sd1_clip.py 21.5 KB
Newer Older
comfyanonymous's avatar
comfyanonymous committed
1
2
import os

3
from transformers import CLIPTokenizer
4
import comfy.ops
comfyanonymous's avatar
comfyanonymous committed
5
import torch
6
import traceback
7
import zipfile
8
from . import model_management
9
10
import comfy.clip_model
import json
11
import logging
Mario Klingemann's avatar
Mario Klingemann committed
12
import numbers
comfyanonymous's avatar
comfyanonymous committed
13

14
15
16
17
18
19
20
21
22
23
24
25
def gen_empty_tokens(special_tokens, length):
    start_token = special_tokens.get("start", None)
    end_token = special_tokens.get("end", None)
    pad_token = special_tokens.get("pad")
    output = []
    if start_token is not None:
        output.append(start_token)
    if end_token is not None:
        output.append(end_token)
    output += [pad_token] * (length - len(output))
    return output

comfyanonymous's avatar
comfyanonymous committed
26
27
class ClipTokenWeightEncoder:
    def encode_token_weights(self, token_weight_pairs):
28
29
30
        to_encode = list()
        max_token_len = 0
        has_weights = False
comfyanonymous's avatar
comfyanonymous committed
31
        for x in token_weight_pairs:
32
            tokens = list(map(lambda a: a[0], x))
33
34
            max_token_len = max(len(tokens), max_token_len)
            has_weights = has_weights or not all(map(lambda a: a[1] == 1.0, x))
35
36
            to_encode.append(tokens)

37
38
39
40
        sections = len(to_encode)
        if has_weights or sections == 0:
            to_encode.append(gen_empty_tokens(self.special_tokens, max_token_len))

41
42
43
        o = self.encode(to_encode)
        out, pooled = o[:2]

44
        if pooled is not None:
45
            first_pooled = pooled[0:1].to(model_management.intermediate_device())
46
        else:
47
            first_pooled = pooled
48
49

        output = []
50
        for k in range(0, sections):
51
            z = out[k:k+1]
52
53
54
55
56
57
58
            if has_weights:
                z_empty = out[-1]
                for i in range(len(z)):
                    for j in range(len(z[i])):
                        weight = token_weight_pairs[k][j][1]
                        if weight != 1.0:
                            z[i][j] = (z[i][j] - z_empty[j]) * weight + z_empty[j]
59
60
            output.append(z)

comfyanonymous's avatar
comfyanonymous committed
61
        if (len(output) == 0):
62
63
64
            r = (out[-1:].to(model_management.intermediate_device()), first_pooled)
        else:
            r = (torch.cat(output, dim=-2).to(model_management.intermediate_device()), first_pooled)
65
66
67
68
69
70
71
72
73
74

        if len(o) > 2:
            extra = {}
            for k in o[2]:
                v = o[2][k]
                if k == "attention_mask":
                    v = v[:sections].flatten().unsqueeze(dim=0).to(model_management.intermediate_device())
                extra[k] = v

            r = r + (extra,)
75
        return r
comfyanonymous's avatar
comfyanonymous committed
76

77
class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
comfyanonymous's avatar
comfyanonymous committed
78
79
80
81
82
83
84
    """Uses the CLIP transformer encoder for text (from huggingface)"""
    LAYERS = [
        "last",
        "pooled",
        "hidden"
    ]
    def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
85
                 freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel,
86
                 special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False,
87
                 return_projected_pooled=True, return_attention_masks=False):  # clip-vit-base-patch32
comfyanonymous's avatar
comfyanonymous committed
88
89
        super().__init__()
        assert layer in self.LAYERS
90
91
92
93
94
95
96

        if textmodel_json_config is None:
            textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json")

        with open(textmodel_json_config) as f:
            config = json.load(f)

97
        self.transformer = model_class(config, dtype, device, comfy.ops.manual_cast)
98
        self.num_layers = self.transformer.num_layers
99

comfyanonymous's avatar
comfyanonymous committed
100
101
102
103
104
        self.max_length = max_length
        if freeze:
            self.freeze()
        self.layer = layer
        self.layer_idx = None
105
        self.special_tokens = special_tokens
106

107
        self.logit_scale = torch.nn.Parameter(torch.tensor(4.6055))
108
        self.enable_attention_masks = enable_attention_masks
109
        self.zero_out_masked = zero_out_masked
110

