colors = torch.tensor(
[
[255, 255, 255],
- [0, 0, 0],
[255, 0, 0],
- [0, 128, 0],
+ [0, 192, 0],
[0, 0, 255],
- [255, 255, 0],
+ [255, 192, 0],
+ [0, 255, 255],
+ [255, 0, 255],
+ [192, 255, 192],
+ [255, 192, 192],
+ [192, 192, 255],
[192, 192, 192],
]
)
-token2char = "_X" + "".join([str(n) for n in range(len(colors) - 2)]) + ">"
+token_background = 0
+first_bird_token = 1
+nb_bird_tokens = colors.size(0) - 1
+token_forward = first_bird_token + nb_bird_tokens
+token_backward = token_forward + 1
+token2char = "_" + "".join([chr(ord("A") + n) for n in range(len(colors) - 1)]) + "><"
-def generate(
+
+def generate_seq(
+ nb, height, width, nb_birds=3, nb_iterations=2, return_iterations=False
+):
+ pairs = []
+ kept_iterations = []
+
+ for _ in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world generation"):
+ while True:
+ iterations = []
+
+ f_start = torch.zeros(height, width, dtype=torch.int64)
+
+ i, j, vi, vj = (
+ torch.empty(nb_birds, dtype=torch.int64),
+ torch.empty(nb_birds, dtype=torch.int64),
+ torch.empty(nb_birds, dtype=torch.int64),
+ torch.empty(nb_birds, dtype=torch.int64),
+ )
+
+ col = torch.randperm(colors.size(0) - 1)[:nb_birds].sort().values + 1
+
+ for n in range(nb_birds):
+ c = col[n]
+
+ while True:
+ i[n], j[n] = (
+ torch.randint(height, (1,))[0],
+ torch.randint(width, (1,))[0],
+ )
+ vm = torch.randint(4, (1,))[0]
+ vi[n], vj[n] = (vm % 2) * 2 - 1, (vm // 2) * 2 - 1
+ if (
+ i[n] - vi[n] >= 0
+ and i[n] - vi[n] < height
+ and j[n] - vj[n] >= 0
+ and j[n] - vj[n] < width
+ and f_start[i[n], j[n]] == 0
+ and f_start[i[n] - vi[n], j[n]] == 0
+ and f_start[i[n], j[n] - vj[n]] == 0
+ ):
+ break
+
+ f_start[i[n], j[n]] = c
+ f_start[i[n] - vi[n], j[n]] = c
+ f_start[i[n], j[n] - vj[n]] = c
+
+ f_end = f_start.clone()
+
+ for l in range(nb_iterations):
+ iterations.append(f_end.clone())
+ f_end[...] = 0
+ nb_collisions = 0
+ for n in range(nb_birds):
+ c = col[n]
+
+ pi, pj, pvi, pvj = (
+ i[n].item(),
+ j[n].item(),
+ vi[n].item(),
+ vj[n].item(),
+ )
+
+ if (i[n] == 0 and vi[n] == -1) or (
+ i[n] == height - 1 and vi[n] == 1
+ ):
+ vi[n] = -vi[n]
+ if (j[n] == 0 and vj[n] == -1) or (
+ j[n] == width - 1 and vj[n] == 1
+ ):
+ vj[n] = -vj[n]
+
+ i[n] += vi[n]
+ j[n] += vj[n]
+
+ if not (
+ f_end[i[n], j[n]] == 0
+ and f_end[i[n] - vi[n], j[n]] == 0
+ and f_end[i[n], j[n] - vj[n]] == 0
+ ):
+ nb_collisions += 1
+
+ f_end[i[n], j[n]] = c
+ f_end[i[n] - vi[n], j[n]] = c
+ f_end[i[n], j[n] - vj[n]] = c
+
+ iterations.append(f_end.clone())
+
+ if nb_collisions == 0:
+ break
+
+ kept_iterations.append(iterations)
+ pairs.append((f_start, f_end))
+
+ 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, :]
+ )
+
+ if return_iterations:
+ # iterations = torch.cat([ torch.cat([ x[None, None] for x in l], dim = 1) for l in kept_iterations ], dim=0)
+ return torch.cat(result, dim=0), kept_iterations
+ else:
+ return torch.cat(result, dim=0)
+
+
+######################################################################
+
+
+def generate_seq_old(
nb,
height,
width,
- max_nb_obj=colors.size(0) - 2,
+ nb_birds=3,
nb_iterations=2,
):
- f_start = torch.zeros(nb, height, width, dtype=torch.int64)
- f_end = torch.zeros(nb, height, width, dtype=torch.int64)
