#!/usr/bin/env python
-import math
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
+import math, sys, tqdm
import torch, torchvision
from torch import nn
from torch.nn import functional as F
-import cairo
-
-
-class Box:
- def __init__(self, x, y, w, h, r, g, b):
- self.x = x
- self.y = y
- self.w = w
- self.h = h
- self.r = r
- self.g = g
- self.b = b
-
- def collision(self, scene):
- for c in scene:
- if (
- self is not c
- and max(self.x, c.x) <= min(self.x + self.w, c.x + c.w)
- and max(self.y, c.y) <= min(self.y + self.h, c.y + c.h)
- ):
- return True
- return False
-
-
-def scene2tensor(xh, yh, scene, size=64):
- 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, b.g, b.b, 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).float() / 255
-
-
-def random_scene():
- scene = []
- colors = [
- (1.00, 0.00, 0.00),
- (0.00, 1.00, 0.00),
- (0.60, 0.60, 1.00),
- (1.00, 1.00, 0.00),
- (0.75, 0.75, 0.75),
- ]
-
- for k in range(10):
- wh = torch.rand(2) * 0.2 + 0.2
- xy = torch.rand(2) * (1 - wh)
- c = colors[torch.randint(len(colors), (1,))]
- b = Box(
- xy[0].item(), xy[1].item(), wh[0].item(), wh[1].item(), c[0], c[1], c[2]
- )
- if not b.collision(scene):
- scene.append(b)
-
- return scene
-
-
-def sequence(nb_steps=10, all_frames=False):
- delta = 0.1
- effects = [
- (False, 0, 0),
- (False, delta, 0),
- (False, 0, delta),
- (False, -delta, 0),
- (False, 0, -delta),
- (True, delta, 0),
- (True, 0, delta),
- (True, -delta, 0),
- (True, 0, -delta),
- ]
-
- while True:
- frames = []
-
- scene = random_scene()
- xh, yh = tuple(x.item() for x in torch.rand(2))
-
- frames.append(scene2tensor(xh, yh, scene))
-
- actions = torch.randint(len(effects), (nb_steps,))
- change = False
-
- for a in actions:
- g, dx, dy = effects[a]
- if g:
- for b in scene:
- if b.x <= xh and b.x + b.w >= xh and b.y <= yh and b.y + b.h >= yh:
- x, y = b.x, b.y
- b.x += dx
- b.y += dy
- if (
- b.x < 0
- or b.y < 0
- or b.x + b.w > 1
- or b.y + b.h > 1
- or b.collision(scene)
- ):
- b.x, b.y = x, y
- else:
- xh += dx
- yh += dy
- change = True
- else:
- x, y = xh, yh
- xh += dx
- yh += dy
- if xh < 0 or xh > 1 or yh < 0 or yh > 1:
- xh, yh = x, y
-
- if all_frames:
- frames.append(scene2tensor(xh, yh, scene))
-
- if not all_frames:
- frames.append(scene2tensor(xh, yh, scene))
-
- if change:
- break
-
- return frames, actions
-
######################################################################
-# ||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
+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)]) + "><"
+
+
+def generate_seq(
+ nb,
+ height,
+ width,
+ nb_birds=3,
+ nb_iterations=2,
+):
+ pairs = []
+
+ for _ in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world generation"):
+ while True:
+ 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),
+ )
- 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)
+ 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):
+ 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
+
+ if nb_collisions == 0:
+ break
+
+ 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:
- 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)
+ result.append(
+ torch.cat(
+ [p[1].flatten(), torch.tensor([token_backward]), p[0].flatten()],
+ dim=0,
+ )[None, :]
+ )
- c = x[torch.randperm(x.size(0))[:nb_centroids]]
- t = torch.full((x.size(0),), -1)
- n = 0
+ return torch.cat(result, dim=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 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))
-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
+ 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,
)
- .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
+ 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
+
+ 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))
+