111
        self.layer_norm_hidden_state = layer_norm_hidden_state
112
        self.return_projected_pooled = return_projected_pooled
113
        self.return_attention_masks = return_attention_masks
114

comfyanonymous's avatar
comfyanonymous committed
115
116
        if layer == "hidden":
            assert layer_idx is not None
117
            assert abs(layer_idx) < self.num_layers
118
119
            self.set_clip_options({"layer": layer_idx})
        self.options_default = (self.layer, self.layer_idx, self.return_projected_pooled)
comfyanonymous's avatar
comfyanonymous committed
120
121
122
123
124
125
126

    def freeze(self):
        self.transformer = self.transformer.eval()
        #self.train = disabled_train
        for param in self.parameters():
            param.requires_grad = False

127
128
129
130
    def set_clip_options(self, options):
        layer_idx = options.get("layer", self.layer_idx)
        self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled)
        if layer_idx is None or abs(layer_idx) > self.num_layers:
comfyanonymous's avatar
comfyanonymous committed
131
132
133
134
135
            self.layer = "last"
        else:
            self.layer = "hidden"
            self.layer_idx = layer_idx

136
137
138
139
    def reset_clip_options(self):
        self.layer = self.options_default[0]
        self.layer_idx = self.options_default[1]
        self.return_projected_pooled = self.options_default[2]
140

141
142
    def set_up_textual_embeddings(self, tokens, current_embeds):
        out_tokens = []
143
        next_new_token = token_dict_size = current_embeds.weight.shape[0]
144
145
146
147
148
        embedding_weights = []

        for x in tokens:
            tokens_temp = []
            for y in x:
Mario Klingemann's avatar
Mario Klingemann committed
149
150
                if isinstance(y, numbers.Integral):
                    tokens_temp += [int(y)]
151
                else:
152
153
154
155
156
                    if y.shape[0] == current_embeds.weight.shape[1]:
                        embedding_weights += [y]
                        tokens_temp += [next_new_token]
                        next_new_token += 1
                    else:
157
                        logging.warning("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored {} != {}".format(y.shape[0], current_embeds.weight.shape[1]))
158
            while len(tokens_temp) < len(x):
159
                tokens_temp += [self.special_tokens["pad"]]
160
161
            out_tokens += [tokens_temp]

162
        n = token_dict_size
163
        if len(embedding_weights) > 0:
164
            new_embedding = torch.nn.Embedding(next_new_token + 1, current_embeds.weight.shape[1], device=current_embeds.weight.device, dtype=current_embeds.weight.dtype)
165
            new_embedding.weight[:token_dict_size] = current_embeds.weight
166
167
168
169
            for x in embedding_weights:
                new_embedding.weight[n] = x
                n += 1
            self.transformer.set_input_embeddings(new_embedding)
170
171
172
173
174
175

        processed_tokens = []
        for x in out_tokens:
            processed_tokens += [list(map(lambda a: n if a == -1 else a, x))] #The EOS token should always be the largest one

        return processed_tokens
176

comfyanonymous's avatar
comfyanonymous committed
177
    def forward(self, tokens):
178
        backup_embeds = self.transformer.get_input_embeddings()
179
        device = backup_embeds.weight.device
180
        tokens = self.set_up_textual_embeddings(tokens, backup_embeds)
181
182
        tokens = torch.LongTensor(tokens).to(device)

183
        attention_mask = None
184
        if self.enable_attention_masks or self.zero_out_masked or self.return_attention_masks:
185
            attention_mask = torch.zeros_like(tokens)
186
            end_token = self.special_tokens.get("end", -1)
187
188
189
            for x in range(attention_mask.shape[0]):
                for y in range(attention_mask.shape[1]):
                    attention_mask[x, y] = 1
190
                    if tokens[x, y] == end_token:
191
192
                        break

193
194
195
196
197
        attention_mask_model = None
        if self.enable_attention_masks:
            attention_mask_model = attention_mask

        outputs = self.transformer(tokens, attention_mask_model, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state)
198
199
200
        self.transformer.set_input_embeddings(backup_embeds)

        if self.layer == "last":
201
            z = outputs[0].float()
comfyanonymous's avatar
comfyanonymous committed
202
        else:
203
204
            z = outputs[1].float()

205
        if self.zero_out_masked:
206
            z *= attention_mask.unsqueeze(-1).float()
207

208
209
210
211
212
213
        pooled_output = None
        if len(outputs) >= 3:
            if not self.return_projected_pooled and len(outputs) >= 4 and outputs[3] is not None:
                pooled_output = outputs[3].float()
            elif outputs[2] is not None:
                pooled_output = outputs[2].float()
214