- n = torch.arange(f_start.size(0))
+ pairs = []
- for n in range(nb):
- nb_fish = torch.randint(max_nb_obj, (1,)).item() + 1
- for c in torch.randperm(colors.size(0) - 2)[:nb_fish].sort().values:
+ 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))
+
+ for c in (
+ (torch.randperm(nb_bird_tokens) + first_bird_token)[:nb_birds].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)
- f_start[n, i, j] = c + 2
- f_start[n, i - vi, j - vj] = c + 2
- f_start[n, i + vj, j - vi] = c + 2
- f_start[n, i - vj, j + vi] = c + 2
+ 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
i += vi
j += vj
- f_end[n, i, j] = c + 2
- f_end[n, i - vi, j - vj] = c + 2
- f_end[n, i + vj, j - vi] = c + 2
- f_end[n, i - vj, j + vi] = c + 2
+ 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
- return torch.cat(
- [
- f_end.flatten(1),
- torch.full((f_end.size(0), 1), len(colors)),
- f_start.flatten(1),
- ],
- dim=1,
- )
+ pairs.append((f_start, f_end))
+
+ 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):
- f_start = seq[:, : height * width].reshape(-1, height, width)
- f_start = (f_start >= len(colors)).long() + (f_start < len(colors)).long() * f_start
- f_end = seq[:, height * width + 1 :].reshape(-1, height, width)
- f_end = (f_end >= len(colors)).long() + (f_end < len(colors)).long() * f_end
- img_f_start, img_f_end = colors[f_start], colors[f_end]
+def frame2img(x, height, width, upscale=15):
+ x = x.reshape(-1, height, width)
+ m = torch.logical_and(x >= 0, x < first_bird_token + nb_bird_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)
- img = torch.cat(
+ x[:, :, :, torch.arange(0, x.size(3), upscale)] = 0
+ x[:, :, torch.arange(0, x.size(2), upscale), :] = 0
+ x = x[:, :, 1:, 1:]
+
+ 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
+
+
+def seq2img(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]
+
+ direction_symbol = torch.full((direction.size(0), height * upscale - 1, upscale), 0)
+ direction_symbol = colors[direction_symbol].permute(0, 3, 1, 2)
+ separator = torch.full((direction.size(0), 3, height * upscale - 1, 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 + upscale // 2 - 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 + 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
+
+ return torch.cat(
[
- img_f_start,
- torch.full(
- (img_f_start.size(0), img_f_start.size(1), 1, img_f_start.size(3)), 1
- ),
- img_f_end,
+ frame2img(f_first, height, width, upscale),
+ separator,
+ direction_symbol,
+ separator,
+ frame2img(f_second, height, width, 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, max_nb_obj=3)
+ seq, it = generate_seq(
+ nb=64, height=height, width=width, nb_iterations=100, return_iterations=True
+ )
delay = time.perf_counter() - start_time
print(f"{seq.size(0)/delay:02f} samples/s")
print(seq2str(seq[:4]))
- img = sample2img(seq, height, width)
+ for t in range(len(it[0])):
+ img = torch.cat([frame2img(f[t], height, width) for f in it], dim=0)
+ torchvision.utils.save_image(
+ img.float() / 255.0,
+ f"/tmp/frame_{t:03d}.png",
+ nrow=8,
+ padding=6,
+ pad_value=0,
+ )
+
+ # m = (torch.rand(seq.size()) < 0.05).long()
+ # seq = (1 - m) * seq + m * 23
+
+ img = seq2img(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=6, pad_value=0
+ )