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
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)
+
+
+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
+
+class DiscreteSampler2d(nn.Module):
+ def __init__(self):
+ super().__init__()
+
+ def forward(self, x):
+ s = (x >= x.max(-3,keepdim=True).values).float()
+
+ if self.training:
+ u = x.softmax(dim=-3)
+ 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,
+ nb_bits_per_token,
+ dim_hidden=48,
+ lambda_entropy=0.0,
+ lr_start=1e-3,
+ lr_end=1e-4,
+ nb_epochs=10,
+ batch_size=25,
+ logger=None,
+ device=torch.device("cpu"),
+):
+ 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),
+ )
+
+ model = nn.Sequential(encoder, decoder)
+
+ 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 = 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)
+
+ if lambda_entropy > 0:
+ train_loss = train_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 = quantizer(z)
+ output = decoder(zq)
+
+ output = output.reshape(
+ output.size(0), -1, 3, output.size(2), output.size(3)
+ )
+
+ 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)
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.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.close_path()
ctx.fill()
- return pixel_map[None, :, :, :3].flip(-1).permute(0, 3, 1, 2).float() / 255
+ return (
+ pixel_map[None, :, :, :3]
+ .flip(-1)
+ .permute(0, 3, 1, 2)
+ .long()
+ .mul(Box.nb_rgb_levels)
+ .floor_divide(256)
+ )
-def random_scene():
+def random_scene(nb_insert_attempts=3):
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),
+ ((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):
+ 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,))]
return scene
-def generate_sequence(nb_steps=10, all_frames=False, size=64):
+def generate_episode(steps, size=64):
delta = 0.1
effects = [
(False, 0, 0),
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), (len(steps),))
+ nb_changes = 0
+
+ for s, a in zip(steps, actions):
+ if s:
+ frames.append(scene2tensor(xh, yh, scene, size=size))
- actions = torch.randint(len(effects), (nb_steps,))
- change = False
+ grasp, dx, dy = effects[a]
- for a in actions:
- g, dx, dy = effects[a]
- if g:
+ 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
else:
xh += dx
yh += dy
- change = True
+ nb_changes += 1
else:
x, y = xh, yh
xh += dx
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, size=size))
-
- if not all_frames:
- frames.append(scene2tensor(xh, yh, scene, size=size))
-
- if change:
+ if nb_changes > len(steps) // 3:
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
-
-
-######################################################################
-
-
-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
- )
- .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
- )
- .permute(0, 3, 1, 4, 2, 5)
- .reshape(invert_size)
- )
-
-
-class Normalizer(nn.Module):
- def __init__(self, mu, std):
- super().__init__()
- self.mu = nn.Parameter(mu)
- self.log_var = nn.Parameter(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 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 train_encoder(
- train_input,
- dim_hidden=64,
- block_size=16,
- nb_bits_per_block=10,
- lr_start=1e-3,
- lr_end=1e-5,
+def create_data_and_processors(
+ nb_train_samples,
+ nb_test_samples,
+ mode,
+ nb_steps,
+ depth=3,
+ nb_bits_per_token=8,
nb_epochs=10,
- batch_size=25,
device=torch.device("cpu"),
+ device_storage=torch.device("cpu"),
+ logger=None,
):
- mu, std = train_input.mean(), train_input.std()
+ assert mode in ["first_last"]
- encoder = nn.Sequential(
- Normalizer(mu, std),
- nn.Conv2d(3, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(
- dim_hidden,
- nb_bits_per_block,
- kernel_size=block_size,
- stride=block_size,
- padding=0,
- ),
- SignSTE(),
- )
+ if mode == "first_last":
+ steps = [True] + [False] * (nb_steps + 1) + [True]
- decoder = nn.Sequential(
- nn.ConvTranspose2d(
- nb_bits_per_block,
- dim_hidden,
- kernel_size=block_size,
- stride=block_size,
- padding=0,
- ),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, dim_hidden, kernel_size=5, stride=1, padding=2),
- nn.ReLU(),
- nn.Conv2d(dim_hidden, 3, kernel_size=5, stride=1, padding=2),
+ 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
)
- model = nn.Sequential(encoder, decoder)
-
- nb_parameters = sum(p.numel() for p in model.parameters())
-
- print(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
- )
- print(f"lr {lr}")
- optimizer = torch.optim.Adam(model.parameters(), lr=lr)
- acc_loss, nb_samples = 0.0, 0
-
- for input in tqdm.tqdm(
- train_input.split(batch_size),
- dynamic_ncols=True,
- desc="vqae-train",
- total=train_input.size(0) // batch_size,
- ):
- output = model(input)
- loss = F.mse_loss(output, input)
- acc_loss += loss.item() * input.size(0)
- nb_samples += input.size(0)
+ encoder, quantizer, decoder = train_encoder(
+ train_input,
+ test_input,
+ depth=depth,
+ nb_bits_per_token=nb_bits_per_token,
+ 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()
- optimizer.zero_grad()
- loss.backward()
- optimizer.step()
+ frames.append(output)
- print(f"loss {k} {acc_loss/nb_samples}")
- sys.stdout.flush()
+ return torch.cat(frames, dim=0)
- return encoder, decoder
+ return train_input, train_actions, test_input, test_actions, frame2seq, seq2frame
######################################################################
if __name__ == "__main__":
- import time
-
- all_frames = []
- nb = 25000
- start_time = time.perf_counter()
- for n in tqdm.tqdm(
- range(nb),
- dynamic_ncols=True,
- desc="world-data",
- ):
- frames, actions = generate_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)
- encoder, decoder = train_encoder(input)
-
- # 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())
-
- z = encoder(input)
- results = decoder(z)
+ (
+ train_input,
+ train_actions,
+ test_input,
+ test_actions,
+ frame2seq,
+ seq2frame,
+ ) = create_data_and_processors(
+ 25000, 1000,
+ nb_epochs=5,
+ mode="first_last",
+ nb_steps=20,
+ )
- print(f"{input.size()=} {z.size()=} {results.size()=}")
+ input = test_input[:256]
- torchvision.utils.save_image(input[:64], "orig.png", nrow=8)
+ seq = frame2seq(input)
+ output = seq2frame(seq)
- torchvision.utils.save_image(results[:64], "qtiz.png", nrow=8)
+ torchvision.utils.save_image(
+ input.float() / (Box.nb_rgb_levels - 1), "orig.png", nrow=16
+ )
- # frames, actions = generate_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(
+ output.float() / (Box.nb_rgb_levels - 1), "qtiz.png", nrow=16
+ )