File size: 14,447 Bytes
5d3f98b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2021-2022 The Alibaba Fundamental Vision Team Authors. All rights reserved.
import os.path as osp
from typing import Any, Dict

import json
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from modelscope.metainfo import Models
from modelscope.models import Model
from modelscope.models.builder import MODELS
from modelscope.models.multi_modal.diffusion.diffusion import (
    GaussianDiffusion, beta_schedule)
from modelscope.models.multi_modal.diffusion.structbert import (BertConfig,
                                                                BertModel)
from modelscope.models.multi_modal.diffusion.tokenizer import FullTokenizer
from modelscope.models.multi_modal.diffusion.unet_generator import \
    DiffusionGenerator
from modelscope.models.multi_modal.diffusion.unet_upsampler_256 import \
    SuperResUNet256
from modelscope.models.multi_modal.diffusion.unet_upsampler_1024 import \
    SuperResUNet1024
from modelscope.utils.constant import ModelFile, Tasks
from modelscope.utils.device import create_device
from modelscope.utils.logger import get_logger

logger = get_logger()

__all__ = ['DiffusionForTextToImageSynthesis']


def make_diffusion(schedule,
                   num_timesteps=1000,
                   init_beta=None,
                   last_beta=None,
                   var_type='fixed_small'):
    betas = beta_schedule(schedule, num_timesteps, init_beta, last_beta)
    diffusion = GaussianDiffusion(betas, var_type=var_type)
    return diffusion


class Tokenizer(object):

    def __init__(self, vocab_file, seq_len=64):
        self.vocab_file = vocab_file
        self.seq_len = seq_len
        self.tokenizer = FullTokenizer(
            vocab_file=vocab_file, do_lower_case=True)

    def __call__(self, text):
        # tokenization
        tokens = self.tokenizer.tokenize(text)
        tokens = ['[CLS]'] + tokens[:self.seq_len - 2] + ['[SEP]']
        input_ids = self.tokenizer.convert_tokens_to_ids(tokens)
        input_mask = [1] * len(input_ids)
        segment_ids = [0] * len(input_ids)

        # padding
        input_ids += [0] * (self.seq_len - len(input_ids))
        input_mask += [0] * (self.seq_len - len(input_mask))
        segment_ids += [0] * (self.seq_len - len(segment_ids))
        assert len(input_ids) == len(input_mask) == len(
            segment_ids) == self.seq_len

        # convert to tensors
        input_ids = torch.LongTensor(input_ids)
        input_mask = torch.LongTensor(input_mask)
        segment_ids = torch.LongTensor(segment_ids)
        return input_ids, segment_ids, input_mask


class DiffusionModel(nn.Module):

    def __init__(self, model_dir):
        super(DiffusionModel, self).__init__()
        # including text and generator config
        model_config = json.load(
            open('{}/model_config.json'.format(model_dir), encoding='utf-8'))

        # text encoder
        text_config = model_config['text_config']
        self.text_encoder = BertModel(BertConfig.from_dict(text_config))

        # generator (64x64)
        generator_config = model_config['generator_config']
        self.unet_generator = DiffusionGenerator(**generator_config)

        # upsampler (256x256)
        upsampler_256_config = model_config['upsampler_256_config']
        self.unet_upsampler_256 = SuperResUNet256(**upsampler_256_config)

        # upsampler (1024x1024)
        upsampler_1024_config = model_config['upsampler_1024_config']
        self.unet_upsampler_1024 = SuperResUNet1024(**upsampler_1024_config)

    def forward(self, noise, timesteps, input_ids, token_type_ids,
                attention_mask):
        context, y = self.text_encoder(
            input_ids=input_ids,
            token_type_ids=token_type_ids,
            attention_mask=attention_mask)
        context = context[-1]
        x = self.unet_generator(noise, timesteps, y, context, attention_mask)
        x = self.unet_upsampler_256(noise, timesteps, x,
                                    torch.zeros_like(timesteps), y, context,
                                    attention_mask)
        x = self.unet_upsampler_1024(x, t, x)
        return x


@MODELS.register_module(
    Tasks.text_to_image_synthesis, module_name=Models.diffusion)
class DiffusionForTextToImageSynthesis(Model):

    def __init__(self, model_dir, device='gpu', **kwargs):
        device = 'gpu' if torch.cuda.is_available() else 'cpu'
        super().__init__(model_dir=model_dir, device=device, **kwargs)
        diffusion_model = DiffusionModel(model_dir=model_dir)
        pretrained_params = torch.load(
            osp.join(model_dir, ModelFile.TORCH_MODEL_BIN_FILE), 'cpu')
        diffusion_model.load_state_dict(pretrained_params)
        diffusion_model.eval().to()

        self.device = create_device(device)
        diffusion_model.to(self.device)

