File size: 8,263 Bytes
3c19795
990d40a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3c19795
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import List, Optional
import torch 
import torch.nn as nn
import torch.nn.functional as F
import numpy as np


def extract(v, t, x_shape):
    """
    Extract some coefficients at specified timesteps, then reshape to
    [batch_size, 1, 1, 1, 1, ...] for broadcasting purposes.
    """
    v = v.to("cuda")
    out = torch.gather(v, index=t, dim=0).float().to("cuda")
    return out.view([t.shape[0]] + [1] * (len(x_shape) - 1))


class GaussianDiffusionTrainer(nn.Module):
    def __init__(self, beta_1, beta_T, T, model) -> None:
        super().__init__()
        self.model = model
        self.register_buffer(
            'betas',
            torch.linspace(beta_1,beta_T,T).double()
        )

        self.T = T
        self.alphas = 1 - self.betas
        self.beta_alphas =  torch.cumprod(self.alphas,dim=0)

        # Calculation for Algorithm 1 := sqrt(alpha_bar), sqrt(1-alpha_bar)
        self.register_buffer(
            "sqrt_beta_alphas",
            torch.sqrt(self.beta_alphas)
        )
        self.register_buffer(
            "sqrt_one_minus_beta_alphas",
            torch.sqrt(1 - self.beta_alphas)
        )
    
    def forward(self,x_0):
        t = torch.randint(self.T,size=(x_0.shape[0],),device=x_0.device)
        noise = torch.randn_like(x_0)
        x_t = (
            extract(self.sqrt_beta_alphas,t,x_0.shape) * x_0 + 
            extract(self.sqrt_one_minus_beta_alphas,t,x_0.shape) * noise
        )
        loss = F.mse_loss(self.model(x_t,t),noise,reduction='mean')
        return loss
    

class GaussianDiffusionSampler(nn.Module):
    def __init__(self,beta_1,beta_t,model, T) -> None:
        super().__init__()
        self.model = model 
        self.T = T
        self.register_buffer(
            "betas",
            torch.linspace(beta_1,beta_t,self.T).double()
        )
        self.alphas = 1 - self.betas
        self.beta_alphas = torch.cumprod(self.alphas,dim=0)

        """
         This line of code pads the tensor self.beta_alphas by adding a single element with the value 1 to the beginning of the tensor. 
         The resulting tensor is stored in self.beta_alphas_prev.
        """
        self.beta_alphas_prev = F.pad(self.beta_alphas,[1,0],value=1)[:T]

        self.register_buffer(
            "coeff1",
            (1 / torch.sqrt(self.alphas))
        )

        self.register_buffer(
            "coeff2",
            self.coeff1 * ((1- self.alphas) / (torch.sqrt(1-self.beta_alphas)))
        )

        self.register_buffer(
            "posterior_coeff",
            (1 - self.beta_alphas_prev) / (1-self.beta_alphas) * self.betas
        )

    def pred_xt_prev_mean_from_eps(self,x_t,t,eps):
        return (
            extract(self.coeff1,t,x_t.shape) * x_t - 
            extract(self.coeff2,t,x_t.shape) * eps
        )
    
    def p_mean_variance(self,x_t,t):
        var = torch.cat([self.posterior_coeff[1:2],self.betas[1:]])
        var = extract(var,t,x_t.shape)

        eps = self.model(x_t,t)
        xt_prev_mean = self.pred_xt_prev_mean_from_eps(x_t,t,eps)
        return xt_prev_mean,var
    
    def forward(self,x_T):
        x_t=x_T.to("cuda")
        for timestep in reversed(range(self.T)):
            print(f"Sampling timestep: {timestep}")

            t = x_t.new_ones([x_t.shape[0],], dtype=torch.long) * timestep
            mean, var = self.p_mean_variance(x_t,t)
            mean , var = mean.to("cuda"), var.to("cuda")
            if timestep > 0:
                noise = torch.randn_like(x_t).to("cuda")
            else:
                noise = 0
            x_t = mean + torch.sqrt(var) * noise
        
        x_0 = x_t
        return torch.clip(x_0,-1,1)


class DDIMSampler(nn.Module):
    def __init__(self,model,n_steps,beta_1,beta_T,ddim_discretize: str = "uniform", ddim_eta = 0.1) -> None:
        super().__init__()
        self.steps = n_steps
        self.model = model
        
        if ddim_discretize == "uniform":
            c = self.steps // n_steps
            self.time_steps = torch.asarray(list(range(0,self.steps,c))) + 1
            print(f"Discreatization uniform : {self.time_steps}")

