5 import torch, torchvision
8 from torch.nn import functional as F
11 ######################################################################
17 def __init__(self, x, y, w, h, r, g, b):
26 def collision(self, scene):
30 and max(self.x, c.x) <= min(self.x + self.w, c.x + c.w)
31 and max(self.y, c.y) <= min(self.y + self.h, c.y + c.h)
37 ######################################################################
40 class Normalizer(nn.Module):
41 def __init__(self, mu, std):
43 self.register_buffer("mu", mu)
44 self.register_buffer("log_var", 2 * torch.log(std))
47 return (x - self.mu) / torch.exp(self.log_var / 2.0)
50 class SignSTE(nn.Module):
55 # torch.sign() takes three values
56 s = (x >= 0).float() * 2 - 1
60 return s + u - u.detach()
65 def loss_H(binary_logits, h_threshold=1):
66 p = binary_logits.sigmoid().mean(0)
67 h = (-p.xlogy(p) - (1 - p).xlogy(1 - p)) / math.log(2)
68 h.clamp_(max=h_threshold)
69 return h_threshold - h.mean()
84 device=torch.device("cpu"),
87 logger = lambda s: print(s)
89 mu, std = train_input.float().mean(), train_input.float().std()
91 def encoder_core(depth, dim):
95 dim * 2**k, dim * 2**k, kernel_size=5, stride=1, padding=2
98 nn.Conv2d(dim * 2**k, dim * 2 ** (k + 1), kernel_size=2, stride=2),
101 for k in range(depth)
104 return nn.Sequential(*[x for m in l for x in m])
106 def decoder_core(depth, dim):
110 dim * 2 ** (k + 1), dim * 2**k, kernel_size=2, stride=2
114 dim * 2**k, dim * 2**k, kernel_size=5, stride=1, padding=2
118 for k in range(depth - 1, -1, -1)
121 return nn.Sequential(*[x for m in l for x in m])
123 encoder = nn.Sequential(
125 nn.Conv2d(3, dim_hidden, kernel_size=1, stride=1),
128 encoder_core(depth=depth, dim=dim_hidden),
130 nn.Conv2d(dim_hidden * 2**depth, nb_bits_per_token, kernel_size=1, stride=1),
133 quantizer = SignSTE()
135 decoder = nn.Sequential(
136 nn.Conv2d(nb_bits_per_token, dim_hidden * 2**depth, kernel_size=1, stride=1),
138 decoder_core(depth=depth, dim=dim_hidden),
140 nn.ConvTranspose2d(dim_hidden, 3 * Box.nb_rgb_levels, kernel_size=1, stride=1),
143 model = nn.Sequential(encoder, decoder)
145 nb_parameters = sum(p.numel() for p in model.parameters())
147 logger(f"nb_parameters {nb_parameters}")
151 for k in range(nb_epochs):
153 math.log(lr_start) + math.log(lr_end / lr_start) / (nb_epochs - 1) * k
155 optimizer = torch.optim.Adam(model.parameters(), lr=lr)
159 for input in tqdm.tqdm(train_input.split(batch_size), desc="vqae-train"):
160 input = input.to(device)
162 zq = z if k < 2 else quantizer(z)
165 output = output.reshape(
166 output.size(0), -1, 3, output.size(2), output.size(3)
169 train_loss = F.cross_entropy(output, input)
171 if lambda_entropy > 0:
172 train_loss = train_loss + lambda_entropy * loss_H(z, h_threshold=0.5)
174 acc_train_loss += train_loss.item() * input.size(0)
176 optimizer.zero_grad()
177 train_loss.backward()
182 for input in tqdm.tqdm(test_input.split(batch_size), desc="vqae-test"):
183 input = input.to(device)
185 zq = z if k < 1 else quantizer(z)
188 output = output.reshape(
189 output.size(0), -1, 3, output.size(2), output.size(3)
192 test_loss = F.cross_entropy(output, input)
194 acc_test_loss += test_loss.item() * input.size(0)
196 train_loss = acc_train_loss / train_input.size(0)
197 test_loss = acc_test_loss / test_input.size(0)
199 logger(f"train_ae {k} lr {lr} train_loss {train_loss} test_loss {test_loss}")
202 return encoder, quantizer, decoder
205 ######################################################################
208 def scene2tensor(xh, yh, scene, size):
209 width, height = size, size
210 pixel_map = torch.ByteTensor(width, height, 4).fill_(255)
211 data = pixel_map.numpy()
212 surface = cairo.ImageSurface.create_for_data(
213 data, cairo.FORMAT_ARGB32, width, height
216 ctx = cairo.Context(surface)
217 ctx.set_fill_rule(cairo.FILL_RULE_EVEN_ODD)
220 ctx.move_to(b.x * size, b.y * size)
221 ctx.rel_line_to(b.w * size, 0)
