stable-diffusion-webui-toke.../scripts/tokenizer.py

184 lines
7.0 KiB
Python

import html
from ldm.modules.encoders.modules import FrozenCLIPEmbedder
from modules import script_callbacks, shared
import gradio as gr
css = """
@media (prefers-color-scheme: dark) {
.tokenizer-token{
cursor: pointer;
}
.tokenizer-token-0 {background: rgba(255, 0, 0, 0.2);}
.tokenizer-token-0:hover {background: rgba(255, 0, 0, 0.4);}
.tokenizer-token-1 {background: rgba(0, 255, 0, 0.2);}
.tokenizer-token-1:hover {background: rgba(0, 255, 0, 0.4);}
.tokenizer-token-2 {background: rgba(0, 0, 255, 0.2);}
.tokenizer-token-2:hover {background: rgba(0, 0, 255, 0.4);}
.tokenizer-token-3 {background: rgba(255, 156, 0, 0.2);}
.tokenizer-token-3:hover {background: rgba(255, 156, 0, 0.4);}
}
@media (prefers-color-scheme: light) {
.tokenizer-token{
cursor: pointer;
}
.tokenizer-token-0 {background: rgba(255, 0, 0, 0.1);}
.tokenizer-token-0:hover {background: rgba(255, 0, 0, 0.2);}
.tokenizer-token-1 {background: rgba(0, 255, 0, 0.1);}
.tokenizer-token-1:hover {background: rgba(0, 255, 0, 0.2);}
.tokenizer-token-2 {background: rgba(0, 0, 255, 0.1);}
.tokenizer-token-2:hover {background: rgba(0, 0, 255, 0.2);}
.tokenizer-token-3 {background: rgba(255, 156, 0, 0.1);}
.tokenizer-token-3:hover {background: rgba(255, 156, 0, 0.2);}
}
"""
def tokenize(text, current_step=1, total_step=1, AND_block=0, simple_input=False, input_is_ids=False):
clip: FrozenCLIPEmbedder = shared.sd_model.cond_stage_model.wrapped
token_count = None
if input_is_ids:
tokens = [int(x.strip()) for x in text.split(",")]
elif simple_input:
tokens = clip.tokenizer(text, truncation=False, add_special_tokens=False)["input_ids"]
else:
from modules import sd_hijack, prompt_parser
from functools import reduce
_, prompt_flat_list, _ = prompt_parser.get_multicond_prompt_list([text])
prompt_schedules = prompt_parser.get_learned_conditioning_prompt_schedules(prompt_flat_list, int(total_step))
flat_prompts = reduce(lambda list1, list2: list1+list2, prompt_schedules)
prompts = [prompt_text for step, prompt_text in flat_prompts]
def find_current_prompt_idx(c_step, a_block):
_idx = 0
for i, prompts_block in enumerate(prompt_schedules):
for step_prompt_chunk in prompts_block:
if i == a_block:
if c_step <= step_prompt_chunk[0]:
return _idx
_idx += 1
idx = find_current_prompt_idx(current_step, AND_block)
tokens, token_count, max_length = [sd_hijack.model_hijack.tokenize(prompt) for prompt in prompts][idx]
vocab = {v: k for k, v in clip.tokenizer.get_vocab().items()}
code = ''
ids = []
current_ids = []
class_index = 0
def dump(last=False):
nonlocal code, ids, current_ids
words = [vocab.get(x, "") for x in current_ids]
def wordscode(ids, word):
nonlocal class_index
if ids != [clip.tokenizer.eos_token_id]:
res = f"""<span class='tokenizer-token tokenizer-token-{class_index%4}' title='{html.escape(", ".join([str(x) for x in ids]))}'>{html.escape(word)}</span>"""
else:
res = f"""<span class='tokenizer-token tokenizer-token-4' title='{html.escape(", ".join([str(x) for x in ids]))}'>{html.escape(word)}</span>"""
class_index += 1
return res
try:
word = bytearray([clip.tokenizer.byte_decoder[x] for x in ''.join(words)]).decode("utf-8")
except UnicodeDecodeError:
if last:
word = "" * len(current_ids)
elif len(current_ids) > 4:
id = current_ids[0]
ids += [id]
local_ids = current_ids[1:]
code += wordscode([id], "")
current_ids = []
for id in local_ids:
current_ids.append(id)
dump()
return
else:
return
word = word.replace("</w>", " ")
code += wordscode(current_ids, word)
ids += current_ids
current_ids = []
for token in tokens:
token = int(token)
current_ids.append(token)
dump()
dump(last=True)
if token_count is None:
token_count = len(ids)
ids_html = f"""
<p>
Token count: {token_count}/{len(ids)}<br>
{", ".join([str(x) for x in ids])}
</p>
"""
return code, ids_html
def add_tab():
with gr.Blocks(analytics_enabled=False, css=css) as ui:
gr.HTML(f"""
<style>{css}</style>
<p>
Before your text is sent to the neural network, it gets turned into numbers in a process called tokenization. These tokens are how the neural network reads and interprets text. Thanks to our great friends at Shousetsu愛 for inspiration for this feature.<br>
Depending on your setting, text will be first parsed by webui to calculate prompt attention like (text) and [text], and scheduler like [a:b:0.5], and the capital AND like a AND b before tokenization. This extension processes your text like this as well.<br>
To disable this feature, check on "Don't parse webui special grammar".
</p>
""")
with gr.Tabs() as tabs:
with gr.Tab("Text input", id="input_text"):
prompt = gr.Textbox(label="Prompt", elem_id="tokenizer_prompt", show_label=False, lines=8, placeholder="Prompt for tokenization")
go = gr.Button(value="Tokenize", variant="primary")
is_simple = gr.Checkbox(label="Don't parse webui special grammar", interactive=True)
with gr.Row():
current_step = gr.Number(label='Current sampling steps', value=1, step=1, interactive=True)
total_step = gr.Number(label='Total sampling steps', value=28, step=1, interactive=True)
and_block = gr.Number(label='Which block of prompts (separated by AND) to tokenize', value=0, step=1, interactive=True)
with gr.Tab("ID input", id="input_ids"):
prompt_ids = gr.Textbox(label="Prompt", elem_id="tokenizer_prompt", show_label=False, lines=8, placeholder="Ids for tokenization (example: 9061, 631, 736)")
go_ids = gr.Button(value="Tokenize", variant="primary")
with gr.Tabs():
with gr.Tab("Text"):
tokenized_text = gr.HTML(elem_id="tokenized_text")
with gr.Tab("Tokens"):
tokens = gr.HTML(elem_id="tokenized_tokens")
go.click(
fn=tokenize,
inputs=[prompt, current_step, total_step, and_block, is_simple],
outputs=[tokenized_text, tokens],
)
go_ids.click(
fn=lambda x: tokenize(x, input_is_ids=True),
inputs=[prompt_ids],
outputs=[tokenized_text, tokens],
)
return [(ui, "Tokenizer", "tokenizer")]
script_callbacks.on_ui_tabs(add_tab)