+ 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, :]
)
- .permute(0, 3, 1, 4, 2, 5)
- .reshape(invert_size)
- )
-
-
-def train_encoder(input, device=torch.device("cpu")):
- class SomeLeNet(nn.Module):
- def __init__(self):
- super().__init__()
- self.conv1 = nn.Conv2d(1, 32, kernel_size=5)
- self.conv2 = nn.Conv2d(32, 64, kernel_size=5)
- self.fc1 = nn.Linear(256, 200)
- self.fc2 = nn.Linear(200, 10)
-
- def forward(self, x):
- x = F.relu(F.max_pool2d(self.conv1(x), kernel_size=3))
- x = F.relu(F.max_pool2d(self.conv2(x), kernel_size=2))
- x = x.view(x.size(0), -1)
- x = F.relu(self.fc1(x))
- x = self.fc2(x)
- return x
-
- ######################################################################
-
- model = SomeLeNet()
- nb_parameters = sum(p.numel() for p in model.parameters())
+ 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_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)
+
+ 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
+
+ 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(
+ [
+ mosaic(f_first, upscale),
+ separator,
+ direction_symbol,
+ separator,
+ mosaic(f_second, upscale),
+ ],
+ dim=3,
+ )
- print(f"nb_parameters {nb_parameters}")
- optimizer = torch.optim.SGD(model.parameters(), lr=lr)
- criterion = nn.CrossEntropyLoss()
+def seq2str(seq):
+ result = []
+ for s in seq:
+ result.append("".join([token2char[v] for v in s]))
+ return result
- model.to(device)
- criterion.to(device)
- train_input, train_targets = train_input.to(device), train_targets.to(device)
- test_input, test_targets = test_input.to(device), test_targets.to(device)
+######################################################################
- mu, std = train_input.mean(), train_input.std()
- train_input.sub_(mu).div_(std)
- test_input.sub_(mu).div_(std)
+if __name__ == "__main__":
+ import time
+ height, width = 6, 8
start_time = time.perf_counter()
+ seq = generate_seq(nb=90, height=height, width=width)
+ delay = time.perf_counter() - start_time
+ print(f"{seq.size(0)/delay:02f} samples/s")
- for k in range(nb_epochs):
- acc_loss = 0.0
-
- for input, targets in zip(
- train_input.split(batch_size), train_targets.split(batch_size)
- ):
- output = model(input)
- loss = criterion(output, targets)
- acc_loss += loss.item()
-
- optimizer.zero_grad()
- loss.backward()
- optimizer.step()
-
- nb_test_errors = 0
- for input, targets in zip(
- test_input.split(batch_size), test_targets.split(batch_size)
- ):
- wta = model(input).argmax(1)
- nb_test_errors += (wta != targets).long().sum()
- test_error = nb_test_errors / test_input.size(0)
- duration = time.perf_counter() - start_time
-
- print(f"loss {k} {duration:.02f}s {acc_loss:.02f} {test_error*100:.02f}%")
+ 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())
-if __name__ == "__main__":
- import time
-
- all_frames = []
- nb = 1000
- start_time = time.perf_counter()
- for n in range(nb):
- frames, actions = sequence(nb_steps=31)
- all_frames += frames
- end_time = time.perf_counter()
- print(f"{nb / (end_time - start_time):.02f} samples per second")
-
- input = torch.cat(all_frames, 0)
-
- # x = patchify(input, 8)
- # y = x.reshape(x.size(0), -1)
- # print(f"{x.size()=} {y.size()=}")
- # centroids, t = kmeans(y, 4096)
- # results = centroids[t]
- # results = results.reshape(x.size())
- # results = patchify(results, 8, input.size())
-
- print(f"{input.size()=} {results.size()=}")
-
- torchvision.utils.save_image(input[:64], "orig.png", nrow=8)
- torchvision.utils.save_image(results[:64], "qtiz.png", nrow=8)
-
- # frames, actions = sequence(nb_steps=31, all_frames=True)
- # frames = torch.cat(frames, 0)
- # torchvision.utils.save_image(frames, "seq.png", nrow=8)
+ torchvision.utils.save_image(
+ img.float() / 255.0, "/tmp/world.png", nrow=6, padding=6, pad_value=0
+ )