        elif ddim_discretize == "quad":
            self.time_steps = (torch.linspace(0,torch.sqrt(self.steps * .8),n_steps) ** 2).type(torch.int) + 1
            print(f"Quad descreatization : {self.time_steps}")
        else:
            raise NotImplementedError(ddim_discretize)
        
        
        self.register_buffer(
            "betas",
            torch.linspace(beta_1,beta_T,self.steps).double()
        ) 

        self.alphas = 1 - self.betas
        self.alpha_bar = torch.cumprod(self.alphas,dim=0)

        with torch.no_grad():
            self.ddim_alpha = self.alpha_bar[self.time_steps].clone().to(torch.float32)
            self.ddim_alpha_sqrt = torch.sqrt(self.ddim_alpha)
            self.ddim_alpha_prev = torch.cat([self.alpha_bar[0:1], self.alpha_bar[self.time_steps:-1]])
            self.ddim_sigma = (ddim_eta * (
                (1 - self.ddim_alpha_prev) / (1 - self.ddim_alpha) * 
                (1 - self.ddim_alpha / self.ddim_alpha_prev)
                ) ** .5)
            self.ddim_sqrt_one_minus_alpha  = torch.sqrt(1 - self.ddim_alpha)

    def get_eps(self, x: torch.Tensor, t: torch.Tensor, c: torch.Tensor, *,
                uncond_scale: float, uncond_cond: Optional[torch.Tensor]):
        if uncond_cond is None or uncond_scale == 1.:
            return self.model(x, t, c)

        x_in = torch.cat([x] * 2)
        t_in = torch.cat([t] * 2)

        c_in = torch.cat([uncond_cond, c])

        e_t_uncond, e_t_cond = self.model(x_in, t_in, c_in).chunk(2)

        e_t = e_t_uncond + uncond_scale * (e_t_cond - e_t_uncond)

        return e_t 

    @torch.no_grad()
    def forward(self, 
                shape: List[int],
                cond: torch.Tensor,
                repeat_noise: bool = False,
                temperature: float = 1.,
                x_last: Optional[torch.Tensor] = None,
                uncond_scale: float = 1.,
                uncond_cond: Optional[torch.Tensor] = None,
                skip_steps: int = 0,
                ):

        device = self.model.device
        batch_size = shape[0]

        # Get xtS
        x = x_last if x_last is not None else torch.randn(shape,device=device)
        time_steps = torch.flip(self.time_steps,[0])[skip_steps:]
        for i, step in enumerate(reversed(range(time_steps))):
            index  = len(time_steps) - i - 1
            ts = x.new_full((batch_size,),step,dtype=torch.long)
            x, pred_x0, e_t = self.p_sample(x=x,c=cond,t=ts,step=step,index=index,
                                            repeat_noise=repeat_noise,temperature=temperature,uncond_scale=uncond_scale,
                                            uncond_cond=uncond_cond) # type: ignore
        return x # Return x0
    
    @torch.no_grad()
    def p_sample(self, x: torch.Tensor, c:torch.Tensor, t:torch.Tensor, step , index,
                 repeat_noise: bool=False, temperature:float =1.,
                 uncond_scale: float = 1.,
                 uncond_cond: Optional[torch.Tensor] = None):
        e_t = self.get_eps(x,t,c,uncond_cond=uncond_cond,uncond_scale=uncond_scale)

        x_prev ,pred_x0 = self.get_x_prev_and_pred_x0(e_t,index,x,
                                                      temperature,repeat_noise)
        
        return x_prev, pred_x0, e_t
    
    def get_x_prev_and_pred_x0(self, e_t: torch.Tensor, index: int, x: torch.Tensor, temperature: float, repeat_noise: bool):
        alpha = self.ddim_alpha[index]
        alpha_prev = self.ddim_alpha_prev[index]
        sigma = self.ddim_sigma[index]
        sqrt_one_minus_alpha = self.ddim_sqrt_one_minus_alpha[index]

        pred_x0 = (x - sqrt_one_minus_alpha * e_t) / (alpha ** 0.5)
        dir_xt = (1. - alpha_prev - sigma ** 2).sqrt() * e_t
        if sigma == 0.:
            noise = torch.zeros_like(x)

        if repeat_noise:
            noise = torch.randn((1,*x.shape[1:]), device=x.device)

        else:
            noise = torch.randn(x.shape, device=x.device)

        noise = noise * temperature

        x_prev = (alpha_prev ** 0.5) * pred_x0 + dir_xt + sigma * noise

        return x_prev, pred_x0