215
        extra = {}
216
        if self.return_attention_masks:
217
218
219
220
            extra["attention_mask"] = attention_mask

        if len(extra) > 0:
            return z, pooled_output, extra
221

222
        return z, pooled_output
comfyanonymous's avatar
comfyanonymous committed
223
224
225
226

    def encode(self, tokens):
        return self(tokens)

227
228
229
    def load_sd(self, sd):
        return self.transformer.load_state_dict(sd, strict=False)

comfyanonymous's avatar
comfyanonymous committed
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
def parse_parentheses(string):
    result = []
    current_item = ""
    nesting_level = 0
    for char in string:
        if char == "(":
            if nesting_level == 0:
                if current_item:
                    result.append(current_item)
                    current_item = "("
                else:
                    current_item = "("
            else:
                current_item += char
            nesting_level += 1
        elif char == ")":
            nesting_level -= 1
            if nesting_level == 0:
                result.append(current_item + ")")
                current_item = ""
            else:
                current_item += char
        else:
            current_item += char
    if current_item:
        result.append(current_item)
    return result

def token_weights(string, current_weight):
    a = parse_parentheses(string)
    out = []
    for x in a:
        weight = current_weight
        if len(x) >= 2 and x[-1] == ')' and x[0] == '(':
            x = x[1:-1]
            xx = x.rfind(":")
            weight *= 1.1
            if xx > 0:
                try:
                    weight = float(x[xx+1:])
                    x = x[:xx]
                except:
                    pass
            out += token_weights(x, weight)
        else:
            out += [(x, current_weight)]
    return out

def escape_important(text):
    text = text.replace("\\)", "\0\1")
    text = text.replace("\\(", "\0\2")
    return text

def unescape_important(text):
    text = text.replace("\0\1", ")")
    text = text.replace("\0\2", "(")
    return text

288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
def safe_load_embed_zip(embed_path):
    with zipfile.ZipFile(embed_path) as myzip:
        names = list(filter(lambda a: "data/" in a, myzip.namelist()))
        names.reverse()
        for n in names:
            with myzip.open(n) as myfile:
                data = myfile.read()
                number = len(data) // 4
                length_embed = 1024 #sd2.x
                if number < 768:
                    continue
                if number % 768 == 0:
                    length_embed = 768 #sd1.x
                num_embeds = number // length_embed
                embed = torch.frombuffer(data, dtype=torch.float)
                out = embed.reshape((num_embeds, length_embed)).clone()
                del embed
                return out

307
308
309
310
311
312
313
def expand_directory_list(directories):
    dirs = set()
    for x in directories:
        dirs.add(x)
        for root, subdir, file in os.walk(x, followlinks=True):
            dirs.add(root)
    return list(dirs)
314

315
def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=None):
316
317
318
    if isinstance(embedding_directory, str):
        embedding_directory = [embedding_directory]

319
320
    embedding_directory = expand_directory_list(embedding_directory)

321
322
    valid_file = None
    for embed_dir in embedding_directory:
323
324
325
326
327
328
329
        embed_path = os.path.abspath(os.path.join(embed_dir, embedding_name))
        embed_dir = os.path.abspath(embed_dir)
        try:
            if os.path.commonpath((embed_dir, embed_path)) != embed_dir:
                continue
        except:
            continue
330
331
332
333
334
335
336
        if not os.path.isfile(embed_path):
            extensions = ['.safetensors', '.pt', '.bin']
            for x in extensions:
                t = embed_path + x
                if os.path.isfile(t):
                    valid_file = t
                    break
337
        else:
338
339
340
341
342
343
344
345
            valid_file = embed_path
        if valid_file is not None:
            break

    if valid_file is None:
        return None

    embed_path = valid_file
346

347
348
    embed_out = None

349
350
351
352
    try:
        if embed_path.lower().endswith(".safetensors"):
            import safetensors.torch
            embed = safetensors.torch.load_file(embed_path, device="cpu")
comfyanonymous's avatar
comfyanonymous committed
353
        else:
354
            if 'weights_only' in torch.load.__code__.co_varnames:
355
356
357
358
                try:
                    embed = torch.load(embed_path, weights_only=True, map_location="cpu")
                except:
                    embed_out = safe_load_embed_zip(embed_path)
359
360
361
            else:
                embed = torch.load(embed_path, map_location="cpu")
    except Exception as e:
362
        logging.warning("{}\n\nerror loading embedding, skipping loading: {}".format(traceback.format_exc(), embedding_name))
363
364
        return None