        # modules
        self.text_encoder = diffusion_model.text_encoder
        self.unet_generator = diffusion_model.unet_generator
        self.unet_upsampler_256 = diffusion_model.unet_upsampler_256
        self.unet_upsampler_1024 = diffusion_model.unet_upsampler_1024

        # text tokenizer
        vocab_path = f'{model_dir}/{ModelFile.VOCAB_FILE}'
        self.tokenizer = Tokenizer(vocab_file=vocab_path, seq_len=64)

        # diffusion process
        diffusion_params = json.load(
            open(
                '{}/diffusion_config.json'.format(model_dir),
                encoding='utf-8'))
        self.diffusion_generator = make_diffusion(
            **diffusion_params['generator_config'])
        self.diffusion_upsampler_256 = make_diffusion(
            **diffusion_params['upsampler_256_config'])
        self.diffusion_upsampler_1024 = make_diffusion(
            **diffusion_params['upsampler_1024_config'])

    def forward(self, input: Dict[str, Any]) -> Dict[str, Any]:
        if not all([key in input for key in ('text', 'noise', 'timesteps')]):
            raise ValueError(
                f'input should contains "text", "noise", and "timesteps", but got {input.keys()}'
            )
        input_ids, token_type_ids, attention_mask = self.tokenizer(
            input['text'])
        input_ids = input_ids.to(self.device).unsqueeze(0)
        token_type_ids = token_type_ids.to(self.device).unsqueeze(0)
        attention_mask = attention_mask.to(self.device).unsqueeze(0)
        context, y = self.text_encoder(
            input_ids=input_ids,
            token_type_ids=token_type_ids,
            attention_mask=attention_mask)
        context = context[-1]
        x = self.unet_generator(noise, timesteps, y, context, attention_mask)
        x = self.unet_upsampler_256(noise, timesteps, x,
                                    torch.zeros_like(timesteps), y, context,
                                    attention_mask)
        x = self.unet_upsampler_1024(x, t, x)
        img = x.clamp(-1, 1).add(1).mul(127.5)
        img = img.squeeze(0).permute(1, 2, 0).cpu().numpy().astype(np.uint8)
        return img

    def postprocess(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
        return inputs

    @torch.no_grad()
    def generate(self, input: Dict[str, Any]) -> Dict[str, Any]:
        if 'text' not in input:
            raise ValueError(
                f'input should contain "text", but got {input.keys()}')

        # encode text
        input_ids, token_type_ids, attention_mask = self.tokenizer(
            input['text'])
        input_ids = input_ids.to(self.device).unsqueeze(0)
        token_type_ids = token_type_ids.to(self.device).unsqueeze(0)
        attention_mask = attention_mask.to(self.device).unsqueeze(0)
        context, y = self.text_encoder(
            input_ids=input_ids,
            token_type_ids=token_type_ids,
            attention_mask=attention_mask)
        context = context[-1]

        # choose a proper solver
        solver = input.get('solver', 'dpm-solver')
        if solver == 'dpm-solver':
            # generation
            img = self.diffusion_generator.dpm_solver_sample_loop(
                noise=torch.randn(1, 3, 64, 64).to(self.device),
                model=self.unet_generator,
                model_kwargs=[{
                    'y': y,
                    'context': context,
                    'mask': attention_mask
                }, {
                    'y': torch.zeros_like(y),
                    'context': torch.zeros_like(context),
                    'mask': attention_mask
                }],
                percentile=input.get('generator_percentile', 0.995),
                guide_scale=input.get('generator_guide_scale', 5.0),
                dpm_solver_timesteps=input.get('dpm_solver_timesteps', 20),
                order=3,
                skip_type='logSNR',
                method='singlestep',
                t_start=0.9946)

