File size: 12,508 Bytes
fb53ec8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import numpy as np
from tqdm import tqdm
from functools import partial

from .diffusion_utils import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like

from .ddim import DDIMSampler

class DDIMSampler_VD(DDIMSampler):
    @torch.no_grad()
    def sample(self,
               steps,
               shape,
               xt=None,
               conditioning=None,
               unconditional_guidance_scale=1.,
               unconditional_conditioning=None,
               xtype='image',
               ctype='prompt',
               eta=0.,
               temperature=1.,
               noise_dropout=0.,
               verbose=True,
               log_every_t=100,):

        self.make_schedule(ddim_num_steps=steps, ddim_eta=eta, verbose=verbose)
        print(f'Data shape for DDIM sampling is {shape}, eta {eta}')
        samples, intermediates = self.ddim_sampling(
            shape,
            xt=xt,
            conditioning=conditioning, 
            unconditional_guidance_scale=unconditional_guidance_scale,
            unconditional_conditioning=unconditional_conditioning,
            xtype=xtype,
            ctype=ctype,
            ddim_use_original_steps=False,
            noise_dropout=noise_dropout,
            temperature=temperature,
            log_every_t=log_every_t,)
        return samples, intermediates

    @torch.no_grad()
    def ddim_sampling(self, 
                      shape,
                      xt=None,
                      conditioning=None,
                      unconditional_guidance_scale=1., 
                      unconditional_conditioning=None,
                      xtype='image',
                      ctype='prompt',
                      ddim_use_original_steps=False,
                      timesteps=None, 
                      noise_dropout=0., 
                      temperature=1., 
                      log_every_t=100,):

        device = self.model.device
        bs = shape[0]
        if xt is None:
            xt = torch.randn(shape, device=device)

        if timesteps is None:
            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
        elif timesteps is not None and not ddim_use_original_steps:
            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
            timesteps = self.ddim_timesteps[:subset_end]

        intermediates = {'pred_xt': [], 'pred_x0': []}
        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
        # print(f"Running DDIM Sampling with {total_steps} timesteps")

        pred_xt = xt
        iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
        for i, step in enumerate(iterator):
            index = total_steps - i - 1
            ts = torch.full((bs,), step, device=device, dtype=torch.long)

            outs = self.p_sample_ddim(
                pred_xt, conditioning, ts, index, 
                unconditional_guidance_scale=unconditional_guidance_scale,
                unconditional_conditioning=unconditional_conditioning, 
                xtype=xtype,
                ctype=ctype,
                use_original_steps=ddim_use_original_steps,
                noise_dropout=noise_dropout,
                temperature=temperature,)
            pred_xt, pred_x0 = outs

            if index % log_every_t == 0 or index == total_steps - 1:
                intermediates['pred_xt'].append(pred_xt)
                intermediates['pred_x0'].append(pred_x0)

        return pred_xt, intermediates

    @torch.no_grad()
    def p_sample_ddim(self, x, conditioning, t, index, 
                      unconditional_guidance_scale=1., 
                      unconditional_conditioning=None, 
                      xtype='image',
                      ctype='prompt',
                      repeat_noise=False, 
                      use_original_steps=False, 
                      noise_dropout=0.,
                      temperature=1.,):

        b, *_, device = *x.shape, x.device

        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
            e_t = self.model.apply_model(x, t, conditioning, xtype=xtype, ctype=ctype)
        else:
            x_in = torch.cat([x] * 2)
            t_in = torch.cat([t] * 2)
            c_in = torch.cat([unconditional_conditioning, conditioning])
            e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in, xtype=xtype, ctype=ctype).chunk(2)
            e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)

        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
        # select parameters corresponding to the currently considered timestep

        if xtype == 'image':
            extended_shape = (b, 1, 1, 1)
        elif xtype == 'text':
            extended_shape = (b, 1)

        a_t = torch.full(extended_shape, alphas[index], device=device)
        a_prev = torch.full(extended_shape, alphas_prev[index], device=device)
        sigma_t = torch.full(extended_shape, sigmas[index], device=device)
        sqrt_one_minus_at = torch.full(extended_shape, sqrt_one_minus_alphas[index],device=device)