365
366
367
    if embed_out is None:
        if 'string_to_param' in embed:
            values = embed['string_to_param'].values()
368
369
370
371
372
373
374
375
376
377
            embed_out = next(iter(values))
        elif isinstance(embed, list):
            out_list = []
            for x in range(len(embed)):
                for k in embed[x]:
                    t = embed[x][k]
                    if t.shape[-1] != embedding_size:
                        continue
                    out_list.append(t.reshape(-1, t.shape[-1]))
            embed_out = torch.cat(out_list, dim=0)
378
379
        elif embed_key is not None and embed_key in embed:
            embed_out = embed[embed_key]
380
381
        else:
            values = embed.values()
382
            embed_out = next(iter(values))
383
    return embed_out
384

385
class SDTokenizer:
386
    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, pad_to_max_length=True, min_length=None, pad_token=None, tokenizer_data={}):
comfyanonymous's avatar
comfyanonymous committed
387
388
        if tokenizer_path is None:
            tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
389
        self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path)
comfyanonymous's avatar
comfyanonymous committed
390
        self.max_length = max_length
391
        self.min_length = min_length
392

comfyanonymous's avatar
comfyanonymous committed
393
        empty = self.tokenizer('')["input_ids"]
394
395
396
397
398
399
400
401
        if has_start_token:
            self.tokens_start = 1
            self.start_token = empty[0]
            self.end_token = empty[1]
        else:
            self.tokens_start = 0
            self.start_token = None
            self.end_token = empty[0]
402
403
404
405
406
407
408
409

        if pad_token is not None:
            self.pad_token = pad_token
        elif pad_with_end:
            self.pad_token = self.end_token
        else:
            self.pad_token = 0

comfyanonymous's avatar
comfyanonymous committed
410
        self.pad_with_end = pad_with_end
411
412
        self.pad_to_max_length = pad_to_max_length

comfyanonymous's avatar
comfyanonymous committed
413
414
        vocab = self.tokenizer.get_vocab()
        self.inv_vocab = {v: k for k, v in vocab.items()}
415
416
        self.embedding_directory = embedding_directory
        self.max_word_length = 8
417
        self.embedding_identifier = "embedding:"
418
        self.embedding_size = embedding_size
419
        self.embedding_key = embedding_key
420

421
    def _try_get_embedding(self, embedding_name:str):
422
423
424
425
        '''
        Takes a potential embedding name and tries to retrieve it.
        Returns a Tuple consisting of the embedding and any leftover string, embedding can be None.
        '''
426
        embed = load_embed(embedding_name, self.embedding_directory, self.embedding_size, self.embedding_key)
427
428
429
        if embed is None:
            stripped = embedding_name.strip(',')
            if len(stripped) < len(embedding_name):
430
                embed = load_embed(stripped, self.embedding_directory, self.embedding_size, self.embedding_key)
431
432
433
434
                return (embed, embedding_name[len(stripped):])
        return (embed, "")


435
    def tokenize_with_weights(self, text:str, return_word_ids=False):
436
437
438
439
440
441
        '''
        Takes a prompt and converts it to a list of (token, weight, word id) elements.
        Tokens can both be integer tokens and pre computed CLIP tensors.
        Word id values are unique per word and embedding, where the id 0 is reserved for non word tokens.
        Returned list has the dimensions NxM where M is the input size of CLIP
        '''
comfyanonymous's avatar
comfyanonymous committed
442
443
444
445

        text = escape_important(text)
        parsed_weights = token_weights(text, 1.0)

446
        #tokenize words
comfyanonymous's avatar
comfyanonymous committed
447
        tokens = []
448
449
450
451
452
453
        for weighted_segment, weight in parsed_weights:
            to_tokenize = unescape_important(weighted_segment).replace("\n", " ").split(' ')
            to_tokenize = [x for x in to_tokenize if x != ""]
            for word in to_tokenize:
                #if we find an embedding, deal with the embedding
                if word.startswith(self.embedding_identifier) and self.embedding_directory is not None:
454
455
                    embedding_name = word[len(self.embedding_identifier):].strip('\n')
                    embed, leftover = self._try_get_embedding(embedding_name)
456
                    if embed is None:
457
                        logging.warning(f"warning, embedding:{embedding_name} does not exist, ignoring")
458
                    else:
459
                        if len(embed.shape) == 1:
460
                            tokens.append([(embed, weight)])
461
                        else:
462
463
464
465
                            tokens.append([(embed[x], weight) for x in range(embed.shape[0])])
                    #if we accidentally have leftover text, continue parsing using leftover, else move on to next word
                    if leftover != "":
                        word = leftover
466
                    else:
467
468
                        continue
                #parse word
469
                tokens.append([(t, weight) for t in self.tokenizer(word)["input_ids"][self.tokens_start:-1]])
470

