File size: 9,781 Bytes
ed8c297
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
import torch
import torch.nn.functional as F
import math

def linear_beta_schedule(timesteps):
    scale = 1.0 # for 100 steps
    beta_start = scale * 0.0001
    beta_end = scale   * 0.02
    return torch.linspace(beta_start, beta_end, timesteps, dtype=torch.float64)

class GaussianDiffusion:
    def __init__(
        self,
        device,
        fix_mode=False,
        text_emb=False,
        fixed_frames=2,
        seq_len=16,
        timesteps=100,
        beta_schedule='linear',
    ):
        self.device = device
        self.fix_mode = fix_mode            # autoregressive
        self.fixed_frames = fixed_frames    # number of frames to fix
        self.timesteps = timesteps
        self.text_emb = text_emb
        self.seq_len = seq_len
        
        
        if beta_schedule == 'linear':
            betas = linear_beta_schedule(timesteps)
        elif beta_schedule == 'cosine':
            raise NotImplementedError('cosine schedule is not implemented yet!')
        else:
            raise ValueError(f'unknown beta schedule {beta_schedule}')
        
        self.betas = betas.to(self.device)
        self.alphas = (1. - self.betas).to(self.device)
        self.alphas_cumprod = torch.cumprod(self.alphas, axis=0).to(self.device)
        self.alphas_cumprod_prev = F.pad(self.alphas_cumprod[:-1], (1, 0), value=1.).to(self.device)
        
        # calculations for diffusion q(x_t | x_{t-1}) and others
        self.sqrt_alphas_cumprod = torch.sqrt(self.alphas_cumprod).to(self.device)
        self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - self.alphas_cumprod).to(self.device)
        self.log_one_minus_alphas_cumprod = torch.log(1.0 - self.alphas_cumprod).to(self.device)
        self.sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / self.alphas_cumprod).to(self.device)
        self.sqrt_recipm1_alphas_cumprod = torch.sqrt(1.0 / self.alphas_cumprod - 1).to(self.device)
        
        # calculations for posterior q(x_{t-1} | x_t, x_0)
        self.posterior_variance = (
            self.betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
        ).to(self.device)
        # below: log calculation clipped because the posterior variance is 0 at the beginning
        # of the diffusion chain
        self.posterior_log_variance_clipped = torch.log(self.posterior_variance.clamp(min =1e-20)).to(self.device)
        
        self.posterior_mean_coef1 = (
            self.betas * torch.sqrt(self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
        ).to(self.device)
        self.posterior_mean_coef2 = (
            (1.0 - self.alphas_cumprod_prev)
            * torch.sqrt(self.alphas)
            / (1.0 - self.alphas_cumprod)
        ).to(self.device)
    
    # get the param of given timestep t
    def _extract(self, a, t, x_shape):
        batch_size = t.shape[0]
        out = a.to(t.device).gather(0, t).float()
        out = out.reshape(batch_size, *((1,) * (len(x_shape) - 1))).to(self.device)
        return out
    
    # forward diffusion (using the nice property): q(x_t | x_0)
    def q_sample(self, x_start, t, noise=None):
        if noise is None:
            noise = torch.randn_like(x_start)

        sqrt_alphas_cumprod_t = self._extract(self.sqrt_alphas_cumprod, t, x_start.shape)
        sqrt_one_minus_alphas_cumprod_t = self._extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape)
        return sqrt_alphas_cumprod_t * x_start + sqrt_one_minus_alphas_cumprod_t * noise
    
    # Get the mean and variance of q(x_t | x_0).
    def q_mean_variance(self, x_start, t):
        mean = self._extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
        variance = self._extract(1.0 - self.alphas_cumprod, t, x_start.shape)
        log_variance = self._extract(self.log_one_minus_alphas_cumprod, t, x_start.shape)
        return mean, variance, log_variance
    
    # Compute the mean and variance of the diffusion posterior: q(x_{t-1} | x_t, x_0)
    def q_posterior_mean_variance(self, x_start, x_t, t):
        posterior_mean = (
            self._extract(self.posterior_mean_coef1, t, x_t.shape) * x_start
            + self._extract(self.posterior_mean_coef2, t, x_t.shape) * x_t
        )
        posterior_variance = self._extract(self.posterior_variance, t, x_t.shape)
        posterior_log_variance_clipped = self._extract(self.posterior_log_variance_clipped, t, x_t.shape)
        return posterior_mean, posterior_variance, posterior_log_variance_clipped
    
    # compute x_0 from x_t and pred noise: the reverse of `q_sample`
    def predict_start_from_noise(self, x_t, t, noise):
        return (
            self._extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
            self._extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
        )
    