            # upsampling (64->256)
            if not input.get('debug', False):
                img = F.interpolate(
                    img,
                    scale_factor=4.0,
                    mode='bilinear',
                    align_corners=False)
            img = self.diffusion_upsampler_256.dpm_solver_sample_loop(
                noise=torch.randn_like(img),
                model=self.unet_upsampler_256,
                model_kwargs=[{
                    'lx': img,
                    'lt': torch.zeros(1).to(self.device),
                    'y': y,
                    'context': context,
                    'mask': attention_mask
                }, {
                    'lx': img,
                    'lt': torch.zeros(1).to(self.device),
                    'y': torch.zeros_like(y),
                    'context': torch.zeros_like(context),
                    'mask': torch.zeros_like(attention_mask)
                }],
                percentile=input.get('upsampler_256_percentile', 0.995),
                guide_scale=input.get('upsampler_256_guide_scale', 5.0),
                dpm_solver_timesteps=input.get('dpm_solver_timesteps', 20),
                order=3,
                skip_type='logSNR',
                method='singlestep',
                t_start=0.9946)

            # upsampling (256->1024)
            if not input.get('debug', False):
                img = F.interpolate(
                    img,
                    scale_factor=4.0,
                    mode='bilinear',
                    align_corners=False)
            img = self.diffusion_upsampler_1024.dpm_solver_sample_loop(
                noise=torch.randn_like(img),
                model=self.unet_upsampler_256,
                model_kwargs=[{
                    'lx': img,
                    'lt': torch.zeros(1).to(self.device),
                    'y': y,
                    'context': context,
                    'mask': attention_mask
                }, {
                    'lx': img,
                    'lt': torch.zeros(1).to(self.device),
                    'y': torch.zeros_like(y),
                    'context': torch.zeros_like(context),
                    'mask': torch.zeros_like(attention_mask)
                }],
                percentile=input.get('upsampler_256_percentile', 0.995),
                guide_scale=input.get('upsampler_256_guide_scale', 5.0),
                dpm_solver_timesteps=input.get('dpm_solver_timesteps', 10),
                order=3,
                skip_type='logSNR',
                method='singlestep',
                t_start=None)
        elif solver == 'ddim':
            # generation
            img = self.diffusion_generator.ddim_sample_loop(
                noise=torch.randn(1, 3, 64, 64).to(self.device),
                model=self.unet_generator,
                model_kwargs=[{
                    'y': y,
                    'context': context,
                    'mask': attention_mask
                }, {
                    'y': torch.zeros_like(y),
                    'context': torch.zeros_like(context),
                    'mask': attention_mask
                }],
                percentile=input.get('generator_percentile', 0.995),
                guide_scale=input.get('generator_guide_scale', 5.0),
                ddim_timesteps=input.get('generator_ddim_timesteps', 250),
                eta=input.get('generator_ddim_eta', 0.0))

            # upsampling (64->256)
            if not input.get('debug', False):
                img = F.interpolate(
                    img,
                    scale_factor=4.0,
                    mode='bilinear',
                    align_corners=False)
            img = self.diffusion_upsampler_256.ddim_sample_loop(
                noise=torch.randn_like(img),
                model=self.unet_upsampler_256,
                model_kwargs=[{
                    'lx': img,
                    'lt': torch.zeros(1).to(self.device),
                    'y': y,
                    'context': context,
                    'mask': attention_mask
                }, {
                    'lx': img,
                    'lt': torch.zeros(1).to(self.device),
                    'y': torch.zeros_like(y),
                    'context': torch.zeros_like(context),
                    'mask': torch.zeros_like(attention_mask)
                }],
                percentile=input.get('upsampler_256_percentile', 0.995),
                guide_scale=input.get('upsampler_256_guide_scale', 5.0),
                ddim_timesteps=input.get('upsampler_256_ddim_timesteps', 50),
                eta=input.get('upsampler_256_ddim_eta', 0.0))

            # upsampling (256->1024)
            if not input.get('debug', False):
                img = F.interpolate(
                    img,
                    scale_factor=4.0,
                    mode='bilinear',
                    align_corners=False)
            img = self.diffusion_upsampler_1024.ddim_sample_loop(
                noise=torch.randn_like(img),
                model=self.unet_upsampler_1024,
                model_kwargs={'concat': img},
                percentile=input.get('upsampler_1024_percentile', 0.995),
                ddim_timesteps=input.get('upsampler_1024_ddim_timesteps', 20),
                eta=input.get('upsampler_1024_ddim_eta', 0.0))
        else:
            raise ValueError(
                'currently only supports "ddim" and "dpm-solve" solvers')

        # output
        img = img.clamp(-1, 1).add(1).mul(127.5).squeeze(0).permute(
            1, 2, 0).cpu().numpy().astype(np.uint8)
        return img