#!/usr/bin/env python
+# 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:
- nb_rgb_levels = 10
-
- 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
-
######################################################################
-class Normalizer(nn.Module):
- def __init__(self, mu, std):
- super().__init__()
- self.register_buffer("mu", mu)
- self.register_buffer("log_var", 2 * torch.log(std))
-
- def forward(self, x):
- return (x - self.mu) / torch.exp(self.log_var / 2.0)
+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
-class SignSTE(nn.Module):
- def __init__(self):
- super().__init__()
+token2char = "_" + "".join([chr(ord("A") + n) for n in range(len(colors) - 1)]) + "><"
- def forward(self, x):
- # torch.sign() takes three values
- s = (x >= 0).float() * 2 - 1
- if self.training:
- u = torch.tanh(x)
- return s + u - u.detach()
- else:
- return s
-
-
-def loss_H(binary_logits, h_threshold=1):
- p = binary_logits.sigmoid().mean(0)
- h = (-p.xlogy(p) - (1 - p).xlogy(1 - p)) / math.log(2)
- h.clamp_(max=h_threshold)
- return h_threshold - h.mean()
-
-
-def train_encoder(
- train_input,
- test_input,
- depth=2,
- dim_hidden=48,
- nb_bits_per_token=8,
- lambda_entropy=0.0,
- lr_start=1e-3,
- lr_end=1e-4,
- nb_epochs=10,
- batch_size=25,
- logger=None,
- device=torch.device("cpu"),
+def generate_seq(
+ nb,
+ height,
+ width,
+ nb_birds=3,
+ nb_iterations=2,
):
- if logger is None:
- logger = lambda s: print(s)
-
- 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),
- )
+ pairs = []
- model = nn.Sequential(encoder, decoder)
+ for _ in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world generation"):
+ while True:
+ f_start = torch.zeros(height, width, dtype=torch.int64)
- nb_parameters = sum(p.numel() for p in model.parameters())
-
- logger(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"):
- input = input.to(device)
- z = encoder(input)
- zq = z if k < 2 else quantizer(z)
- output = decoder(zq)
-
- output = output.reshape(
- output.size(0), -1, 3, output.size(2), output.size(3)
+ 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),
)
- train_loss = F.cross_entropy(output, input)
-
- if lambda_entropy > 0:
- loss = loss + lambda_entropy * loss_H(z, h_threshold=0.5)
-
- 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"):
- input = input.to(device)
- 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)
+ 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:
+ result.append(
+ torch.cat(
+ [p[1].flatten(), torch.tensor([token_backward]), p[0].flatten()],
+ dim=0,
+ )[None, :]
)
- 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)
-
- logger(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)
- )
-
-
-def random_scene(nb_insert_attempts=3):
- scene = []
- colors = [
- ((Box.nb_rgb_levels - 1), 0, 0),
- (0, (Box.nb_rgb_levels - 1), 0),
- (0, 0, (Box.nb_rgb_levels - 1)),
- ((Box.nb_rgb_levels - 1), (Box.nb_rgb_levels - 1), 0),
- (
- (Box.nb_rgb_levels * 2) // 3,
- (Box.nb_rgb_levels * 2) // 3,
- (Box.nb_rgb_levels * 2) // 3,
- ),
- ]
-
- for k in range(nb_insert_attempts):
- 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 generate_episode(steps, size=64):
- 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))
-
- actions = torch.randint(len(effects), (len(steps),))
- nb_changes = 0
-
- for s, a in zip(steps, actions):
- if s:
- frames.append(scene2tensor(xh, yh, scene, size=size))
-
- grasp, dx, dy = effects[a]
-
- if grasp:
- 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
- nb_changes += 1
- 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 nb_changes > len(steps) // 3:
- break
-
- return frames, actions
+ return torch.cat(result, dim=0)
######################################################################
