# Written by Francois Fleuret <francois@fleuret.org>
-import math, sys, tqdm
+import math, sys, tqdm, os
import torch, torchvision
######################################################################
+import problem
+
+
+class Sky(problem.Problem):
+ colors = torch.tensor(
+ [
+ [255, 255, 255],
+ [255, 0, 0],
+ [0, 192, 0],
+ [0, 0, 255],
+ [255, 192, 0],
+ [0, 255, 255],
+ [255, 0, 255],
+ [192, 255, 192],
+ [255, 192, 192],
+ [192, 192, 255],
+ [192, 192, 192],
+ ]
+ )
-colors = torch.tensor(
- [
- [255, 255, 255],
- [255, 0, 0],
- [0, 192, 0],
- [0, 0, 255],
- [255, 192, 0],
- [0, 255, 255],
- [255, 0, 255],
- [192, 255, 192],
- [255, 192, 192],
- [192, 192, 255],
- [192, 192, 192],
- ]
-)
-
-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)]) + "><"
-
-
-class Sky:
- def __init__(self, height, width):
- self.heigh = heigh
- self.width = width
+ 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_seq(
- nb, height, width, nb_birds=3, nb_iterations=2, return_iterations=False
+ def __init__(
+ self,
+ height=6,
+ width=8,
+ nb_birds=3,
+ speed=2,
+ nb_iterations=2,
+ avoid_collision=True,
):
- pairs = []
- kept_iterations = []
+ self.height = height
+ self.width = width
+ self.nb_birds = nb_birds
+ self.speed = speed
+ self.nb_iterations = nb_iterations
+ self.avoid_collision = avoid_collision
+
+ def direction_tokens(self):
+ return self.token_forward, self.token_backward
+
+ def generate_frame_sequences(self, nb):
+ frame_sequences = []
for _ in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world generation"):
- while True:
- iterations = []
+ i, j, vi, vj = (
+ torch.empty(self.nb_birds, dtype=torch.int64),
+ torch.empty(self.nb_birds, dtype=torch.int64),
+ torch.empty(self.nb_birds, dtype=torch.int64),
+ torch.empty(self.nb_birds, dtype=torch.int64),
+ )
- f_start = torch.zeros(height, width, dtype=torch.int64)
+ def collision_okay():
+ if not self.avoid_collision:
+ return True
- 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),
- )
+ count = torch.zeros(self.height, self.width, 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
+ for n in range(self.nb_birds):
+ count[i[n], j[n]] += 1
+ count[i[n] - vi[n], j[n]] += 1
+ count[i[n], j[n] - vj[n]] += 1
- 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
+ return count.max() <= 1
- f_end = f_start.clone()
+ col = (
+ torch.randperm(self.colors.size(0) - 1)[: self.nb_birds].sort().values
+ + 1
+ )
- 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]
+ while True:
+ while True:
+ for n in range(self.nb_birds):
+ while True:
+ i[n] = torch.randint(self.height, (1,))
+ j[n] = torch.randint(self.width, (1,))
+ vm = torch.randint(4, (1,))
+ vi[n], vj[n] = (vm % 2) * 2 - 1, (vm // 2) * 2 - 1
+ if (
+ i[n] - vi[n] >= 0
+ and i[n] - vi[n] < self.height
+ and j[n] - vj[n] >= 0
+ and j[n] - vj[n] < self.width
+ ):
+ break
+
+ if collision_okay():
+ break
+
+ result = torch.zeros(
+ self.nb_iterations * self.speed,
+ self.height,
+ self.width,
+ dtype=torch.int64,
+ )
+
+ fine = torch.empty(self.nb_iterations * self.speed)
- pi, pj, pvi, pvj = (
- i[n].item(),
- j[n].item(),
- vi[n].item(),
- vj[n].item(),
- )
+ t_to_keep = (
+ torch.arange(self.nb_iterations, device=result.device) * self.speed
+ )
+
+ for l in range(self.nb_iterations * self.speed):
+ fine[l] = collision_okay()
+ for n in range(self.nb_birds):
+ c = col[n]
+ result[l, i[n], j[n]] = c
+ result[l, i[n] - vi[n], j[n]] = c
+ result[l, i[n], j[n] - vj[n]] = c
if (i[n] == 0 and vi[n] == -1) or (
- i[n] == height - 1 and vi[n] == 1
+ i[n] == self.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
+ j[n] == self.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())
+ result = result[t_to_keep]
+ fine = fine[t_to_keep]
- if nb_collisions == 0:
+ if fine[-1]:
break
- kept_iterations.append(iterations)
- pairs.append((f_start, f_end))
+ frame_sequences.append(result)
- 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)
+ return frame_sequences
######################################################################
- def generate_seq_old(
- nb,
- height,
- width,
- nb_birds=3,
- nb_iterations=2,
- ):
- 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))
-
- 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[i, j] = c
- f_start[i - vi, j - vj] = c
- f_start[i + vj, j - vi] = c
- f_start[i - vj, j + vi] = c
+ def generate_prompts_and_answers(self, nb):
+ frame_sequences = self.generate_frame_sequences(nb)
+ prompts = frame_sequences[:, : frame_sequences.size(0) // 2].flatten(1)
+ answers = frame_sequences[:, frame_sequences.size(0) // 2 :].flatten(1)
+ return prompts, answers
- for l in range(nb_iterations):
- i += vi
- j += vj
- if i < 0 or i >= height or j < 0 or j >= width:
- i -= vi
- j -= vj
- vi, vj = -vi, -vj
- i += vi
- j += vj
-
- 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 generate_token_sequences(self, nb):
