update_name

graph_dit/diffusion/diffusion_utils.py

@@ -0,0 +1,224 @@
+import torch
+from torch.nn import functional as F
+import numpy as np
+from utils import PlaceHolder
+
+
+def sum_except_batch(x):
+    return x.reshape(x.size(0), -1).sum(dim=-1)
+
+def assert_correctly_masked(variable, node_mask):
+    assert (variable * (1 - node_mask.long())).abs().max().item() < 1e-4, \
+        'Variables not masked properly.'
+
+def cosine_beta_schedule_discrete(timesteps, s=0.008):
+    """ Cosine schedule as proposed in https://openreview.net/forum?id=-NEXDKk8gZ. """
+    steps = timesteps + 2
+    x = np.linspace(0, steps, steps)
+
+    alphas_cumprod = np.cos(0.5 * np.pi * ((x / steps) + s) / (1 + s)) ** 2
+    alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
+    alphas = (alphas_cumprod[1:] / alphas_cumprod[:-1])
+    betas = 1 - alphas
+    return betas.squeeze()
+
+def custom_beta_schedule_discrete(timesteps, average_num_nodes=30, s=0.008):
+    """ Cosine schedule as proposed in https://openreview.net/forum?id=-NEXDKk8gZ. """
+    steps = timesteps + 2
+    x = np.linspace(0, steps, steps)
+
+    alphas_cumprod = np.cos(0.5 * np.pi * ((x / steps) + s) / (1 + s)) ** 2
+    alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
+    alphas = (alphas_cumprod[1:] / alphas_cumprod[:-1])
+    betas = 1 - alphas
+
+    assert timesteps >= 100
+
+    p = 4 / 5       # 1 - 1 / num_edge_classes
+    num_edges = average_num_nodes * (average_num_nodes - 1) / 2
+
+    # First 100 steps: only a few updates per graph
+    updates_per_graph = 1.2
+    beta_first = updates_per_graph / (p * num_edges)
+
+    betas[betas < beta_first] = beta_first
+    return np.array(betas)
+
+
+def check_mask_correct(variables, node_mask):
+    for i, variable in enumerate(variables):
+        if len(variable) > 0:
+            assert_correctly_masked(variable, node_mask)
+
+
+def check_tensor_same_size(*args):
+    for i, arg in enumerate(args):
+        if i == 0:
+            continue
+        assert args[0].size() == arg.size()
+
+
+
+def reverse_tensor(x):
+    return x[torch.arange(x.size(0) - 1, -1, -1)]
+
+
+def sample_discrete_features(probX, probE, node_mask, step=None, add_nose=True):
+    ''' Sample features from multinomial distribution with given probabilities (probX, probE, proby)
+        :param probX: bs, n, dx_out        node features
+        :param probE: bs, n, n, de_out     edge features
+        :param proby: bs, dy_out           global features.
+    '''
+    bs, n, _ = probX.shape
+
+    # Noise X
+    # The masked rows should define probability distributions as well
+    probX[~node_mask] = 1 / probX.shape[-1]
+
+    # Flatten the probability tensor to sample with multinomial
+    probX = probX.reshape(bs * n, -1)       # (bs * n, dx_out)
+
+    # Sample X
+    probX = probX + 1e-12
+    probX = probX / probX.sum(dim=-1, keepdim=True)
+    X_t = probX.multinomial(1)      # (bs * n, 1)
+    X_t = X_t.reshape(bs, n)        # (bs, n)
+
+    # Noise E
+    # The masked rows should define probability distributions as well
+    inverse_edge_mask = ~(node_mask.unsqueeze(1) * node_mask.unsqueeze(2))
+    diag_mask = torch.eye(n).unsqueeze(0).expand(bs, -1, -1)
+
+    probE[inverse_edge_mask] = 1 / probE.shape[-1]
+    probE[diag_mask.bool()] = 1 / probE.shape[-1]
+    probE = probE.reshape(bs * n * n, -1)           # (bs * n * n, de_out)
+    probE = probE + 1e-12
+    probE = probE / probE.sum(dim=-1, keepdim=True)
+
+    # Sample E
+    E_t = probE.multinomial(1).reshape(bs, n, n)    # (bs, n, n)
+    E_t = torch.triu(E_t, diagonal=1)
+    E_t = (E_t + torch.transpose(E_t, 1, 2))
+
+    return PlaceHolder(X=X_t, E=E_t, y=torch.zeros(bs, 0).type_as(X_t))
+
+
+def compute_batched_over0_posterior_distribution(X_t, Qt, Qsb, Qtb):
+    """ M: X or E
+        Compute xt @ Qt.T * x0 @ Qsb / x0 @ Qtb @ xt.T for each possible value of x0
+        X_t: bs, n, dt          or bs, n, n, dt
+        Qt: bs, d_t-1, dt
+        Qsb: bs, d0, d_t-1
+        Qtb: bs, d0, dt.
