- mu, std = train_input.float().mean(), train_input.float().std()
-
- def encoder_core(depth, dim):
- l = [
- [
- nn.Conv2d(
- dim * 2**k, dim * 2**k, kernel_size=5, stride=1, padding=2
- ),
- nn.ReLU(),
- nn.Conv2d(dim * 2**k, dim * 2 ** (k + 1), kernel_size=2, stride=2),
- nn.ReLU(),
- ]
- for k in range(depth)
- ]
-
- return nn.Sequential(*[x for m in l for x in m])
-
- def decoder_core(depth, dim):
- l = [
- [
- nn.ConvTranspose2d(
- dim * 2 ** (k + 1), dim * 2**k, kernel_size=2, stride=2
- ),
- nn.ReLU(),
- nn.ConvTranspose2d(
- dim * 2**k, dim * 2**k, kernel_size=5, stride=1, padding=2
- ),
- nn.ReLU(),
- ]
- for k in range(depth - 1, -1, -1)
- ]
-
- return nn.Sequential(*[x for m in l for x in m])
-
- encoder = nn.Sequential(
- Normalizer(mu, std),
- nn.Conv2d(3, dim_hidden, kernel_size=1, stride=1),
- nn.ReLU(),
- # 64x64
- encoder_core(depth=depth, dim=dim_hidden),
- # 8x8
- nn.Conv2d(dim_hidden * 2**depth, nb_bits_per_token, kernel_size=1, stride=1),
- )
-
- quantizer = SignSTE()
-
- decoder = nn.Sequential(
- nn.Conv2d(nb_bits_per_token, dim_hidden * 2**depth, kernel_size=1, stride=1),
- # 8x8
- decoder_core(depth=depth, dim=dim_hidden),
- # 64x64
- nn.ConvTranspose2d(dim_hidden, 3 * Box.nb_rgb_levels, kernel_size=1, stride=1),
- )
-
- 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
- )
- optimizer = torch.optim.Adam(model.parameters(), lr=lr)
-
- acc_train_loss = 0.0
-
- for input in tqdm.tqdm(train_input.split(batch_size), desc="vqae-train"):
- z = encoder(input)
- zq = z if k < 1 else quantizer(z)
- output = decoder(zq)
-
- output = output.reshape(
- output.size(0), -1, 3, output.size(2), output.size(3)
- )
-
- train_loss = F.cross_entropy(output, input)
-
- acc_train_loss += train_loss.item() * input.size(0)
-
- optimizer.zero_grad()
- train_loss.backward()
- optimizer.step()
-
- acc_test_loss = 0.0
-
- for input in tqdm.tqdm(test_input.split(batch_size), desc="vqae-test"):
- z = encoder(input)
- zq = z if k < 1 else quantizer(z)
- output = decoder(zq)
-
- output = output.reshape(
- output.size(0), -1, 3, output.size(2), output.size(3)
- )
-
- test_loss = F.cross_entropy(output, input)
-
- acc_test_loss += test_loss.item() * input.size(0)
-
- train_loss = acc_train_loss / train_input.size(0)
- test_loss = acc_test_loss / test_input.size(0)
-
- print(f"train_ae {k} lr {lr} train_loss {train_loss} test_loss {test_loss}")
- sys.stdout.flush()
-
- return encoder, quantizer, decoder
-
-
-######################################################################
-
-
-def scene2tensor(xh, yh, scene, size):
- width, height = size, size
- pixel_map = torch.ByteTensor(width, height, 4).fill_(255)
- data = pixel_map.numpy()
- surface = cairo.ImageSurface.create_for_data(
- data, cairo.FORMAT_ARGB32, width, height
- )
-
- ctx = cairo.Context(surface)
- ctx.set_fill_rule(cairo.FILL_RULE_EVEN_ODD)
-
- for b in scene:
- ctx.move_to(b.x * size, b.y * size)
- ctx.rel_line_to(b.w * size, 0)
- ctx.rel_line_to(0, b.h * size)
- ctx.rel_line_to(-b.w * size, 0)
- ctx.close_path()
- ctx.set_source_rgba(
- b.r / (Box.nb_rgb_levels - 1),
- b.g / (Box.nb_rgb_levels - 1),
- b.b / (Box.nb_rgb_levels - 1),
- 1.0,
- )
- ctx.fill()
-
- hs = size * 0.1
- ctx.set_source_rgba(0.0, 0.0, 0.0, 1.0)
- ctx.move_to(xh * size - hs / 2, yh * size - hs / 2)
- ctx.rel_line_to(hs, 0)
- ctx.rel_line_to(0, hs)
- ctx.rel_line_to(-hs, 0)
- ctx.close_path()
- ctx.fill()
-
- return (
- pixel_map[None, :, :, :3]
- .flip(-1)
- .permute(0, 3, 1, 2)
- .long()
- .mul(Box.nb_rgb_levels)
- .floor_divide(256)
+ intact = torch.zeros(nb, height, width, dtype=torch.int64)
+ n = torch.arange(intact.size(0))
+
+ for n in range(nb):
+ for c in torch.randperm(colors.size(0) - 2)[:nb_obj] + 2:
+ z = intact[n].flatten()
+ m = (torch.rand(z.size()) * (z == 0)).argmax(dim=0)
+ i, j = m // width, m % width
+ vm = torch.randint(4, (1,))[0]
+ vi, vj = (vm // 2) * (2 * (vm % 2) - 1), (1 - vm // 2) * (2 * (vm % 2) - 1)
+ for l in range(obj_length):
+ intact[n, i, j] = c
+ i += vi
+ j += vj
+ if i < 0 or i >= height or j < 0 or j >= width or intact[n, i, j] != 0:
+ i -= vi
+ j -= vj
+ vi, vj = -vj, vi
+ i += vi
+ j += vj
+ if (
+ i < 0
+ or i >= height
+ or j < 0
+ or j >= width
+ or intact[n, i, j] != 0
+ ):
+ break
+
+ masked = intact.clone()
+
+ for n in range(nb):
+ i = torch.randint(height - mask_height + 1, (1,))[0]
+ j = torch.randint(width - mask_width + 1, (1,))[0]
+ masked[n, i : i + mask_height, j : j + mask_width] = 1
+
+ return torch.cat(
+ [
+ masked.flatten(1),
+ torch.full((masked.size(0), 1), len(colors)),
+ intact.flatten(1),
+ ],
+ dim=1,