        # current prediction for x_0
        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
        if noise_dropout > 0.:
            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
        return x_prev, pred_x0

class DDIMSampler_VD_DualContext(DDIMSampler_VD):
    @torch.no_grad()
    def sample_dc(self,
               steps,
               shape,
               xt=None,
               first_conditioning=None,
               second_conditioning=None,
               unconditional_guidance_scale=1.,
               xtype='image',
               first_ctype='prompt',
               second_ctype='prompt',
               eta=0.,
               temperature=1.,
               mixed_ratio=0.5,
               noise_dropout=0.,
               verbose=True,
               log_every_t=100,):

        self.make_schedule(ddim_num_steps=steps, ddim_eta=eta, verbose=verbose)
        print(f'Data shape for DDIM sampling is {shape}, eta {eta}')
        samples, intermediates = self.ddim_sampling_dc(
            shape,
            xt=xt,
            first_conditioning=first_conditioning,
            second_conditioning=second_conditioning,
            unconditional_guidance_scale=unconditional_guidance_scale,
            xtype=xtype,
            first_ctype=first_ctype,
            second_ctype=second_ctype,
            ddim_use_original_steps=False,
            noise_dropout=noise_dropout,
            temperature=temperature,
            log_every_t=log_every_t,
            mixed_ratio=mixed_ratio, )
        return samples, intermediates

    @torch.no_grad()
    def ddim_sampling_dc(self, 
                      shape,
                      xt=None,
                      first_conditioning=None,
                      second_conditioning=None,
                      unconditional_guidance_scale=1., 
                      xtype='image',
                      first_ctype='prompt',
                      second_ctype='prompt',
                      ddim_use_original_steps=False,
                      timesteps=None, 
                      noise_dropout=0., 
                      temperature=1.,
                      mixed_ratio=0.5,
                      log_every_t=100,):

        device = self.model.device
        bs = shape[0]
        if xt is None:
            xt = torch.randn(shape, device=device)

        if timesteps is None:
            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
        elif timesteps is not None and not ddim_use_original_steps:
            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
            timesteps = self.ddim_timesteps[:subset_end]

        intermediates = {'pred_xt': [], 'pred_x0': []}
        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
        # print(f"Running DDIM Sampling with {total_steps} timesteps")

        pred_xt = xt
        iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
        for i, step in enumerate(iterator):
            index = total_steps - i - 1
            ts = torch.full((bs,), step, device=device, dtype=torch.long)

            outs = self.p_sample_ddim_dc(
                pred_xt, 
                first_conditioning, 
                second_conditioning, 
                ts, index, 
                unconditional_guidance_scale=unconditional_guidance_scale,
                xtype=xtype,
                first_ctype=first_ctype,
                second_ctype=second_ctype,
                use_original_steps=ddim_use_original_steps,
                noise_dropout=noise_dropout,
                temperature=temperature,
                mixed_ratio=mixed_ratio,)
            pred_xt, pred_x0 = outs

            if index % log_every_t == 0 or index == total_steps - 1:
                intermediates['pred_xt'].append(pred_xt)
                intermediates['pred_x0'].append(pred_x0)

        return pred_xt, intermediates

    @torch.no_grad()
    def p_sample_ddim_dc(self, x, 
                      first_conditioning,
                      second_conditioning,
                      t, index, 
                      unconditional_guidance_scale=1., 
                      xtype='image',
                      first_ctype='prompt',
                      second_ctype='prompt',
                      repeat_noise=False, 
                      use_original_steps=False, 
                      noise_dropout=0.,
                      temperature=1.,
                      mixed_ratio=0.5,):

        b, *_, device = *x.shape, x.device

        x_in = torch.cat([x] * 2)
        t_in = torch.cat([t] * 2)
        first_c = torch.cat(first_conditioning)
        second_c = torch.cat(second_conditioning)

        e_t_uncond, e_t = self.model.apply_model_dc(
            x_in, t_in, first_c, second_c, xtype=xtype, first_ctype=first_ctype, second_ctype=second_ctype, mixed_ratio=mixed_ratio).chunk(2)

        # e_t_uncond, e_t = self.model.apply_model(x_in, t_in, first_c, xtype='image', ctype='vision').chunk(2)

        e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)

        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
        # select parameters corresponding to the currently considered timestep

        if xtype == 'image':
            extended_shape = (b, 1, 1, 1)
        elif xtype == 'text':
            extended_shape = (b, 1)

        a_t = torch.full(extended_shape, alphas[index], device=device)
        a_prev = torch.full(extended_shape, alphas_prev[index], device=device)
        sigma_t = torch.full(extended_shape, sigmas[index], device=device)
        sqrt_one_minus_at = torch.full(extended_shape, sqrt_one_minus_alphas[index],device=device)

        # current prediction for x_0
        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
        if noise_dropout > 0.:
            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
        return x_prev, pred_x0