+######################################################################
+
+
+# ||x_i - c_j||^2 = ||x_i||^2 + ||c_j||^2 - 2<x_i, c_j>
+def sq2matrix(x, c):
+ nx = x.pow(2).sum(1)
+ nc = c.pow(2).sum(1)
+ return nx[:, None] + nc[None, :] - 2 * x @ c.t()
+
+
+def update_centroids(x, c, nb_min=1):
+ _, b = sq2matrix(x, c).min(1)
+ b.squeeze_()
+ nb_resets = 0
+
+ for k in range(0, c.size(0)):
+ i = b.eq(k).nonzero(as_tuple=False).squeeze()
+ if i.numel() >= nb_min:
+ c[k] = x.index_select(0, i).mean(0)
+ else:
+ n = torch.randint(x.size(0), (1,))
+ nb_resets += 1
+ c[k] = x[n]
+
+ return c, b, nb_resets
+
+
+def kmeans(x, nb_centroids, nb_min=1):
+ if x.size(0) < nb_centroids * nb_min:
+ print("Not enough points!")
+ exit(1)
+
+ c = x[torch.randperm(x.size(0))[:nb_centroids]]
+ t = torch.full((x.size(0),), -1)
+ n = 0
+
+ while True:
+ c, u, nb_resets = update_centroids(x, c, nb_min)
+ n = n + 1
+ nb_changes = (u - t).sign().abs().sum() + nb_resets
+ t = u
+ if nb_changes == 0:
+ break
+
+ return c, t
+
+
+######################################################################
+
+
+def patchify(x, factor, invert_size=None):
+ if invert_size is None:
+ return (
+ x.reshape(
+ x.size(0), # 0
+ x.size(1), # 1
+ factor, # 2
+ x.size(2) // factor, # 3
+ factor, # 4
+ x.size(3) // factor, # 5
+ )
+ .permute(0, 2, 4, 1, 3, 5)
+ .reshape(-1, x.size(1), x.size(2) // factor, x.size(3) // factor)
+ )
+ else:
+ return (
+ x.reshape(
+ invert_size[0], # 0
+ factor, # 1
+ factor, # 2
+ invert_size[1], # 3
+ invert_size[2] // factor, # 4
+ invert_size[3] // factor, # 5
+ )
+ .permute(0, 3, 1, 4, 2, 5)
+ .reshape(invert_size)
+ )
+
+
+class Normalizer(nn.Module):
+ def __init__(self, mu, std):
+ super().__init__()
+ self.mu = nn.Parameter(mu)
+ self.log_var = nn.Parameter(2 * torch.log(std))
+
+ def forward(self, x):
+ return (x - self.mu) / torch.exp(self.log_var / 2.0)
+
+
+class SignSTE(nn.Module):
+ def __init__(self):
+ super().__init__()
+
+ def forward(self, x):
+ # torch.sign() takes three values
+ s = (x >= 0).float() * 2 - 1
+ if self.training:
+ u = torch.tanh(x)
+ return s + u - u.detach()
+ else:
+ return s
+
+
+def train_encoder(
+ train_input,
+ dim_hidden=64,
+ block_size=16,
+ nb_bits_per_block=10,
+ lr_start=1e-3,
+ lr_end=1e-5,
+ nb_epochs=10,
+ batch_size=25,
+ device=torch.device("cpu"),
+):
+ mu, std = train_input.mean(), train_input.std()
+
+ encoder = nn.Sequential(
+ Normalizer(mu, std),
+ nn.Conv2d(3, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(
+ dim_hidden,
+ nb_bits_per_block,
+ kernel_size=block_size,
+ stride=block_size,
+ padding=0,
+ ),
+ SignSTE(),
+ )
+
+ decoder = nn.Sequential(
+ nn.ConvTranspose2d(
+ nb_bits_per_block,
+ dim_hidden,
+ kernel_size=block_size,
+ stride=block_size,
+ padding=0,
+ ),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
+ nn.ReLU(),
+ nn.Conv2d(dim_hidden, 3, kernel_size=5, stride=1, padding=2),
+ )
+
+ model = nn.Sequential(encoder, decoder)
+
+ nb_parameters = sum(p.numel() for p in model.parameters())
+
+ print(f"nb_parameters {nb_parameters}")
+
+ model.to(device)
+
+ for k in range(nb_epochs):
+ lr = math.exp(
+ math.log(lr_start) + math.log(lr_end / lr_start) / (nb_epochs - 1) * k
+ )
+ print(f"lr {lr}")
+ optimizer = torch.optim.Adam(model.parameters(), lr=lr)
+ acc_loss, nb_samples = 0.0, 0
+
+ for input in tqdm.tqdm(
+ train_input.split(batch_size),
+ dynamic_ncols=True,
+ desc="vqae-train",
+ total=train_input.size(0) // batch_size,
+ ):
+ output = model(input)
+ loss = F.mse_loss(output, input)
+ acc_loss += loss.item() * input.size(0)
+ nb_samples += input.size(0)
+
+ optimizer.zero_grad()
+ loss.backward()
+ optimizer.step()
+
+ print(f"loss {k} {acc_loss/nb_samples}")
+ sys.stdout.flush()
+
+ return encoder, decoder
+
+
+######################################################################
+