-def generate_episodes(nb, steps):
- all_frames, all_actions = [], []
- for n in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world-data"):
- frames, actions = generate_episode(steps)
- all_frames += frames
- all_actions += [actions[None, :]]
- return torch.cat(all_frames, 0).contiguous(), torch.cat(all_actions, 0)
-
-
-def create_data_and_processors(
- nb_train_samples,
- nb_test_samples,
- mode,
- nb_steps,
- nb_epochs=10,
- device=torch.device("cpu"),
- device_storage=torch.device("cpu"),
- logger=None,
+def generate_seq_(
+ nb,
+ height,
+ width,
+ nb_birds=3,
+ nb_iterations=2,
):
- assert mode in ["first_last"]
-
- if mode == "first_last":
- steps = [True] + [False] * (nb_steps + 1) + [True]
+ 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
+
+ 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, :]
+ )
- train_input, train_actions = generate_episodes(nb_train_samples, steps)
- train_input, train_actions = train_input.to(device_storage), train_actions.to(
- device_storage
- )
- test_input, test_actions = generate_episodes(nb_test_samples, steps)
- test_input, test_actions = test_input.to(device_storage), test_actions.to(
- device_storage
+ 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,
)
- encoder, quantizer, decoder = train_encoder(
- train_input,
- test_input,
- lambda_entropy=1.0,
- nb_epochs=nb_epochs,
- logger=logger,
- device=device,
- )
- encoder.train(False)
- quantizer.train(False)
- decoder.train(False)
-
- z = encoder(train_input[:1].to(device))
- pow2 = (2 ** torch.arange(z.size(1), device=device))[None, None, :]
- z_h, z_w = z.size(2), z.size(3)
-
- def frame2seq(input, batch_size=25):
- seq = []
- p = pow2.to(device)
- for x in input.split(batch_size):
- x = x.to(device)
- z = encoder(x)
- ze_bool = (quantizer(z) >= 0).long()
- output = (
- ze_bool.permute(0, 2, 3, 1).reshape(
- ze_bool.size(0), -1, ze_bool.size(1)
- )
- * p
- ).sum(-1)
-
- seq.append(output)
-
- return torch.cat(seq, dim=0)
-
- def seq2frame(input, batch_size=25, T=1e-2):
- frames = []
- p = pow2.to(device)
- for seq in input.split(batch_size):
- seq = seq.to(device)
- zd_bool = (seq[:, :, None] // p) % 2
- zd_bool = zd_bool.reshape(zd_bool.size(0), z_h, z_w, -1).permute(0, 3, 1, 2)
- logits = decoder(zd_bool * 2.0 - 1.0)
- logits = logits.reshape(
- logits.size(0), -1, 3, logits.size(2), logits.size(3)
- ).permute(0, 2, 3, 4, 1)
- output = torch.distributions.categorical.Categorical(
- logits=logits / T
- ).sample()
-
- frames.append(output)
-
- return torch.cat(frames, dim=0)
-
- return train_input, train_actions, test_input, test_actions, frame2seq, seq2frame
+
+def seq2str(seq):
+ result = []
+ for s in seq:
+ result.append("".join([token2char[v] for v in s]))
+ return result
######################################################################
if __name__ == "__main__":
- (
- train_input,
- train_actions,
- test_input,
- test_actions,
- frame2seq,
- seq2frame,
- ) = create_data_and_processors(
- # 10000, 1000,
- 100,
- 100,
- nb_epochs=2,
- mode="first_last",
- nb_steps=20,
- )
+ import time
- input = test_input[:64]
+ 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")
- seq = frame2seq(input)
+ print(seq2str(seq[:4]))
- print(f"{seq.size()=} {seq.dtype=} {seq.min()=} {seq.max()=}")
+ # m = (torch.rand(seq.size()) < 0.05).long()
+ # seq = (1 - m) * seq + m * 23
- output = seq2frame(seq)
-
- torchvision.utils.save_image(
- input.float() / (Box.nb_rgb_levels - 1), "orig.png", nrow=8
- )
+ img = sample2img(seq, height, width)
+ print(img.size())
torchvision.utils.save_image(
- output.float() / (Box.nb_rgb_levels - 1), "qtiz.png", nrow=8
+ img.float() / 255.0, "/tmp/world.png", nrow=6, padding=6, pad_value=0
)