+    """
+    X_t = X_t.float()
+    Qt_T = Qt.transpose(-1, -2).float()                                       # bs, N, dt
+    assert Qt.dim() == 3
+    left_term = X_t @ Qt_T
+    left_term = left_term.unsqueeze(dim=2)                    # bs, N, 1, d_t-1
+    right_term = Qsb.unsqueeze(1) 
+    numerator = left_term * right_term                        # bs, N, d0, d_t-1
+    
+    denominator = Qtb @ X_t.transpose(-1, -2)                 # bs, d0, N
+    denominator = denominator.transpose(-1, -2)               # bs, N, d0
+    denominator = denominator.unsqueeze(-1)                   # bs, N, d0, 1
+
+    denominator[denominator == 0] = 1.
+    return numerator / denominator
+
+
+def mask_distributions(true_X, true_E, pred_X, pred_E, node_mask):
+    # Add a small value everywhere to avoid nans
+    pred_X = pred_X.clamp_min(1e-18)
+    pred_X = pred_X / torch.sum(pred_X, dim=-1, keepdim=True)
+
+    pred_E = pred_E.clamp_min(1e-18)
+    pred_E = pred_E / torch.sum(pred_E, dim=-1, keepdim=True)
+
+    # Set masked rows to arbitrary distributions, so it doesn't contribute to loss
+    row_X = torch.ones(true_X.size(-1), dtype=true_X.dtype, device=true_X.device)
+    row_E = torch.zeros(true_E.size(-1), dtype=true_E.dtype, device=true_E.device).clamp_min(1e-18)
+    row_E[0] = 1.
+
+    diag_mask = ~torch.eye(node_mask.size(1), device=node_mask.device, dtype=torch.bool).unsqueeze(0)
+    true_X[~node_mask] = row_X
+    true_E[~(node_mask.unsqueeze(1) * node_mask.unsqueeze(2) * diag_mask), :] = row_E
+    pred_X[~node_mask] = row_X
+    pred_E[~(node_mask.unsqueeze(1) * node_mask.unsqueeze(2) * diag_mask), :] = row_E
+
+    return true_X, true_E, pred_X, pred_E
+
+def posterior_distributions(X, X_t, Qt, Qsb, Qtb, X_dim):
+    bs, n, d = X.shape
+    X = X.float()
+    Qt_X_T = torch.transpose(Qt.X, -2, -1).float()                  # (bs, d, d)
+    left_term = X_t @ Qt_X_T                                        # (bs, N, d)
+    right_term = X @ Qsb.X                                          # (bs, N, d)
+    
+    numerator = left_term * right_term                              # (bs, N, d)
+    denominator = X @ Qtb.X                                         # (bs, N, d) @ (bs, d, d) = (bs, N, d)
+    denominator = denominator * X_t 
+    
+    num_X = numerator[:, :, :X_dim]
+    num_E = numerator[:, :, X_dim:].reshape(bs, n*n, -1)
+
+    deno_X = denominator[:, :, :X_dim]
+    deno_E = denominator[:, :, X_dim:].reshape(bs, n*n, -1)
+
