sd1_clip.py 15.7 KB
Newer Older
comfyanonymous's avatar
comfyanonymous committed
1
2
import os

3
from transformers import CLIPTokenizer, CLIPTextModel, CLIPTextConfig, modeling_utils
4
import comfy.ops
comfyanonymous's avatar
comfyanonymous committed
5
import torch
6
import traceback
7
import zipfile
8
9
from . import model_management
import contextlib
comfyanonymous's avatar
comfyanonymous committed
10
11
12

class ClipTokenWeightEncoder:
    def encode_token_weights(self, token_weight_pairs):
13
        to_encode = list(self.empty_tokens)
comfyanonymous's avatar
comfyanonymous committed
14
        for x in token_weight_pairs:
15
16
17
18
19
20
21
22
23
24
25
            tokens = list(map(lambda a: a[0], x))
            to_encode.append(tokens)

        out, pooled = self.encode(to_encode)
        z_empty = out[0:1]
        if pooled.shape[0] > 1:
            first_pooled = pooled[1:2]
        else:
            first_pooled = pooled[0:1]

        output = []
26
27
        for k in range(1, out.shape[0]):
            z = out[k:k+1]
comfyanonymous's avatar
comfyanonymous committed
28
29
            for i in range(len(z)):
                for j in range(len(z[i])):
30
                    weight = token_weight_pairs[k - 1][j][1]
comfyanonymous's avatar
comfyanonymous committed
31
                    z[i][j] = (z[i][j] - z_empty[0][j]) * weight + z_empty[0][j]
32
33
            output.append(z)

comfyanonymous's avatar
comfyanonymous committed
34
        if (len(output) == 0):
35
            return z_empty, first_pooled
36
        return torch.cat(output, dim=-2).cpu(), first_pooled.cpu()
comfyanonymous's avatar
comfyanonymous committed
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

class SD1ClipModel(torch.nn.Module, ClipTokenWeightEncoder):
    """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,
                 freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, textmodel_path=None):  # clip-vit-base-patch32
        super().__init__()
        assert layer in self.LAYERS
        if textmodel_path is not None:
            self.transformer = CLIPTextModel.from_pretrained(textmodel_path)
        else:
            if textmodel_json_config is None:
                textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json")
            config = CLIPTextConfig.from_json_file(textmodel_json_config)
55
56
57
            with comfy.ops.use_comfy_ops():
                with modeling_utils.no_init_weights():
                    self.transformer = CLIPTextModel(config)
comfyanonymous's avatar
comfyanonymous committed
58
59
60
61
62
63
64

        self.max_length = max_length
        if freeze:
            self.freeze()
        self.layer = layer
        self.layer_idx = None
        self.empty_tokens = [[49406] + [49407] * 76]
65
66
        self.text_projection = None
        self.layer_norm_hidden_state = True
comfyanonymous's avatar
comfyanonymous committed
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
        if layer == "hidden":
            assert layer_idx is not None
            assert abs(layer_idx) <= 12
            self.clip_layer(layer_idx)

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

    def clip_layer(self, layer_idx):
        if abs(layer_idx) >= 12:
            self.layer = "last"
        else:
            self.layer = "hidden"
            self.layer_idx = layer_idx

85
86
87
88
89
90
91
92
93
94
95
    def set_up_textual_embeddings(self, tokens, current_embeds):
        out_tokens = []
        next_new_token = token_dict_size = current_embeds.weight.shape[0]
        embedding_weights = []

        for x in tokens:
            tokens_temp = []
            for y in x:
                if isinstance(y, int):
                    tokens_temp += [y]
                else:
96
97
98
99
100
101
                    if y.shape[0] == current_embeds.weight.shape[1]:
                        embedding_weights += [y]
                        tokens_temp += [next_new_token]
                        next_new_token += 1
                    else:
                        print("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored", y.shape[0], current_embeds.weight.shape[1])
102
103
            while len(tokens_temp) < len(x):
                tokens_temp += [self.empty_tokens[0][-1]]
104
105
106
            out_tokens += [tokens_temp]