222 ctx.rel_line_to(0, b.h * size)
223 ctx.rel_line_to(-b.w * size, 0)
226 b.r / (Box.nb_rgb_levels - 1),
227 b.g / (Box.nb_rgb_levels - 1),
228 b.b / (Box.nb_rgb_levels - 1),
234 ctx.set_source_rgba(0.0, 0.0, 0.0, 1.0)
235 ctx.move_to(xh * size - hs / 2, yh * size - hs / 2)
236 ctx.rel_line_to(hs, 0)
237 ctx.rel_line_to(0, hs)
238 ctx.rel_line_to(-hs, 0)
243 pixel_map[None, :, :, :3]
247 .mul(Box.nb_rgb_levels)
252 def random_scene(nb_insert_attempts=3):
255 ((Box.nb_rgb_levels - 1), 0, 0),
256 (0, (Box.nb_rgb_levels - 1), 0),
257 (0, 0, (Box.nb_rgb_levels - 1)),
258 ((Box.nb_rgb_levels - 1), (Box.nb_rgb_levels - 1), 0),
260 (Box.nb_rgb_levels * 2) // 3,
261 (Box.nb_rgb_levels * 2) // 3,
262 (Box.nb_rgb_levels * 2) // 3,
266 for k in range(nb_insert_attempts):
267 wh = torch.rand(2) * 0.2 + 0.2
268 xy = torch.rand(2) * (1 - wh)
269 c = colors[torch.randint(len(colors), (1,))]
271 xy[0].item(), xy[1].item(), wh[0].item(), wh[1].item(), c[0], c[1], c[2]
273 if not b.collision(scene):
279 def generate_episode(steps, size=64):
296 scene = random_scene()
297 xh, yh = tuple(x.item() for x in torch.rand(2))
299 actions = torch.randint(len(effects), (len(steps),))
302 for s, a in zip(steps, actions):
304 frames.append(scene2tensor(xh, yh, scene, size=size))
306 grasp, dx, dy = effects[a]
310 if b.x <= xh and b.x + b.w >= xh and b.y <= yh and b.y + b.h >= yh:
319 or b.collision(scene)
330 if xh < 0 or xh > 1 or yh < 0 or yh > 1:
333 if nb_changes > len(steps) // 3:
336 return frames, actions
339 ######################################################################
342 def generate_episodes(nb, steps):
343 all_frames, all_actions = [], []
344 for n in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world-data"):
345 frames, actions = generate_episode(steps)
347 all_actions += [actions[None, :]]
348 return torch.cat(all_frames, 0).contiguous(), torch.cat(all_actions, 0)
351 def create_data_and_processors(
357 device=torch.device("cpu"),
358 device_storage=torch.device("cpu"),
361 assert mode in ["first_last"]
363 if mode == "first_last":
364 steps = [True] + [False] * (nb_steps + 1) + [True]
366 train_input, train_actions = generate_episodes(nb_train_samples, steps)
367 train_input, train_actions = train_input.to(device_storage), train_actions.to(
370 test_input, test_actions = generate_episodes(nb_test_samples, steps)
371 test_input, test_actions = test_input.to(device_storage), test_actions.to(
375 encoder, quantizer, decoder = train_encoder(
384 quantizer.train(False)
387 z = encoder(train_input[:1].to(device))
388 pow2 = (2 ** torch.arange(z.size(1), device=device))[None, None, :]
389 z_h, z_w = z.size(2), z.size(3)
391 def frame2seq(input, batch_size=25):
394 for x in input.split(batch_size):
397 ze_bool = (quantizer(z) >= 0).long()
399 ze_bool.permute(0, 2, 3, 1).reshape(
400 ze_bool.size(0), -1, ze_bool.size(1)
407 return torch.cat(seq, dim=0)
409 def seq2frame(input, batch_size=25, T=1e-2):
412 for seq in input.split(batch_size):
414 zd_bool = (seq[:, :, None] // p) % 2
415 zd_bool = zd_bool.reshape(zd_bool.size(0), z_h, z_w, -1).permute(0, 3, 1, 2)
416 logits = decoder(zd_bool * 2.0 - 1.0)
417 logits = logits.reshape(
418 logits.size(0), -1, 3, logits.size(2), logits.size(3)
419 ).permute(0, 2, 3, 4, 1)
420 output = torch.distributions.categorical.Categorical(
424 frames.append(output)
426 return torch.cat(frames, dim=0)
428 return train_input, train_actions, test_input, test_actions, frame2seq, seq2frame
431 ######################################################################
433 if __name__ == "__main__":
441 ) = create_data_and_processors(
448 input = test_input[:256]
450 seq = frame2seq(input)
451 output = seq2frame(seq)
453 torchvision.utils.save_image(
454 input.float() / (Box.nb_rgb_levels - 1), "orig.png", nrow=16
457 torchvision.utils.save_image(
458 output.float() / (Box.nb_rgb_levels - 1), "qtiz.png", nrow=16