+    # denominator = (denominator * X_t).sum(dim=-1)                   # (bs, N, d) * (bs, N, d) + sum = (bs, N)
+    denominator = denominator.unsqueeze(-1)                         # (bs, N, 1)
+
+    deno_X = deno_X.sum(dim=-1).unsqueeze(-1)
+    deno_E = deno_E.sum(dim=-1).unsqueeze(-1)
+
+    deno_X[deno_X == 0.] = 1
+    deno_E[deno_E == 0.] = 1
+    prob_X = num_X / deno_X
+    prob_E = num_E / deno_E
+    
+    prob_E = prob_E / prob_E.sum(dim=-1, keepdim=True)
+    prob_X = prob_X / prob_X.sum(dim=-1, keepdim=True)
+    return PlaceHolder(X=prob_X, E=prob_E, y=None)
+
+
+def sample_discrete_feature_noise(limit_dist, node_mask):
+    """ Sample from the limit distribution of the diffusion process"""
+    bs, n_max = node_mask.shape
+    x_limit = limit_dist.X[None, None, :].expand(bs, n_max, -1)
+    U_X = x_limit.flatten(end_dim=-2).multinomial(1).reshape(bs, n_max)
+    U_X = F.one_hot(U_X.long(), num_classes=x_limit.shape[-1]).float()
+
+    e_limit = limit_dist.E[None, None, None, :].expand(bs, n_max, n_max, -1)
+    U_E = e_limit.flatten(end_dim=-2).multinomial(1).reshape(bs, n_max, n_max)
+    U_E = F.one_hot(U_E.long(), num_classes=e_limit.shape[-1]).float()
+
+    U_X = U_X.to(node_mask.device)
+    U_E = U_E.to(node_mask.device)
+
+    # Get upper triangular part of edge noise, without main diagonal
+    upper_triangular_mask = torch.zeros_like(U_E)
+    indices = torch.triu_indices(row=U_E.size(1), col=U_E.size(2), offset=1)
+    upper_triangular_mask[:, indices[0], indices[1], :] = 1
+
+    U_E = U_E * upper_triangular_mask
+    U_E = (U_E + torch.transpose(U_E, 1, 2))
+
+    assert (U_E == torch.transpose(U_E, 1, 2)).all()
+    return PlaceHolder(X=U_X, E=U_E, y=None).mask(node_mask)
+
+def index_QE(X, q_e, n_bond=5):
+    bs, n, n_atom = X.shape
+    node_indices = X.argmax(-1)  # (bs, n)
+
+    exp_ind1 = node_indices[ :, :, None, None, None].expand(bs, n, n_atom, n_bond, n_bond)
+    exp_ind2 = node_indices[ :, :, None, None, None].expand(bs, n, n, n_bond, n_bond)
+    
+    q_e = torch.gather(q_e, 1, exp_ind1)
+    q_e = torch.gather(q_e, 2, exp_ind2) # (bs, n, n, n_bond, n_bond)
+
+
+    node_mask = X.sum(-1) != 0
+    no_edge = (~node_mask)[:, :, None] & (~node_mask)[:, None, :]
+    q_e[no_edge] = torch.tensor([1, 0, 0, 0, 0]).type_as(q_e)
+
+    return q_e

graph_dit/diffusion/distributions.py

@@ -0,0 +1,30 @@
+import torch
+
+class DistributionNodes:
+    def __init__(self, histogram):
+        """ Compute the distribution of the number of nodes in the dataset, and sample from this distribution.
