colors = torch.tensor(
[
[255, 255, 255],
- [0, 0, 0],
[255, 0, 0],
[0, 128, 0],
[0, 0, 255],
- [255, 255, 0],
+ [255, 200, 0],
[192, 192, 192],
]
)
-token2char = "_X01234>"
+token_background = 0
+first_fish_token = 1
+nb_fish_tokens = len(colors) - 1
+token_forward = first_fish_token + nb_fish_tokens
+token_backward = token_forward + 1
+
+token2char = "_" + "".join([str(n) for n in range(len(colors) - 1)]) + "><"
def generate(
nb,
height,
width,
- obj_length=6,
- mask_height=3,
- mask_width=3,
- nb_obj=3,
+ max_nb_obj=2,
+ nb_iterations=2,
):
- 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
+ pairs = []
+
+ for n in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world generation"):
+ f_start = torch.zeros(height, width, dtype=torch.int64)
+ f_end = torch.zeros(height, width, dtype=torch.int64)
+ n = torch.arange(f_start.size(0))
+
+ nb_fish = torch.randint(max_nb_obj, (1,)).item() + 1
+ for c in (
+ (torch.randperm(nb_fish_tokens) + first_fish_token)[:nb_fish].sort().values
+ ):
+ i, j = (
+ torch.randint(height - 2, (1,))[0] + 1,
+ torch.randint(width - 2, (1,))[0] + 1,
+ )
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
+
+ f_start[i, j] = c
+ f_start[i - vi, j - vj] = c
+ f_start[i + vj, j - vi] = c
+ f_start[i - vj, j + vi] = c
+
+ for l in range(nb_iterations):
i += vi
j += vj
- if i < 0 or i >= height or j < 0 or j >= width or intact[n, i, j] != 0:
+ if i < 0 or i >= height or j < 0 or j >= width:
i -= vi
j -= vj
- vi, vj = -vj, vi
+ vi, vj = -vi, -vj
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,
- )
+ f_end[i, j] = c
+ f_end[i - vi, j - vj] = c
+ f_end[i + vj, j - vi] = c
+ f_end[i - vj, j + vi] = c
+ pairs.append((f_start, f_end))
-def sample2img(seq, height, width):
- intact = seq[:, : height * width].reshape(-1, height, width)
- masked = seq[:, height * width + 1 :].reshape(-1, height, width)
- img_intact, img_masked = colors[intact], colors[masked]
+ result = []
+ for p in pairs:
+ if torch.rand(1) < 0.5:
+ result.append(
+ torch.cat(
+ [p[0].flatten(), torch.tensor([token_forward]), p[1].flatten()],
+ dim=0,
+ )[None, :]
+ )
+ else:
+ result.append(
+ torch.cat(
+ [p[1].flatten(), torch.tensor([token_backward]), p[0].flatten()],
+ dim=0,
+ )[None, :]
+ )
+
+ return torch.cat(result, dim=0)
+
+
+def sample2img(seq, height, width, upscale=15):
+ f_first = seq[:, : height * width].reshape(-1, height, width)
+ f_second = seq[:, height * width + 1 :].reshape(-1, height, width)
+ direction = seq[:, height * width]
+
+ def mosaic(x, upscale):
+ x = x.reshape(-1, height, width)
+ m = torch.logical_and(x >= 0, x < first_fish_token + nb_fish_tokens).long()
+ x = colors[x * m].permute(0, 3, 1, 2)
+ s = x.shape
+ x = x[:, :, :, None, :, None].expand(-1, -1, -1, upscale, -1, upscale)
+ x = x.reshape(s[0], s[1], s[2] * upscale, s[3] * upscale)
+
+ for n in range(m.size(0)):
+ for i in range(m.size(1)):
+ for j in range(m.size(2)):
+ if m[n, i, j] == 0:
+ for k in range(2, upscale - 2):
+ x[n, :, i * upscale + k, j * upscale + k] = 0
+ x[n, :, i * upscale + upscale - 1 - k, j * upscale + k] = 0
+
+ return x
+
+ direction_symbol = torch.full((direction.size(0), height * upscale, upscale), 0)
+ direction_symbol = colors[direction_symbol].permute(0, 3, 1, 2)
+ separator = torch.full((direction.size(0), 3, height * upscale, 1), 0)
+
+ for n in range(direction_symbol.size(0)):
+ if direction[n] == token_forward:
+ for k in range(upscale):
+ direction_symbol[
+ n,
+ :,
+ (height * upscale) // 2 - upscale // 2 + k,
+ 3 + abs(k - upscale // 2),
+ ] = 0
+ elif direction[n] == token_backward:
+ for k in range(upscale):
+ direction_symbol[
+ n,
+ :,
+ (height * upscale) // 2 - upscale // 2 + k,
+ 3 + upscale // 2 - abs(k - upscale // 2),
+ ] = 0
+ else:
+ for k in range(2, upscale - 2):
+ direction_symbol[
+ n, :, (height * upscale) // 2 - upscale // 2 + k, k
+ ] = 0
+ direction_symbol[
+ n, :, (height * upscale) // 2 - upscale // 2 + k, upscale - 1 - k
+ ] = 0
- img = torch.cat(
+ return torch.cat(
[
- img_intact,
- torch.full(
- (img_intact.size(0), img_intact.size(1), 1, img_intact.size(3)), 1
- ),
- img_masked,
+ mosaic(f_first, upscale),
+ separator,
+ direction_symbol,
+ separator,
+ mosaic(f_second, upscale),
],
- dim=2,
+ dim=3,
)
- return img.permute(0, 3, 1, 2)
-
def seq2str(seq):
result = []
height, width = 6, 8
start_time = time.perf_counter()
- seq = generate(nb=64, height=height, width=width)
+ seq = generate(nb=90, height=height, width=width, max_nb_obj=3)
delay = time.perf_counter() - start_time
print(f"{seq.size(0)/delay:02f} samples/s")
print(seq2str(seq[:4]))
+ m = (torch.rand(seq.size()) < 0.05).long()
+ seq = (1 - m) * seq + m * 23
+
img = sample2img(seq, height, width)
print(img.size())
- torchvision.utils.save_image(img.float() / 255.0, "world.png", nrow=8, padding=2)
+ torchvision.utils.save_image(
+ img.float() / 255.0, "/tmp/world.png", nrow=6, padding=4
+ )