    # compute predicted mean and variance of p(x_{t-1} | x_t)
    def p_mean_variance(self, model, x_t, t, clip_denoised=True, **kwargs):
        # predict noise using model
        assert 'text' in kwargs, 'text is required'
        assert 'prog_ind' in kwargs, 'prog_ind is required'
        assert 'joints_orig' in kwargs, 'joints_orig is required'
        pred_noise = model(x_t, t, 
                           text_emb=kwargs['text'], 
                           prog_ind=kwargs['prog_ind'], 
                           joints_orig=kwargs['joints_orig'])
        
        # use cfg for text embedding:
        if kwargs['use_cfg']:
            pred_noise_empty = model(x_t, t, 
                            text_emb=torch.zeros_like(kwargs['text']), 
                            prog_ind=kwargs['prog_ind'], 
                            joints_orig=kwargs['joints_orig'])
            pred_noise = pred_noise_empty + kwargs['cfg_alpha'] * (pred_noise - pred_noise_empty)
            
        # get the predicted x_0: different from the algorithm2 in the paper
        x_recon = self.predict_start_from_noise(x_t, t, pred_noise)
        
        if clip_denoised:
            x_recon = torch.clamp(x_recon, min=-1., max=1.)
        model_mean, posterior_variance, posterior_log_variance = \
                    self.q_posterior_mean_variance(x_recon, x_t, t)
        return model_mean, posterior_variance, posterior_log_variance
    
    # denoise_step: sample x_{t-1} from x_t and pred_noise
    # @torch.no_grad()
    def p_sample(self, model, x_t, t, clip_denoised=True, **kwargs):
        if 'disc_model' in kwargs:
            disc_model = kwargs['disc_model']
            try:
                cg_alpha = kwargs['cg_alpha'] # default 1.0
                cg_diffusion_steps = kwargs['cg_diffusion_steps']
            except:
                print("cg_alpha and cg_diffusion_steps are not provided!")
                print("Using default values: cg_alpha=1.0, cg_diffusion_steps=5")
                cg_alpha = 1.0
                cg_diffusion_steps = 5
        # predict mean and variance
        model_mean, _, model_log_variance = self.p_mean_variance(model, x_t, t,
                                                    clip_denoised=clip_denoised, **kwargs)
        model_mean = torch.tensor(model_mean, requires_grad=True)
        noise = torch.randn_like(x_t)
        # no noise when t == 0
        nonzero_mask = ((t != 0).float().view(-1, *([1] * (len(x_t.shape) - 1))))
        if t.item() < cg_diffusion_steps:
            pred_syn = disc_model(model_mean, t) # y = f(theta, x)  theta fixed 
            pred_syn.backward()
           
            grad = model_mean.grad * cg_alpha
            model_mean = model_mean - nonzero_mask * (0.5 * model_log_variance).exp() * grad

        # compute x_{t-1}
        pred_img = model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
        return pred_img
    
    # denoise: reverse diffusion
    # @torch.no_grad()
    def p_sample_loop(self, model, shape, fixed_points=None, **kwargs):
        batch_size = shape[0]
        device = next(model.parameters()).device
        
        # start from pure noise (for each example in the batch)
        img = torch.randn(shape, device=device)
        # notice that if we are in fixed mode, we need to fix the first 2 frames
        if self.fix_mode:
            assert not (fixed_points is None), 'fixed_points is required for fixed mode'
            img[:, :self.fixed_frames, :] = fixed_points
        imgs = []

        for i in reversed(range(0, self.timesteps)):
            img = self.p_sample(model, img, torch.full((batch_size,), i, device=device, dtype=torch.long), **kwargs)
            if self.fix_mode:
                img[:, :self.fixed_frames, :] = fixed_points
            imgs.append(img)
        return imgs


    # sample new images
    # @torch.no_grad()
    def sample(self, model, batch_size=1, seq_len=16, channels=135, 
               fixed_points=None, clip_denoised=True, **kwargs):
        return self.p_sample_loop(model, shape=(batch_size, seq_len, channels), 
                     fixed_points=fixed_points, clip_denoised=clip_denoised, **kwargs)
       
    # compute train losses
    def train_losses(self, model, x_start, t, mask=None, **kwargs):
        assert not (mask is None and self.fixed_mode), 'mask is required for fixed mode'
        if mask is None:
            mask = torch.zeros_like(x_start).to(dtype=torch.bool, device=self.device)

        mask_inv = torch.logical_not(mask)
        # generate random noise
        noise = torch.randn_like(x_start).to(device=self.device)
        noise[mask] = 0.
        
        # get x_t
        x_noisy = self.q_sample(x_start, t, noise=noise)
        predicted_noise = model(x_noisy, t, text_emb=kwargs['text'], prog_ind=kwargs['prog_ind'], joints_orig=kwargs['joints_orig'])

        loss = F.smooth_l1_loss(noise[mask_inv], predicted_noise[mask_inv])
        return loss