- 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():
- 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(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),))
- change = False
-
- for s, a in zip(steps, actions):
- if s:
- frames.append(scene2tensor(xh, yh, scene, size=size))
-
- 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 change:
- break
-
- return frames, actions
-
-
-######################################################################
-
-
-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"),
- logger=None,
-):
- assert mode in ["first_last"]
-
- if mode == "first_last":
- steps = [True] + [False] * (nb_steps + 1) + [True]
-
- train_input, train_actions = generate_episodes(nb_train_samples, steps)
- train_input, train_actions = train_input.to(device), train_actions.to(device)
- test_input, test_actions = generate_episodes(nb_test_samples, steps)
- test_input, test_actions = test_input.to(device), test_actions.to(device)
-
- encoder, quantizer, decoder = train_encoder(
- train_input, test_input, nb_epochs=nb_epochs, logger=logger, device=device
+ 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,