From 2f1b61d97987ae0a52a7dfc6bc99c68928bdb594 Mon Sep 17 00:00:00 2001 From: dan Date: Mon, 3 Oct 2022 19:25:36 +0800 Subject: Allow nested structures inside schedules --- modules/prompt_parser.py | 119 +++++++++++++++++++++-------------------------- 1 file changed, 53 insertions(+), 66 deletions(-) (limited to 'modules/prompt_parser.py') diff --git a/modules/prompt_parser.py b/modules/prompt_parser.py index e811eb9e..99c8ed99 100644 --- a/modules/prompt_parser.py +++ b/modules/prompt_parser.py @@ -1,20 +1,11 @@ import re from collections import namedtuple import torch +from lark import Lark, Transformer, Visitor +import functools import modules.shared as shared -re_prompt = re.compile(r''' -(.*?) -\[ - ([^]:]+): - (?:([^]:]*):)? - ([0-9]*\.?[0-9]+) -] -| -(.+) -''', re.X) - # a prompt like this: "fantasy landscape with a [mountain:lake:0.25] and [an oak:a christmas tree:0.75][ in foreground::0.6][ in background:0.25] [shoddy:masterful:0.5]" # will be represented with prompt_schedule like this (assuming steps=100): # [25, 'fantasy landscape with a mountain and an oak in foreground shoddy'] @@ -25,61 +16,57 @@ re_prompt = re.compile(r''' def get_learned_conditioning_prompt_schedules(prompts, steps): - res = [] - cache = {} - - for prompt in prompts: - prompt_schedule: list[list[str | int]] = [[steps, ""]] - - cached = cache.get(prompt, None) - if cached is not None: - res.append(cached) - continue - - for m in re_prompt.finditer(prompt): - plaintext = m.group(1) if m.group(5) is None else m.group(5) - concept_from = m.group(2) - concept_to = m.group(3) - if concept_to is None: - concept_to = concept_from - concept_from = "" - swap_position = float(m.group(4)) if m.group(4) is not None else None - - if swap_position is not None: - if swap_position < 1: - swap_position = swap_position * steps - swap_position = int(min(swap_position, steps)) - - swap_index = None - found_exact_index = False - for i in range(len(prompt_schedule)): - end_step = prompt_schedule[i][0] - prompt_schedule[i][1] += plaintext - - if swap_position is not None and swap_index is None: - if swap_position == end_step: - swap_index = i - found_exact_index = True - - if swap_position < end_step: - swap_index = i - - if swap_index is not None: - if not found_exact_index: - prompt_schedule.insert(swap_index, [swap_position, prompt_schedule[swap_index][1]]) - - for i in range(len(prompt_schedule)): - end_step = prompt_schedule[i][0] - must_replace = swap_position < end_step - - prompt_schedule[i][1] += concept_to if must_replace else concept_from - - res.append(prompt_schedule) - cache[prompt] = prompt_schedule - #for t in prompt_schedule: - # print(t) - - return res + grammar = r""" + start: prompt + prompt: (emphasized | scheduled | weighted | plain)* + !emphasized: "(" prompt ")" + | "(" prompt ":" prompt ")" + | "[" prompt "]" + scheduled: "[" (prompt ":")? prompt ":" NUMBER "]" + !weighted: "{" weighted_item ("|" weighted_item)* "}" + !weighted_item: prompt (":" prompt)? + plain: /([^\\\[\](){}:|]|\\.)+/ + %import common.SIGNED_NUMBER -> NUMBER + """ + parser = Lark(grammar, parser='lalr') + def collect_steps(steps, tree): + l = [steps] + class CollectSteps(Visitor): + def scheduled(self, tree): + tree.children[-1] = float(tree.children[-1]) + if tree.children[-1] < 1: + tree.children[-1] *= steps + tree.children[-1] = min(steps, int(tree.children[-1])) + l.append(tree.children[-1]) + CollectSteps().visit(tree) + return sorted(set(l)) + def at_step(step, tree): + class AtStep(Transformer): + def scheduled(self, args): + if len(args) == 2: + before, after, when = (), *args + else: + before, after, when = args + yield before if step <= when else after + def start(self, args): + def flatten(x): + if type(x) == str: + yield x + else: + for gen in x: + yield from flatten(gen) + return ''.join(flatten(args[0])) + def plain(self, args): + yield args[0].value + def __default__(self, data, children, meta): + for child in children: + yield from child + return AtStep().transform(tree) + @functools.cache + def get_schedule(prompt): + tree = parser.parse(prompt) + return [[t, at_step(t, tree)] for t in collect_steps(steps, tree)] + return [get_schedule(prompt) for prompt in prompts] ScheduledPromptConditioning = namedtuple("ScheduledPromptConditioning", ["end_at_step", "cond"]) -- cgit v1.2.3 From 1eb588cbf19924333b88beaa1ac0041904966640 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 4 Oct 2022 18:02:01 +0300 Subject: remove functools.cache as some people are having issues with it --- modules/prompt_parser.py | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'modules/prompt_parser.py') diff --git a/modules/prompt_parser.py b/modules/prompt_parser.py index 99c8ed99..5d58c4ed 100644 --- a/modules/prompt_parser.py +++ b/modules/prompt_parser.py @@ -29,6 +29,7 @@ def get_learned_conditioning_prompt_schedules(prompts, steps): %import common.SIGNED_NUMBER -> NUMBER """ parser = Lark(grammar, parser='lalr') + def collect_steps(steps, tree): l = [steps] class CollectSteps(Visitor): @@ -40,6 +41,7 @@ def get_learned_conditioning_prompt_schedules(prompts, steps): l.append(tree.children[-1]) CollectSteps().visit(tree) return sorted(set(l)) + def at_step(step, tree): class AtStep(Transformer): def scheduled(self, args): @@ -62,11 +64,13 @@ def get_learned_conditioning_prompt_schedules(prompts, steps): for child in children: yield from child return AtStep().transform(tree) - @functools.cache + def get_schedule(prompt): tree = parser.parse(prompt) return [[t, at_step(t, tree)] for t in collect_steps(steps, tree)] - return [get_schedule(prompt) for prompt in prompts] + + promptdict = {prompt: get_schedule(prompt) for prompt in set(prompts)} + return [promptdict[prompt] for prompt in prompts] ScheduledPromptConditioning = namedtuple("ScheduledPromptConditioning", ["end_at_step", "cond"]) -- cgit v1.2.3 From 90e911fd546e76f879b38a764473569911a0f845 Mon Sep 17 00:00:00 2001 From: Rae Fu Date: Tue, 4 Oct 2022 09:49:51 -0600 Subject: prompt_parser: allow spaces in schedules, add test, log/ignore errors Only build the parser once (at import time) instead of for each step. doctest is run by simply executing modules/prompt_parser.py --- modules/prompt_parser.py | 139 ++++++++++++++++++++++++++++++----------------- 1 file changed, 90 insertions(+), 49 deletions(-) (limited to 'modules/prompt_parser.py') diff --git a/modules/prompt_parser.py b/modules/prompt_parser.py index 5d58c4ed..a3b12421 100644 --- a/modules/prompt_parser.py +++ b/modules/prompt_parser.py @@ -1,10 +1,7 @@ import re from collections import namedtuple -import torch -from lark import Lark, Transformer, Visitor -import functools -import modules.shared as shared +import lark # a prompt like this: "fantasy landscape with a [mountain:lake:0.25] and [an oak:a christmas tree:0.75][ in foreground::0.6][ in background:0.25] [shoddy:masterful:0.5]" # will be represented with prompt_schedule like this (assuming steps=100): @@ -14,25 +11,48 @@ import modules.shared as shared # [75, 'fantasy landscape with a lake and an oak in background masterful'] # [100, 'fantasy landscape with a lake and a christmas tree in background masterful'] +schedule_parser = lark.Lark(r""" +!start: (prompt | /[][():]/+)* +prompt: (emphasized | scheduled | plain | WHITESPACE)* +!emphasized: "(" prompt ")" + | "(" prompt ":" prompt ")" + | "[" prompt "]" +scheduled: "[" [prompt ":"] prompt ":" [WHITESPACE] NUMBER "]" +WHITESPACE: /\s+/ +plain: /([^\\\[\]():]|\\.)+/ +%import common.SIGNED_NUMBER -> NUMBER +""") def get_learned_conditioning_prompt_schedules(prompts, steps): - grammar = r""" - start: prompt - prompt: (emphasized | scheduled | weighted | plain)* - !emphasized: "(" prompt ")" - | "(" prompt ":" prompt ")" - | "[" prompt "]" - scheduled: "[" (prompt ":")? prompt ":" NUMBER "]" - !weighted: "{" weighted_item ("|" weighted_item)* "}" - !weighted_item: prompt (":" prompt)? - plain: /([^\\\[\](){}:|]|\\.)