#!/usr/bin/env python
-import math
+import math, sys, tqdm
import torch, torchvision
return False
-def scene2tensor(xh, yh, scene, size=64):
+def scene2tensor(xh, yh, scene, size):
width, height = size, size
pixel_map = torch.ByteTensor(width, height, 4).fill_(255)
data = pixel_map.numpy()
return scene
-def sequence(nb_steps=10, all_frames=False):
+def generate_sequence(nb_steps=10, all_frames=False, 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))
+ frames.append(scene2tensor(xh, yh, scene, size=size))
actions = torch.randint(len(effects), (nb_steps,))
change = False
xh, yh = x, y
if all_frames:
- frames.append(scene2tensor(xh, yh, scene))
+ frames.append(scene2tensor(xh, yh, scene, size=size))
if not all_frames:
- frames.append(scene2tensor(xh, yh, scene))
+ frames.append(scene2tensor(xh, yh, scene, size=size))
if change:
break
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
+ 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
+ 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 train_encoder(
+ train_input,
+ dim_hidden=64,
+ block_size=16,
+ nb_bits_per_block=10,
+ lr_start=1e-3,
+ lr_end=1e-5,
+ nb_epochs=10,
+ batch_size=25,
+ device=torch.device("cpu"),
+):
+ mu, std = train_input.mean(), train_input.std()
+
+ 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(),
+ )
+
+ 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),
+ )
+
+ 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)
+
+ optimizer.zero_grad()
+ loss.backward()
+ optimizer.step()
+
+ print(f"loss {k} {acc_loss/nb_samples}")
+ sys.stdout.flush()
+
+ return encoder, decoder
+
+
+######################################################################
+
if __name__ == "__main__":
import time
all_frames = []
- nb = 1000
+ nb = 25000
start_time = time.perf_counter()
- for n in range(nb):
- frames, actions = sequence(nb_steps=31)
+ 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)
- 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())
+ encoder, decoder = train_encoder(input)
- print(f"{input.size()=} {results.size()=}")
+ # 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)
+
+ print(f"{input.size()=} {z.size()=} {results.size()=}")
torchvision.utils.save_image(input[:64], "orig.png", nrow=8)
+
torchvision.utils.save_image(results[:64], "qtiz.png", nrow=8)
- # frames, actions = sequence(nb_steps=31, all_frames=True)
+ # frames, actions = generate_sequence(nb_steps=31, all_frames=True)
# frames = torch.cat(frames, 0)
# torchvision.utils.save_image(frames, "seq.png", nrow=8)
projects / picoclvr.git / blobdiff