        if len(embedding_weights) > 0:
107
            new_embedding = torch.nn.Embedding(next_new_token, current_embeds.weight.shape[1], device=current_embeds.weight.device, dtype=current_embeds.weight.dtype)
108
109
110
111
112
113
114
115
            new_embedding.weight[:token_dict_size] = current_embeds.weight[:]
            n = token_dict_size
            for x in embedding_weights:
                new_embedding.weight[n] = x
                n += 1
            self.transformer.set_input_embeddings(new_embedding)
        return out_tokens

comfyanonymous's avatar
comfyanonymous committed
116
    def forward(self, tokens):
117
        backup_embeds = self.transformer.get_input_embeddings()
118
        device = backup_embeds.weight.device
119
        tokens = self.set_up_textual_embeddings(tokens, backup_embeds)
120
121
122
123
        tokens = torch.LongTensor(tokens).to(device)

        if backup_embeds.weight.dtype != torch.float32:
            precision_scope = torch.autocast
comfyanonymous's avatar
comfyanonymous committed
124
        else:
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
            precision_scope = contextlib.nullcontext

        with precision_scope(model_management.get_autocast_device(device)):
            outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer=="hidden")
            self.transformer.set_input_embeddings(backup_embeds)

            if self.layer == "last":
                z = outputs.last_hidden_state
            elif self.layer == "pooled":
                z = outputs.pooler_output[:, None, :]
            else:
                z = outputs.hidden_states[self.layer_idx]
                if self.layer_norm_hidden_state:
                    z = self.transformer.text_model.final_layer_norm(z)

            pooled_output = outputs.pooler_output
            if self.text_projection is not None:
                pooled_output = pooled_output @ self.text_projection
        return z.float(), pooled_output.float()
comfyanonymous's avatar
comfyanonymous committed
144
145
146
147

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

148
149
150
    def load_sd(self, sd):
        return self.transformer.load_state_dict(sd, strict=False)

comfyanonymous's avatar
comfyanonymous committed
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
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

209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
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

228
229
230
231
232
233
234
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)
235

236
def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=None):
237
238
239
    if isinstance(embedding_directory, str):
        embedding_directory = [embedding_directory]

240
241
    embedding_directory = expand_directory_list(embedding_directory)

242
243
244
245
246
247
248
249
250
251
    valid_file = None
    for embed_dir in embedding_directory:
        embed_path = os.path.join(embed_dir, embedding_name)
        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
252
        else:
253
254
255
256
257
258
259
260
            valid_file = embed_path
        if valid_file is not None:
            break

    if valid_file is None:
        return None

    embed_path = valid_file
261

262
263
    embed_out = None

264
265
266
267
    try:
        if embed_path.lower().endswith(".safetensors"):
            import safetensors.torch
            embed = safetensors.torch.load_file(embed_path, device="cpu")
comfyanonymous's avatar
comfyanonymous committed
268
        else:
269
            if 'weights_only' in torch.load.__code__.co_varnames:
270
271
272
273
                try:
                    embed = torch.load(embed_path, weights_only=True, map_location="cpu")
                except:
                    embed_out = safe_load_embed_zip(embed_path)
274
275
276
277
278
279
280
281
            else:
                embed = torch.load(embed_path, map_location="cpu")
    except Exception as e:
        print(traceback.format_exc())
        print()
        print("error loading embedding, skipping loading:", embedding_name)
        return None

282
283
284
    if embed_out is None:
        if 'string_to_param' in embed:
            values = embed['string_to_param'].values()
285
286
287
288
289
290
291
292
293
294
            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)
295
296
        elif embed_key is not None and embed_key in embed:
            embed_out = embed[embed_key]
297
298
        else:
            values = embed.values()
299
            embed_out = next(iter(values))
300
    return embed_out
301

comfyanonymous's avatar
comfyanonymous committed
302
class SD1Tokenizer:
303
    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l'):
comfyanonymous's avatar
comfyanonymous committed
304
305
306
307
        if tokenizer_path is None:
            tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
        self.tokenizer = CLIPTokenizer.from_pretrained(tokenizer_path)
        self.max_length = max_length
308
309
        self.max_tokens_per_section = self.max_length - 2

