File size: 13,200 Bytes
5e052b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
import re
from collections import namedtuple
from typing import List
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):
# [25, 'fantasy landscape with a mountain and an oak in foreground shoddy']
# [50, 'fantasy landscape with a lake and an oak in foreground in background shoddy']
# [60, 'fantasy landscape with a lake and an oak in foreground in background masterful']
# [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 | alternate | plain | WHITESPACE)*
!emphasized: "(" prompt ")"
        | "(" prompt ":" prompt ")"
        | "[" prompt "]"
scheduled: "[" [prompt ":"] prompt ":" [WHITESPACE] NUMBER "]"
alternate: "[" prompt ("|" prompt)+ "]"
WHITESPACE: /\s+/
plain: /([^\\\[\]():|]|\\.)+/
%import common.SIGNED_NUMBER -> NUMBER
""")

def get_learned_conditioning_prompt_schedules(prompts, steps):
    """
    >>> 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']]
    >>> g("[a|(b:1.1)]")
    [[1, 'a'], [2, '(b:1.1)'], [3, 'a'], [4, '(b:1.1)'], [5, 'a'], [6, '(b:1.1)'], [7, 'a'], [8, '(b:1.1)'], [9, 'a'], [10, '(b:1.1)']]
    """

    def collect_steps(steps, tree):
        l = [steps]
        class CollectSteps(lark.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])
            def alternate(self, tree):
                l.extend(range(1, steps+1))
        CollectSteps().visit(tree)
        return sorted(set(l))

    def at_step(step, tree):
        class AtStep(lark.Transformer):
            def scheduled(self, args):
                before, after, _, when = args
                yield before or () if step <= when else after
            def alternate(self, args):
                yield next(args[(step - 1)%len(args)])
            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))
            def plain(self, args):
                yield args[0].value
            def __default__(self, data, children, meta):
                for child in children:
                    yield child
        return AtStep().transform(tree)

    def get_schedule(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)}
    return [promptdict[prompt] for prompt in prompts]


ScheduledPromptConditioning = namedtuple("ScheduledPromptConditioning", ["end_at_step", "cond"])


def get_learned_conditioning(model, prompts, steps):
    """converts a list of prompts into a list of prompt schedules - each schedule is a list of ScheduledPromptConditioning, specifying the comdition (cond),
    and the sampling step at which this condition is to be replaced by the next one.

    Input:
    (model, ['a red crown', 'a [blue:green:5] jeweled crown'], 20)

    Output:
    [
        [
            ScheduledPromptConditioning(end_at_step=20, cond=tensor([[-0.3886,  0.0229, -0.0523,  ..., -0.4901, -0.3066,  0.0674], ..., [ 0.3317, -0.5102, -0.4066,  ...,  0.4119, -0.7647, -1.0160]], device='cuda:0'))
        ],
        [
            ScheduledPromptConditioning(end_at_step=5, cond=tensor([[-0.3886,  0.0229, -0.0522,  ..., -0.4901, -0.3067,  0.0673], ..., [-0.0192,  0.3867, -0.4644,  ...,  0.1135, -0.3696, -0.4625]], device='cuda:0')),
            ScheduledPromptConditioning(end_at_step=20, cond=tensor([[-0.3886,  0.0229, -0.0522,  ..., -0.4901, -0.3067,  0.0673], ..., [-0.7352, -0.4356, -0.7888,  ...,  0.6994, -0.4312, -1.2593]], device='cuda:0'))
        ]
    ]
    """
    res = []

    prompt_schedules = get_learned_conditioning_prompt_schedules(prompts, steps)
    cache = {}

    for prompt, prompt_schedule in zip(prompts, prompt_schedules):

        cached = cache.get(prompt, None)
        if cached is not None:
            res.append(cached)
            continue

        texts = [x[1] for x in prompt_schedule]
        conds = model.get_learned_conditioning(texts)

        cond_schedule = []
        for i, (end_at_step, text) in enumerate(prompt_schedule):
            cond_schedule.append(ScheduledPromptConditioning(end_at_step, conds[i]))

        cache[prompt] = cond_schedule
        res.append(cond_schedule)

    return res


re_AND = re.compile(r"\bAND\b")
re_weight = re.compile(r"^(.*?)(?:\s*:\s*([-+]?(?:\d+\.?|\d*\.\d+)))?\s*$")

def get_multicond_prompt_list(prompts):
    res_indexes = []

    prompt_flat_list = []
    prompt_indexes = {}

    for prompt in prompts:
        subprompts = re_AND.split(prompt)