471
472
        #reshape token array to CLIP input size
        batched_tokens = []
473
474
475
        batch = []
        if self.start_token is not None:
            batch.append((self.start_token, 1.0, 0))
476
477
        batched_tokens.append(batch)
        for i, t_group in enumerate(tokens):
478
479
            #determine if we're going to try and keep the tokens in a single batch
            is_large = len(t_group) >= self.max_word_length
BlenderNeko's avatar
BlenderNeko committed
480

481
            while len(t_group) > 0:
BlenderNeko's avatar
BlenderNeko committed
482
483
484
                if len(t_group) + len(batch) > self.max_length - 1:
                    remaining_length = self.max_length - len(batch) - 1
                    #break word in two and add end token
485
486
                    if is_large:
                        batch.extend([(t,w,i+1) for t,w in t_group[:remaining_length]])
BlenderNeko's avatar
BlenderNeko committed
487
                        batch.append((self.end_token, 1.0, 0))
488
                        t_group = t_group[remaining_length:]
BlenderNeko's avatar
BlenderNeko committed
489
                    #add end token and pad
490
                    else:
BlenderNeko's avatar
BlenderNeko committed
491
                        batch.append((self.end_token, 1.0, 0))
492
                        if self.pad_to_max_length:
493
                            batch.extend([(self.pad_token, 1.0, 0)] * (remaining_length))
BlenderNeko's avatar
BlenderNeko committed
494
                    #start new batch
495
496
497
                    batch = []
                    if self.start_token is not None:
                        batch.append((self.start_token, 1.0, 0))
498
                    batched_tokens.append(batch)
499
                else:
500
501
                    batch.extend([(t,w,i+1) for t,w in t_group])
                    t_group = []
502

503
        #fill last batch
504
505
        batch.append((self.end_token, 1.0, 0))
        if self.pad_to_max_length:
506
            batch.extend([(self.pad_token, 1.0, 0)] * (self.max_length - len(batch)))
507
        if self.min_length is not None and len(batch) < self.min_length:
508
            batch.extend([(self.pad_token, 1.0, 0)] * (self.min_length - len(batch)))
comfyanonymous's avatar
comfyanonymous committed
509

510
511
        if not return_word_ids:
            batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens]
comfyanonymous's avatar
comfyanonymous committed
512

513
        return batched_tokens
comfyanonymous's avatar
comfyanonymous committed
514
515
516
517


    def untokenize(self, token_weight_pair):
        return list(map(lambda a: (a, self.inv_vocab[a[0]]), token_weight_pair))
518

519
520
    def state_dict(self):
        return {}
521
522

class SD1Tokenizer:
523
    def __init__(self, embedding_directory=None, tokenizer_data={}, clip_name="l", tokenizer=SDTokenizer):
524
525
        self.clip_name = clip_name
        self.clip = "clip_{}".format(self.clip_name)
526
        setattr(self, self.clip, tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data))
527
528
529
530
531
532
533
534
535

    def tokenize_with_weights(self, text:str, return_word_ids=False):
        out = {}
        out[self.clip_name] = getattr(self, self.clip).tokenize_with_weights(text, return_word_ids)
        return out

    def untokenize(self, token_weight_pair):
        return getattr(self, self.clip).untokenize(token_weight_pair)

536
537
    def state_dict(self):
        return {}
538
539

class SD1ClipModel(torch.nn.Module):
540
    def __init__(self, device="cpu", dtype=None, clip_name="l", clip_model=SDClipModel, name=None, **kwargs):
541
        super().__init__()
542
543
544
545
546
547
548
549

        if name is not None:
            self.clip_name = name
            self.clip = "{}".format(self.clip_name)
        else:
            self.clip_name = clip_name
            self.clip = "clip_{}".format(self.clip_name)

550
        setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs))
551

552
553
554
555
        self.dtypes = set()
        if dtype is not None:
            self.dtypes.add(dtype)

556
557
    def set_clip_options(self, options):
        getattr(self, self.clip).set_clip_options(options)
558

559
560
    def reset_clip_options(self):
        getattr(self, self.clip).reset_clip_options()
561
562
563

    def encode_token_weights(self, token_weight_pairs):
        token_weight_pairs = token_weight_pairs[self.clip_name]
564
565
        out = getattr(self, self.clip).encode_token_weights(token_weight_pairs)
        return out
566
567
568

    def load_sd(self, sd):
        return getattr(self, self.clip).load_sd(sd)