comfyanonymous's avatar
comfyanonymous committed
310
311
312
313
314
315
        empty = self.tokenizer('')["input_ids"]
        self.start_token = empty[0]
        self.end_token = empty[1]
        self.pad_with_end = pad_with_end
        vocab = self.tokenizer.get_vocab()
        self.inv_vocab = {v: k for k, v in vocab.items()}
316
317
        self.embedding_directory = embedding_directory
        self.max_word_length = 8
318
        self.embedding_identifier = "embedding:"
319
        self.embedding_size = embedding_size
320
        self.embedding_key = embedding_key
321

322
    def _try_get_embedding(self, embedding_name:str):
323
324
325
326
        '''
        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.
        '''
327
        embed = load_embed(embedding_name, self.embedding_directory, self.embedding_size, self.embedding_key)
328
329
330
        if embed is None:
            stripped = embedding_name.strip(',')
            if len(stripped) < len(embedding_name):
331
                embed = load_embed(stripped, self.embedding_directory, self.embedding_size, self.embedding_key)
332
333
334
335
                return (embed, embedding_name[len(stripped):])
        return (embed, "")


336
    def tokenize_with_weights(self, text:str, return_word_ids=False):
337
338
339
340
341
342
        '''
        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
        '''
BlenderNeko's avatar
BlenderNeko committed
343
344
345
346
        if self.pad_with_end:
            pad_token = self.end_token
        else:
            pad_token = 0
comfyanonymous's avatar
comfyanonymous committed
347
348
349
350

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

351
        #tokenize words
comfyanonymous's avatar
comfyanonymous committed
352
        tokens = []
353
354
355
356
357
358
        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:
359
360
                    embedding_name = word[len(self.embedding_identifier):].strip('\n')
                    embed, leftover = self._try_get_embedding(embedding_name)
361
                    if embed is None:
362
                        print(f"warning, embedding:{embedding_name} does not exist, ignoring")
363
                    else:
364
                        if len(embed.shape) == 1:
365
                            tokens.append([(embed, weight)])
366
                        else:
367
368
369
370
                            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
371
                    else:
372
373
374
                        continue
                #parse word
                tokens.append([(t, weight) for t in self.tokenizer(word)["input_ids"][1:-1]])
375

376
377
        #reshape token array to CLIP input size
        batched_tokens = []
BlenderNeko's avatar
BlenderNeko committed
378
        batch = [(self.start_token, 1.0, 0)]
379
380
        batched_tokens.append(batch)
        for i, t_group in enumerate(tokens):
381
382
            #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
383

384
            while len(t_group) > 0:
BlenderNeko's avatar
BlenderNeko committed
385
386
387
                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
388
389
                    if is_large:
                        batch.extend([(t,w,i+1) for t,w in t_group[:remaining_length]])
BlenderNeko's avatar
BlenderNeko committed
390
                        batch.append((self.end_token, 1.0, 0))
391
                        t_group = t_group[remaining_length:]
BlenderNeko's avatar
BlenderNeko committed
392
                    #add end token and pad
393
                    else:
BlenderNeko's avatar
BlenderNeko committed
394
395
396
397
                        batch.append((self.end_token, 1.0, 0))
                        batch.extend([(pad_token, 1.0, 0)] * (remaining_length))
                    #start new batch
                    batch = [(self.start_token, 1.0, 0)]
398
                    batched_tokens.append(batch)
399
                else:
400
401
                    batch.extend([(t,w,i+1) for t,w in t_group])
                    t_group = []
402

403
        #fill last batch
BlenderNeko's avatar
BlenderNeko committed
404
        batch.extend([(self.end_token, 1.0, 0)] + [(pad_token, 1.0, 0)] * (self.max_length - len(batch) - 1))
comfyanonymous's avatar
comfyanonymous committed
405

406
407
        if not return_word_ids:
            batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens]
comfyanonymous's avatar
comfyanonymous committed
408

409
        return batched_tokens
comfyanonymous's avatar
comfyanonymous committed
410
411
412
413


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