        indexes = []
        for subprompt in subprompts:
            match = re_weight.search(subprompt)

            text, weight = match.groups() if match is not None else (subprompt, 1.0)

            weight = float(weight) if weight is not None else 1.0

            index = prompt_indexes.get(text, None)
            if index is None:
                index = len(prompt_flat_list)
                prompt_flat_list.append(text)
                prompt_indexes[text] = index

            indexes.append((index, weight))

        res_indexes.append(indexes)

    return res_indexes, prompt_flat_list, prompt_indexes


class ComposableScheduledPromptConditioning:
    def __init__(self, schedules, weight=1.0):
        self.schedules: List[ScheduledPromptConditioning] = schedules
        self.weight: float = weight


class MulticondLearnedConditioning:
    def __init__(self, shape, batch):
        self.shape: tuple = shape  # the shape field is needed to send this object to DDIM/PLMS
        self.batch: List[List[ComposableScheduledPromptConditioning]] = batch

def get_multicond_learned_conditioning(model, prompts, steps) -> MulticondLearnedConditioning:
    """same as get_learned_conditioning, but returns a list of ScheduledPromptConditioning along with the weight objects for each prompt.
    For each prompt, the list is obtained by splitting the prompt using the AND separator.

    https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/
    """

    res_indexes, prompt_flat_list, prompt_indexes = get_multicond_prompt_list(prompts)

    learned_conditioning = get_learned_conditioning(model, prompt_flat_list, steps)

    res = []
    for indexes in res_indexes:
        res.append([ComposableScheduledPromptConditioning(learned_conditioning[i], weight) for i, weight in indexes])

    return MulticondLearnedConditioning(shape=(len(prompts),), batch=res)


def reconstruct_cond_batch(c: List[List[ScheduledPromptConditioning]], current_step):
    param = c[0][0].cond
    res = torch.zeros((len(c),) + param.shape, device=param.device, dtype=param.dtype)
    for i, cond_schedule in enumerate(c):
        target_index = 0
        for current, (end_at, cond) in enumerate(cond_schedule):
            if current_step <= end_at:
                target_index = current
                break
        res[i] = cond_schedule[target_index].cond

    return res


def reconstruct_multicond_batch(c: MulticondLearnedConditioning, current_step):
    param = c.batch[0][0].schedules[0].cond

    tensors = []
    conds_list = []

    for batch_no, composable_prompts in enumerate(c.batch):
        conds_for_batch = []

        for cond_index, composable_prompt in enumerate(composable_prompts):
            target_index = 0
            for current, (end_at, cond) in enumerate(composable_prompt.schedules):
                if current_step <= end_at:
                    target_index = current
                    break

            conds_for_batch.append((len(tensors), composable_prompt.weight))
            tensors.append(composable_prompt.schedules[target_index].cond)

        conds_list.append(conds_for_batch)

    # if prompts have wildly different lengths above the limit we'll get tensors fo different shapes
    # and won't be able to torch.stack them. So this fixes that.
    token_count = max([x.shape[0] for x in tensors])
    for i in range(len(tensors)):
        if tensors[i].shape[0] != token_count:
            last_vector = tensors[i][-1:]
            last_vector_repeated = last_vector.repeat([token_count - tensors[i].shape[0], 1])
            tensors[i] = torch.vstack([tensors[i], last_vector_repeated])

    return conds_list, torch.stack(tensors).to(device=param.device, dtype=param.dtype)


re_attention = re.compile(r"""
\\\(|
\\\)|
\\\[|
\\]|
\\\\|
\\|
\(|
\[|
:([+-]?[.\d]+)\)|
\)|
]|
[^\\()\[\]:]+|
:
""", re.X)

re_break = re.compile(r"\s*\bBREAK\b\s*", re.S)

def parse_prompt_attention(text):
    """
    Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
    Accepted tokens are:
      (abc) - increases attention to abc by a multiplier of 1.1
      (abc:3.12) - increases attention to abc by a multiplier of 3.12
      [abc] - decreases attention to abc by a multiplier of 1.1
      \( - literal character '('
      \[ - literal character '['
      \) - literal character ')'
      \] - literal character ']'
      \\ - literal character '\'
      anything else - just text

    >>> 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 = []
    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

    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 len(round_brackets) > 0:
            multiply_range(round_brackets.pop(), float(weight))
        elif text == ')' and len(round_brackets) > 0:
            multiply_range(round_brackets.pop(), round_bracket_multiplier)
        elif text == ']' and len(square_brackets) > 0:
            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]]

    # 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