#!/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
-
-
-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():
- 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(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 generate_episode(nb_steps=10, 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))
-
- frames.append(scene2tensor(xh, yh, scene, size=size))
-
- 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
-
- frames.append(scene2tensor(xh, yh, scene, size=size))
-
- 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
-
- 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)
- 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)
-
- c = x[torch.randperm(x.size(0))[:nb_centroids]]
- t = torch.full((x.size(0),), -1)
- n = 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
+colors = torch.tensor(
+ [
+ [255, 255, 255],
+ [255, 0, 0],
+ [0, 128, 0],
+ [0, 0, 255],
+ [255, 200, 0],
+ [192, 192, 192],
+ ]
+)
+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 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
+def generate(
+ nb,
+ height,
+ width,
+ max_nb_obj=2,
+ 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))
+
+ 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,
)
- .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, :]
)
- .permute(0, 3, 1, 4, 2, 5)
- .reshape(invert_size)
- )
-
-
-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)
-
-
-class SignSTE(nn.Module):
- def __init__(self):
- super().__init__()
-
- 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 train_encoder(
- train_input,
- test_input,
- depth=2,
- dim_hidden=48,
- nb_bits_per_token=10,
- lr_start=1e-3,
- lr_end=1e-4,
- nb_epochs=10,
- batch_size=25,
- device=torch.device("cpu"),
-):
- 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),
- )
-
- model = nn.Sequential(encoder, decoder)
-
- nb_parameters = sum(p.numel() for p in model.parameters())
-
- print(f"nb_parameters {nb_parameters}")
-
- model.to(device)
-
- g5x5 = torch.exp(-torch.tensor([[-2.0, -1.0, 0.0, 1.0, 2.0]]) ** 2 / 2)
- g5x5 = (g5x5.t() @ g5x5).view(1, 1, 5, 5)
- g5x5 = g5x5 / g5x5.sum()
-
- 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 train_input.split(batch_size):
- 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)
- )
-
- train_loss = F.cross_entropy(output, input)
-
- 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 test_input.split(batch_size):
- 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)
+ 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)
-
- print(f"train_ae {k} lr {lr} train_loss {train_loss} test_loss {test_loss}")
- sys.stdout.flush()
-
- return encoder, quantizer, decoder
-
-def generate_episodes(nb):
- all_frames = []
- for n in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world-data"):
- frames, actions = generate_episode(nb_steps=31)
- all_frames += [ frames[0], frames[-1] ]
- return torch.cat(all_frames, 0).contiguous()
-
-def create_data_and_processors(nb_train_samples, nb_test_samples):
- train_input = generate_episodes(nb_train_samples)
- test_input = generate_episodes(nb_test_samples)
- encoder, quantizer, decoder = train_encoder(train_input, test_input, nb_epochs=2)
+ 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
+
+ return torch.cat(
+ [
+ mosaic(f_first, upscale),
+ separator,
+ direction_symbol,
+ separator,
+ mosaic(f_second, upscale),
+ ],
+ dim=3,
+ )
- input = test_input[:64]
- z = encoder(input.float())
- height, width = z.size(2), z.size(3)
- zq = quantizer(z).long()
- pow2=(2**torch.arange(zq.size(1), device=zq.device))[None,None,:]
- seq = (zq.permute(0,2,3,1).clamp(min=0).reshape(zq.size(0),-1,zq.size(1)) * pow2).sum(-1)
- print(f"{seq.size()=}")
+def seq2str(seq):
+ result = []
+ for s in seq:
+ result.append("".join([token2char[v] for v in s]))
+ return result
- ZZ=zq
- zq = ((seq[:,:,None] // pow2)%2)*2-1
- zq = zq.reshape(zq.size(0), height, width, -1).permute(0,3,1,2)
+######################################################################
- print(ZZ[0])
- print(zq[0])
+if __name__ == "__main__":
+ import time
- print("CHECK", (ZZ-zq).abs().sum())
+ height, width = 6, 8
+ start_time = time.perf_counter()
+ 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")
- results = decoder(zq.float())
- T = 0.1
- results = results.reshape(
- results.size(0), -1, 3, results.size(2), results.size(3)
- ).permute(0, 2, 3, 4, 1)
- results = torch.distributions.categorical.Categorical(logits=results / T).sample()
+ print(seq2str(seq[:4]))
+ m = (torch.rand(seq.size()) < 0.05).long()
+ seq = (1 - m) * seq + m * 23
- 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(
- results.float() / (Box.nb_rgb_levels - 1), "qtiz.png", nrow=8
+ img.float() / 255.0, "/tmp/world.png", nrow=6, padding=4
)
-
-
-######################################################################
-
-if __name__ == "__main__":
- create_data_and_processors(250,100)
-
- # train_input = generate_episodes(2500)
- # test_input = generate_episodes(1000)
-
- # encoder, quantizer, decoder = train_encoder(train_input, test_input)
-
- # input = test_input[torch.randperm(test_input.size(0))[:64]]
- # z = encoder(input.float())
- # zq = quantizer(z)
- # results = decoder(zq)
-
- # T = 0.1
- # results = torch.distributions.categorical.Categorical(logits=results / T).sample()