+/ - %import common.SIGNED_NUMBER -> NUMBER """ - parser = Lark(grammar, parser='lalr') + >>> g = lambda p: get_learned_conditioning_prompt_schedules([p], 10)[0] + >>> g("test") + [[10, 'test']] + >>> g("a [b:3]") + [[3, 'a '], [10, 'a b']] + >>> g("a [b: 3]") + [[3, 'a '], [10, 'a b']] + >>> g("a [[[b]]:2]") + [[2, 'a '], [10, 'a [[b]]']] + >>> g("[(a:2):3]") + [[3, ''], [10, '(a:2)']] + >>> g("a [b : c : 1] d") + [[1, 'a b d'], [10, 'a c d']] + >>> g("a[b:[c:d:2]:1]e") + [[1, 'abe'], [2, 'ace'], [10, 'ade']] + >>> g("a [unbalanced") + [[10, 'a [unbalanced']] + >>> g("a [b:.5] c") + [[5, 'a c'], [10, 'a b c']] + >>> g("a [{b|d{:.5] c") # not handling this right now + [[5, 'a c'], [10, 'a {b|d{ c']] + >>> g("((a][:b:c [d:3]") + [[3, '((a][:b:c '], [10, '((a][:b:c d']] + """ def collect_steps(steps, tree): l = [steps] - class CollectSteps(Visitor): + class CollectSteps(lark.Visitor): def scheduled(self, tree): tree.children[-1] = float(tree.children[-1]) if tree.children[-1] < 1: @@ -43,13 +63,10 @@ def get_learned_conditioning_prompt_schedules(prompts, steps): return sorted(set(l)) def at_step(step, tree): - class AtStep(Transformer): + class AtStep(lark.Transformer): def scheduled(self, args): - if len(args) == 2: - before, after, when = (), *args - else: - before, after, when = args - yield before if step <= when else after + before, after, _, when = args + yield before or () if step <= when else after def start(self, args): def flatten(x): if type(x) == str: @@ -57,16 +74,22 @@ def get_learned_conditioning_prompt_schedules(prompts, steps): else: for gen in x: yield from flatten(gen) - return ''.join(flatten(args[0])) + return ''.join(flatten(args)) def plain(self, args): yield args[0].value def __default__(self, data, children, meta): for child in children: yield from child return AtStep().transform(tree) - + def get_schedule(prompt): - tree = parser.parse(prompt) + try: + tree = schedule_parser.parse(prompt) + except lark.exceptions.LarkError as e: + if 0: + import traceback + traceback.print_exc() + return [[steps, prompt]] return [[t, at_step(t, tree)] for t in collect_steps(steps, tree)] promptdict = {prompt: get_schedule(prompt) for prompt in set(prompts)} @@ -77,8 +100,7 @@ ScheduledPromptConditioning = namedtuple("ScheduledPromptConditioning", ["end_at ScheduledPromptBatch = namedtuple("ScheduledPromptBatch", ["shape", "schedules"]) -def get_learned_conditioning(prompts, steps): - +def get_learned_conditioning(model, prompts, steps): res = [] prompt_schedules = get_learned_conditioning_prompt_schedules(prompts, steps) @@ -92,7 +114,7 @@ def get_learned_conditioning(prompts, steps): continue texts = [x[1] for x in prompt_schedule] - conds = shared.sd_model.get_learned_conditioning(texts) + conds = model.get_learned_conditioning(texts) cond_schedule = [] for i, (end_at_step, text) in enumerate(prompt_schedule): @@ -105,12 +127,13 @@ def get_learned_conditioning(prompts, steps): def reconstruct_cond_batch(c: ScheduledPromptBatch, current_step): - res = torch.zeros(c.shape, device=shared.device, dtype=next(shared.sd_model.parameters()).dtype) + param = c.schedules[0][0].cond + res = torch.zeros(c.shape, device=param.device, dtype=param.dtype) for i, cond_schedule in enumerate(c.schedules): target_index = 0 - for curret_index, (end_at, cond) in enumerate(cond_schedule): + for current, (end_at, cond) in enumerate(cond_schedule): if current_step <= end_at: - target_index = curret_index + target_index = current break res[i] = cond_schedule[target_index].cond @@ -148,23 +171,26 @@ def parse_prompt_attention(text): \\ - literal character '\' anything else - just text - Example: - - 'a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).' - - produces: - - [ - ['a ', 1.0], - ['house', 1.5730000000000004], - [' ', 1.1], - ['on', 1.0], - [' a ', 1.1], - ['hill', 0.55], - [', sun, ', 1.1], - ['sky', 1.4641000000000006], - ['.', 1.1] - ] + >>> parse_prompt_attention('normal text') + [['normal text', 1.0]] + >>> parse_prompt_attention('an (important) word') + [['an ', 1.0], ['important', 1.1], [' word', 1.0]] + >>> parse_prompt_attention('(unbalanced') + [['unbalanced', 1.1]] + >>> parse_prompt_attention('\(literal\]') + [['(literal]', 1.0]] + >>> parse_prompt_attention('(unnecessary)(parens)') + [['unnecessaryparens', 1.1]] + >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).') + [['a ', 1.0], + ['house', 1.5730000000000004], + [' ', 1.1], + ['on', 1.0], + [' a ', 1.1], + ['hill', 0.55], + [', sun, ', 1.1], + ['sky', 1.4641000000000006], + ['.', 1.1]] """ res = [] @@ -206,4 +232,19 @@ def parse_prompt_attention(text): if len(res) == 0: res = [["", 1.0]] + # merge runs of identical weights + 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 + +if __name__ == "__main__": + import doctest + doctest.testmod(optionflags=doctest.NORMALIZE_WHITESPACE) +else: + import torch # doctest faster -- cgit v1.2.3