+ frame_sequences = self.generate_frame_sequences(nb)
result = []
- for p in pairs:
+
+ for frame_sequence in frame_sequences:
+ a = []
if torch.rand(1) < 0.5:
- result.append(
- torch.cat(
- [p[0].flatten(), torch.tensor([token_forward]), p[1].flatten()],
- dim=0,
- )[None, :]
- )
+ for frame in frame_sequence:
+ if len(a) > 0:
+ a.append(torch.tensor([self.token_forward]))
+ a.append(frame.flatten())
else:
- result.append(
- torch.cat(
- [
- p[1].flatten(),
- torch.tensor([token_backward]),
- p[0].flatten(),
- ],
- dim=0,
- )[None, :]
- )
+ for frame in reversed(frame_sequence):
+ if len(a) > 0:
+ a.append(torch.tensor([self.token_backward]))
+ a.append(frame.flatten())
+
+ result.append(torch.cat(a, dim=0)[None, :])
return torch.cat(result, dim=0)
- 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)
+ ######################################################################
+
+ def frame2img(self, x, scale=15):
+ x = x.reshape(-1, self.height, self.width)
+ m = torch.logical_and(
+ x >= 0, x < self.first_bird_token + self.nb_bird_tokens
+ ).long()
+ x = self.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)
+ x = x[:, :, :, None, :, None].expand(-1, -1, -1, scale, -1, scale)
+ x = x.reshape(s[0], s[1], s[2] * scale, s[3] * scale)
- x[:, :, :, torch.arange(0, x.size(3), upscale)] = 0
- x[:, :, torch.arange(0, x.size(2), upscale), :] = 0
+ x[:, :, :, torch.arange(0, x.size(3), scale)] = 0
+ x[:, :, torch.arange(0, x.size(2), scale), :] = 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
+ for k in range(2, scale - 2):
+ for l in [0, 1]:
+ x[n, :, i * scale + k, j * scale + k - l] = 0
+ x[
+ n, :, i * scale + scale - 1 - k, j * scale + k - l
+ ] = 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(
- [
- frame2img(f_first, height, width, upscale),
+ def seq2img(self, seq, scale=15):
+ all = [
+ self.frame2img(
+ seq[:, : self.height * self.width].reshape(-1, self.height, self.width),
+ scale,
+ )
+ ]
+
+ separator = torch.full((seq.size(0), 3, self.height * scale - 1, 1), 0)
+
+ t = self.height * self.width
+
+ while t < seq.size(1):
+ direction_tokens = seq[:, t]
+ t += 1
+
+ direction_images = self.colors[
+ torch.full(
+ (direction_tokens.size(0), self.height * scale - 1, scale), 0
+ )
+ ].permute(0, 3, 1, 2)
+
+ for n in range(direction_tokens.size(0)):
+ if direction_tokens[n] == self.token_forward:
+ for k in range(scale):
+ for l in [0, 1]:
+ direction_images[
+ n,
+ :,
+ (self.height * scale) // 2 - scale // 2 + k - l,
+ 3 + scale // 2 - abs(k - scale // 2),
+ ] = 0
+ elif direction_tokens[n] == self.token_backward:
+ for k in range(scale):
+ for l in [0, 1]:
+ direction_images[
+ n,
+ :,
+ (self.height * scale) // 2 - scale // 2 + k - l,
+ 3 + abs(k - scale // 2),
+ ] = 0
+ else:
+ for k in range(2, scale - 2):
+ for l in [0, 1]:
+ direction_images[
+ n,
+ :,
+ (self.height * scale) // 2 - scale // 2 + k - l,
+ k,
+ ] = 0
+ direction_images[
+ n,
+ :,
+ (self.height * scale) // 2 - scale // 2 + k - l,
+ scale - 1 - k,
+ ] = 0
+
+ all += [
separator,
- direction_symbol,
+ direction_images,
separator,
- frame2img(f_second, height, width, upscale),
- ],
- dim=3,
- )
+ self.frame2img(
+ seq[:, t : t + self.height * self.width].reshape(
+ -1, self.height, self.width
+ ),
+ scale,
+ ),
+ ]
- def seq2str(seq):
+ t += self.height * self.width
+
+ return torch.cat(all, dim=3)
+
+ def seq2str(self, seq):
result = []
for s in seq:
- result.append("".join([token2char[v] for v in s]))
+ result.append("".join([self.token2char[v] for v in s]))
return result
+ def save_image(self, input, result_dir, filename):
+ img = self.seq2img(input.to("cpu"))
+ image_name = os.path.join(result_dir, filename)
+ torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
+
+ def save_quizzes(self, input, result_dir, filename_prefix):
+ self.save_image(input, result_dir, filename_prefix + ".png")
+
######################################################################
if __name__ == "__main__":
import time
- height, width = 6, 8
+ sky = Sky(height=6, width=8, speed=4, nb_iterations=2)
+
start_time = time.perf_counter()
- seq, it = generate_seq(
- nb=64, height=height, width=width, nb_iterations=100, return_iterations=True
- )
+ token_sequences = sky.generate_token_sequences(nb=64)
delay = time.perf_counter() - start_time
- print(f"{seq.size(0)/delay:02f} samples/s")
+ print(f"{token_sequences.size(0)/delay:02f} seq/s")
- print(seq2str(seq[:4]))
+ # print(sky.seq2str(seq[:4]))
- 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,
- )
+ # for t in range(len(it[0])):
+ # img = torch.cat([sky.frame2img(f[t]) 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())
+ # print(seq.size())
+ img = sky.seq2img(token_sequences)
+ # print(img.size())
torchvision.utils.save_image(
img.float() / 255.0, "/tmp/world.png", nrow=6, padding=6, pad_value=0