cond_stage_model = FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
|
2022-09-29 20:40:28 +08:00
|
|
|
|
2022-09-28 03:56:18 +08:00
|
|
|
self.clip = m.cond_stage_model
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
apply_optimizations()
|
2022-09-05 06:41:20 +08:00
|
|
|
|
2022-09-05 08:25:37 +08:00
|
|
|
def flatten(el):
|
|
|
|
flattened = [flatten(children) for children in el.children()]
|
|
|
|
res = [el]
|
|
|
|
for c in flattened:
|
|
|
|
res += c
|
|
|
|
return res
|
|
|
|
|
|
|
|
self.layers = flatten(m)
|
|
|
|
|
2022-09-29 20:40:28 +08:00
|
|
|
def undo_hijack(self, m):
|
|
|
|
if type(m.cond_stage_model) == FrozenCLIPEmbedderWithCustomWords:
|
|
|
|
m.cond_stage_model = m.cond_stage_model.wrapped
|
|
|
|
|
|
|
|
model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
|
|
|
|
if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
|
|
|
|
model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
|
|
|
|
|
2022-09-05 08:25:37 +08:00
|
|
|
def apply_circular(self, enable):
|
|
|
|
if self.circular_enabled == enable:
|
|
|
|
return
|
|
|
|
|
|
|
|
self.circular_enabled = enable
|
|
|
|
|
|
|
|
for layer in [layer for layer in self.layers if type(layer) == torch.nn.Conv2d]:
|
|
|
|
layer.padding_mode = 'circular' if enable else 'zeros'
|
|
|
|
|
2022-09-28 03:56:18 +08:00
|
|
|
def tokenize(self, text):
|
|
|
|
max_length = self.clip.max_length - 2
|
|
|
|
_, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
|
2022-09-28 07:29:53 +08:00
|
|
|
return remade_batch_tokens[0], token_count, max_length
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-09-29 16:31:48 +08:00
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
|
|
|
|
def __init__(self, wrapped, hijack):
|
|
|
|
super().__init__()
|
|
|
|
self.wrapped = wrapped
|
2022-10-02 20:03:39 +08:00
|
|
|
self.hijack: StableDiffusionModelHijack = hijack
|
2022-09-03 17:08:45 +08:00
|
|
|
self.tokenizer = wrapped.tokenizer
|
|
|
|
self.max_length = wrapped.max_length
|
|
|
|
self.token_mults = {}
|
|
|
|
|
|
|
|
tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_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
|
|
|
|
|
2022-09-29 16:31:48 +08:00
|
|
|
def tokenize_line(self, line, used_custom_terms, hijack_comments):
|
|
|
|
id_start = self.wrapped.tokenizer.bos_token_id
|
|
|
|
id_end = self.wrapped.tokenizer.eos_token_id
|
|
|
|
maxlen = self.wrapped.max_length
|
|
|
|
|
|
|
|
if opts.enable_emphasis:
|
|
|
|
parsed = prompt_parser.parse_prompt_attention(line)
|
|
|
|
else:
|
|
|
|
parsed = [[line, 1.0]]
|
|
|
|
|
|
|
|
tokenized = self.wrapped.tokenizer([text for text, _ in parsed], truncation=False, add_special_tokens=False)["input_ids"]
|
|
|
|
|
|
|
|
fixes = []
|
|
|
|
remade_tokens = []
|
|
|
|
multipliers = []
|
|
|
|
|
|
|
|
for tokens, (text, weight) in zip(tokenized, parsed):
|
|
|
|
i = 0
|
|
|
|
while i < len(tokens):
|
|
|
|
token = tokens[i]
|
|
|
|
|
2022-10-03 00:40:51 +08:00
|
|
|
embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
|
2022-09-29 16:31:48 +08:00
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
if embedding is None:
|
2022-09-29 16:31:48 +08:00
|
|
|
remade_tokens.append(token)
|
|
|
|
multipliers.append(weight)
|
2022-10-02 20:03:39 +08:00
|
|
|
i += 1
|
2022-09-29 16:31:48 +08:00
|
|
|
else:
|
2022-10-02 20:03:39 +08:00
|
|
|
emb_len = int(embedding.vec.shape[0])
|
|
|
|
fixes.append((len(remade_tokens), embedding))
|
|
|
|
remade_tokens += [0] * emb_len
|
|
|
|
multipliers += [weight] * emb_len
|
|
|
|
used_custom_terms.append((embedding.name, embedding.checksum()))