+            historgram: dict. The keys are num_nodes, the values are counts
+        """
+
+        if type(histogram) == dict:
+            max_n_nodes = max(histogram.keys())
+            prob = torch.zeros(max_n_nodes + 1)
+            for num_nodes, count in histogram.items():
+                prob[num_nodes] = count
+        else:
+            prob = histogram
+
+        self.prob = prob / prob.sum()
+        self.m = torch.distributions.Categorical(prob)
+
+    def sample_n(self, n_samples, device):
+        idx = self.m.sample((n_samples,))
+        return idx.to(device)
+
+    def log_prob(self, batch_n_nodes):
+        assert len(batch_n_nodes.size()) == 1
+        p = self.prob.to(batch_n_nodes.device)
+
+        probas = p[batch_n_nodes]
+        log_p = torch.log(probas + 1e-30)
+        return log_p

graph_dit/diffusion/noise_schedule.py

@@ -0,0 +1,159 @@
+import torch
+import utils
+from diffusion import diffusion_utils
+    
+class PredefinedNoiseScheduleDiscrete(torch.nn.Module):
+    def __init__(self, noise_schedule, timesteps):
+        super(PredefinedNoiseScheduleDiscrete, self).__init__()
+        self.timesteps = timesteps
+
+        if noise_schedule == 'cosine':
+            betas = diffusion_utils.cosine_beta_schedule_discrete(timesteps)
+        elif noise_schedule == 'custom':
+            betas = diffusion_utils.custom_beta_schedule_discrete(timesteps)
+        else:
+            raise NotImplementedError(noise_schedule)
+
+        self.register_buffer('betas', torch.from_numpy(betas).float())
+
+        # 0.9999
+        self.alphas = 1 - torch.clamp(self.betas, min=0, max=1)
+
+        log_alpha = torch.log(self.alphas)
+        log_alpha_bar = torch.cumsum(log_alpha, dim=0)
+        self.alphas_bar = torch.exp(log_alpha_bar)
+
+    def forward(self, t_normalized=None, t_int=None):
+        assert int(t_normalized is None) + int(t_int is None) == 1
+        if t_int is None:
+            t_int = torch.round(t_normalized * self.timesteps)
+        return self.betas[t_int.long()]
+
+    def get_alpha_bar(self, t_normalized=None, t_int=None):
+        assert int(t_normalized is None) + int(t_int is None) == 1
+        if t_int is None:
+            t_int = torch.round(t_normalized * self.timesteps)
+        ### new
+        self.alphas_bar = self.alphas_bar.to(t_int.device)
+        return self.alphas_bar[t_int.long()]
+
+
+class DiscreteUniformTransition:
+    def __init__(self, x_classes: int, e_classes: int, y_classes: int):
+        self.X_classes = x_classes
+        self.E_classes = e_classes
+        self.y_classes = y_classes
+        self.u_x = torch.ones(1, self.X_classes, self.X_classes)
+        if self.X_classes > 0:
+            self.u_x = self.u_x / self.X_classes
+
+        self.u_e = torch.ones(1, self.E_classes, self.E_classes)
+        if self.E_classes > 0:
+            self.u_e = self.u_e / self.E_classes
+
+        self.u_y = torch.ones(1, self.y_classes, self.y_classes)
+        if self.y_classes > 0:
+            self.u_y = self.u_y / self.y_classes
+
+    def get_Qt(self, beta_t, device, X=None, flatten_e=None):
+        """ Returns one-step transition matrices for X and E, from step t - 1 to step t.
+        Qt = (1 - beta_t) * I + beta_t / K
+
+        beta_t: (bs)                         noise level between 0 and 1
+        returns: qx (bs, dx, dx), qe (bs, de, de), qy (bs, dy, dy).
+        """
+        beta_t = beta_t.unsqueeze(1)
+        beta_t = beta_t.to(device)
+        self.u_x = self.u_x.to(device)
+        self.u_e = self.u_e.to(device)
+        self.u_y = self.u_y.to(device)
+
+        q_x = beta_t * self.u_x + (1 - beta_t) * torch.eye(self.X_classes, device=device).unsqueeze(0)
+        q_e = beta_t * self.u_e + (1 - beta_t) * torch.eye(self.E_classes, device=device).unsqueeze(0)
+        q_y = beta_t * self.u_y + (1 - beta_t) * torch.eye(self.y_classes, device=device).unsqueeze(0)
+
+        return utils.PlaceHolder(X=q_x, E=q_e, y=q_y)
+
+    def get_Qt_bar(self, alpha_bar_t, device, X=None, flatten_e=None):
+        """ Returns t-step transition matrices for X and E, from step 0 to step t.
+        Qt = prod(1 - beta_t) * I + (1 - prod(1 - beta_t)) / K
+
+        alpha_bar_t: (bs)         Product of the (1 - beta_t) for each time step from 0 to t.
+        returns: qx (bs, dx, dx), qe (bs, de, de), qy (bs, dy, dy).