|
2022-10-03 00:40:51 +08:00
|
|
|
i += embedding_length_in_tokens
|
2022-09-29 16:31:48 +08:00
|
|
|
|
|
|
|
if len(remade_tokens) > maxlen - 2:
|
|
|
|
vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
|
|
|
|
ovf = remade_tokens[maxlen - 2:]
|
|
|
|
overflowing_words = [vocab.get(int(x), "") for x in ovf]
|
|
|
|
overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
|
|
|
|
hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
|
|
|
|
|
|
|
|
token_count = len(remade_tokens)
|
|
|
|
remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
|
|
|
|
remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end]
|
|
|
|
|
|
|
|
multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
|
|
|
|
multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
|
|
|
|
|
|
|
|
return remade_tokens, fixes, multipliers, token_count
|
|
|
|
|
|
|
|
def process_text(self, texts):
|
|
|
|
used_custom_terms = []
|
|
|
|
remade_batch_tokens = []
|
|
|
|
hijack_comments = []
|
|
|
|
hijack_fixes = []
|
|
|
|
token_count = 0
|
|
|
|
|
|
|
|
cache = {}
|
|
|
|
batch_multipliers = []
|
|
|
|
for line in texts:
|
|
|
|
if line in cache:
|
|
|
|
remade_tokens, fixes, multipliers = cache[line]
|
|
|
|
else:
|
|
|
|
remade_tokens, fixes, multipliers, token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
|
|
|
|
|
|
|
|
cache[line] = (remade_tokens, fixes, multipliers)
|
|
|
|
|
|
|
|
remade_batch_tokens.append(remade_tokens)
|
|
|
|
hijack_fixes.append(fixes)
|
|
|
|
batch_multipliers.append(multipliers)
|
|
|
|
|
|
|
|
return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
|
|
|
|
|
|
|
|
|
|
|
|
def process_text_old(self, text):
|
2022-09-03 17:08:45 +08:00
|
|
|
id_start = self.wrapped.tokenizer.bos_token_id
|
|
|
|
id_end = self.wrapped.tokenizer.eos_token_id
|
2022-09-25 14:13:03 +08:00
|
|
|
maxlen = self.wrapped.max_length
|
2022-09-03 17:08:45 +08:00
|
|
|
used_custom_terms = []
|
2022-09-28 03:56:18 +08:00
|
|
|
remade_batch_tokens = []
|
|
|
|
overflowing_words = []
|
|
|
|
hijack_comments = []
|
|
|
|
hijack_fixes = []
|
|
|
|
token_count = 0
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
cache = {}
|
|
|
|
batch_tokens = self.wrapped.tokenizer(text, truncation=False, add_special_tokens=False)["input_ids"]
|
|
|
|
batch_multipliers = []
|
|
|
|
for tokens in batch_tokens:
|
|
|
|
tuple_tokens = tuple(tokens)
|
|
|
|
|
|
|
|
if tuple_tokens in cache:
|
|
|
|
remade_tokens, fixes, multipliers = cache[tuple_tokens]
|
|
|
|
else:
|
|
|
|
fixes = []
|
|
|
|
remade_tokens = []
|
|
|
|
multipliers = []
|
|
|
|
mult = 1.0
|
|
|
|
|
|
|
|
i = 0
|
|
|
|
while i < len(tokens):
|
|
|
|
token = tokens[i]
|
|
|
|
|
2022-10-03 00:40:51 +08:00
|
|
|
embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
mult_change = self.token_mults.get(token) if opts.enable_emphasis else None
|
|
|
|
if mult_change is not None:
|
|
|
|
mult *= mult_change
|
2022-10-02 20:03:39 +08:00
|
|
|
i += 1
|
|
|
|
elif embedding is None:
|
2022-09-03 17:08:45 +08:00
|
|
|
remade_tokens.append(token)
|
|
|
|
multipliers.append(mult)
|
2022-10-02 20:03:39 +08:00
|
|
|
i += 1
|
2022-09-03 17:08:45 +08:00
|
|
|
else:
|
2022-10-02 20:03:39 +08:00
|
|
|
emb_len = int(embedding.vec.shape[0])
|
|
|
|
fixes.append((len(remade_tokens), embedding))
|
|
|
|
remade_tokens += [0] * emb_len
|
|
|
|
multipliers += [mult] * emb_len
|
|
|
|
used_custom_terms.append((embedding.name, embedding.checksum()))
|
2022-10-03 00:40:51 +08:00
|
|
|
i += embedding_length_in_tokens
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
if len(remade_tokens) > maxlen - 2:
|
|
|
|
vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
|
|
|
|
ovf = remade_tokens[maxlen - 2:]
|
|
|
|
overflowing_words = [vocab.get(int(x), "") for x in ovf]
|
|
|
|
overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
|
2022-09-28 03:56:18 +08:00
|
|
|
hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
|
2022-10-02 20:03:39 +08:00
|
|
|
|
2022-09-28 03:56:18 +08:00
|
|
|
token_count = len(remade_tokens)
|
2022-09-03 17:08:45 +08:00
|
|
|
remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
|
|
|
|
remade_tokens = [id_start] + remade_tokens[0:maxlen-2] + [id_end]
|
|
|
|
cache[tuple_tokens] = (remade_tokens, fixes, multipliers)
|
|
|
|
|
|
|
|
multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
|
|
|
|
multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
|
|
|
|
|
|
|
|
remade_batch_tokens.append(remade_tokens)
|
2022-09-28 03:56:18 +08:00
|
|
|
hijack_fixes.append(fixes)
|
2022-09-03 17:08:45 +08:00
|
|
|
batch_multipliers.append(multipliers)
|
2022-09-28 03:56:18 +08:00
|
|
|
return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
|
|
|
|
|
|
|
|
def forward(self, text):
|
2022-09-29 16:31:48 +08:00
|
|
|
|
|
|
|
if opts.use_old_emphasis_implementation:
|
|
|
|
batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text)
|
|
|
|
else:
|
|
|
|
batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
|
|
|
|
|
2022-09-28 03:56:18 +08:00
|
|
|
self.hijack.fixes = hijack_fixes
|
|
|
|
self.hijack.comments = hijack_comments
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
if len(used_custom_terms) > 0:
|
2022-09-29 16:31:48 +08:00
|
|
|
self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
tokens = torch.asarray(remade_batch_tokens).to(device)
|
|
|
|
outputs = self.wrapped.transformer(input_ids=tokens)
|
|
|
|
z = outputs.last_hidden_state
|
|
|
|
|
|
|
|
# restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
|
2022-09-08 02:40:32 +08:00
|
|
|
batch_multipliers = torch.asarray(batch_multipliers).to(device)
|
2022-09-03 17:08:45 +08:00
|
|
|
original_mean = z.mean()
|
|
|
|
z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
|
|
|
|
new_mean = z.mean()
|
|
|
|
z *= original_mean / new_mean
|
|
|
|
|
|
|
|
return z
|
|
|
|
|
|
|
|
|
|
|
|
class EmbeddingsWithFixes(torch.nn.Module):
|
|
|
|
def __init__(self, wrapped, embeddings):
|
|
|
|
super().__init__()
|
|
|
|
self.wrapped = wrapped
|
|
|
|
self.embeddings = embeddings
|
|
|
|
|
|
|
|
def forward(self, input_ids):
|
|
|
|
batch_fixes = self.embeddings.fixes
|
|
|
|
self.embeddings.fixes = None
|
|
|
|
|
|
|
|
inputs_embeds = self.wrapped(input_ids)
|
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
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
|
|
|
|
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:]])
|
|
|
|
|
|
|
|
vecs.append(tensor)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
2022-10-02 20:03:39 +08:00
|
|
|
return torch.stack(vecs)
|
2022-09-03 17:08:45 +08:00
|
|
|
|
|
|
|
|
2022-09-05 07:16:36 +08:00
|
|
|
def add_circular_option_to_conv_2d():
|
|
|
|
conv2d_constructor = torch.nn.Conv2d.__init__
|
2022-09-05 06:41:20 +08:00
|
|
|
|
2022-09-05 07:16:36 +08:00
|
|
|
def conv2d_constructor_circular(self, *args, **kwargs):
|
|
|
|
return conv2d_constructor(self, *args, padding_mode='circular', **kwargs)
|
2022-09-05 06:41:20 +08:00
|
|
|
|
2022-09-05 07:16:36 +08:00
|
|
|
torch.nn.Conv2d.__init__ = conv2d_constructor_circular
|
2022-09-05 06:41:20 +08:00
|
|
|
|
|
|
|
|
2022-09-03 17:08:45 +08:00
|
|
|
model_hijack = StableDiffusionModelHijack()
|