+        """
+        alpha_bar_t = alpha_bar_t.unsqueeze(1)
+        alpha_bar_t = alpha_bar_t.to(device)
+        self.u_x = self.u_x.to(device)
+        self.u_e = self.u_e.to(device)
+        self.u_y = self.u_y.to(device)
+
+        q_x = alpha_bar_t * torch.eye(self.X_classes, device=device).unsqueeze(0) + (1 - alpha_bar_t) * self.u_x
+        q_e = alpha_bar_t * torch.eye(self.E_classes, device=device).unsqueeze(0) + (1 - alpha_bar_t) * self.u_e
+        q_y = alpha_bar_t * torch.eye(self.y_classes, device=device).unsqueeze(0) + (1 - alpha_bar_t) * self.u_y
+
+        return utils.PlaceHolder(X=q_x, E=q_e, y=q_y)
+
+    
+class MarginalTransition:
+    def __init__(self, x_marginals, e_marginals, xe_conditions, ex_conditions, y_classes, n_nodes):
+        self.X_classes = len(x_marginals)
+        self.E_classes = len(e_marginals)
+        self.y_classes = y_classes
+        self.x_marginals = x_marginals # Dx
+        self.e_marginals = e_marginals # Dx, De
+        self.xe_conditions = xe_conditions
+
+        self.u_x = x_marginals.unsqueeze(0).expand(self.X_classes, -1).unsqueeze(0) # 1, Dx, Dx
+        self.u_e = e_marginals.unsqueeze(0).expand(self.E_classes, -1).unsqueeze(0) # 1, De, De
+        self.u_xe = xe_conditions.unsqueeze(0) # 1, Dx, De
+        self.u_ex = ex_conditions.unsqueeze(0) # 1, De, Dx
+        self.u = self.get_union_transition(self.u_x, self.u_e, self.u_xe, self.u_ex, n_nodes) # 1, Dx + n*De, Dx + n*De
+
+    def get_union_transition(self, u_x, u_e, u_xe, u_ex, n_nodes):
+        u_e = u_e.repeat(1, n_nodes, n_nodes) # (1, n*de, n*de)
+        u_xe = u_xe.repeat(1, 1, n_nodes) # (1, dx, n*de)
+        u_ex = u_ex.repeat(1, n_nodes, 1) # (1, n*de, dx)
+        u0 = torch.cat([u_x, u_xe], dim=2) # (1, dx, dx + n*de)
+        u1 = torch.cat([u_ex, u_e], dim=2) # (1, n*de, dx + n*de)
+        u = torch.cat([u0, u1], dim=1) # (1, dx + n*de, dx + n*de)
+        return u
+
+    def index_edge_margin(self, X, q_e, n_bond=5):
+        # q_e: (bs, dx, de) --> (bs, n, de)
+        bs, n, n_atom = X.shape
+        node_indices = X.argmax(-1)  # (bs, n)
+        ind = node_indices[ :, :, None].expand(bs, n, n_bond)
+        q_e = torch.gather(q_e, 1, ind)
+        return q_e
+    
+    def get_Qt(self, beta_t, device):
+        """ Returns one-step transition matrices for X and E, from step t - 1 to step t.
+        Qt = (1 - beta_t) * I + beta_t / K
+        beta_t: (bs)
+        returns: q (bs, d0, d0)
+        """
+        bs = beta_t.size(0)
+        d0 = self.u.size(-1)
+        self.u = self.u.to(device)
+        u = self.u.expand(bs, d0, d0)
+
+        beta_t = beta_t.to(device)
+        beta_t = beta_t.view(bs, 1, 1)
+        q = beta_t * u + (1 - beta_t) * torch.eye(d0, device=device).unsqueeze(0)
+
+        return utils.PlaceHolder(X=q, E=None, y=None)
+    
+    def get_Qt_bar(self, alpha_bar_t, device):
+        """ Returns t-step transition matrices for X and E, from step 0 to step t.
+        Qt = prod(1 - beta_t) * I + (1 - prod(1 - beta_t)) * K
+        alpha_bar_t: (bs, 1) roduct of the (1 - beta_t) for each time step from 0 to t.
+        returns: q (bs, d0, d0)
+        """
+        bs = alpha_bar_t.size(0)
+        d0 = self.u.size(-1)
+        alpha_bar_t = alpha_bar_t.to(device)
+        alpha_bar_t = alpha_bar_t.view(bs, 1, 1)
+        self.u = self.u.to(device)
+        q = alpha_bar_t * torch.eye(d0, device=device).unsqueeze(0) + (1 - alpha_bar_t) * self.u
+
+        return utils.PlaceHolder(X=q, E=None, y=None)