5 import torch, torchvision
8 from torch.nn import functional as F
13 def __init__(self, x, y, w, h, r, g, b):
22 def collision(self, scene):
26 and max(self.x, c.x) <= min(self.x + self.w, c.x + c.w)
27 and max(self.y, c.y) <= min(self.y + self.h, c.y + c.h)
33 def scene2tensor(xh, yh, scene, size=64):
34 width, height = size, size
35 pixel_map = torch.ByteTensor(width, height, 4).fill_(255)
36 data = pixel_map.numpy()
37 surface = cairo.ImageSurface.create_for_data(
38 data, cairo.FORMAT_ARGB32, width, height
41 ctx = cairo.Context(surface)
42 ctx.set_fill_rule(cairo.FILL_RULE_EVEN_ODD)
45 ctx.move_to(b.x * size, b.y * size)
46 ctx.rel_line_to(b.w * size, 0)
47 ctx.rel_line_to(0, b.h * size)
48 ctx.rel_line_to(-b.w * size, 0)
50 ctx.set_source_rgba(b.r, b.g, b.b, 1.0)
54 ctx.set_source_rgba(0.0, 0.0, 0.0, 1.0)
55 ctx.move_to(xh * size - hs / 2, yh * size - hs / 2)
56 ctx.rel_line_to(hs, 0)
57 ctx.rel_line_to(0, hs)
58 ctx.rel_line_to(-hs, 0)
62 return pixel_map[None, :, :, :3].flip(-1).permute(0, 3, 1, 2).float() / 255
76 wh = torch.rand(2) * 0.2 + 0.2
77 xy = torch.rand(2) * (1 - wh)
78 c = colors[torch.randint(len(colors), (1,))]
80 xy[0].item(), xy[1].item(), wh[0].item(), wh[1].item(), c[0], c[1], c[2]
82 if not b.collision(scene):
88 def sequence(nb_steps=10, all_frames=False):
105 scene = random_scene()
106 xh, yh = tuple(x.item() for x in torch.rand(2))
108 frames.append(scene2tensor(xh, yh, scene))
110 actions = torch.randint(len(effects), (nb_steps,))
114 g, dx, dy = effects[a]
117 if b.x <= xh and b.x + b.w >= xh and b.y <= yh and b.y + b.h >= yh:
126 or b.collision(scene)
137 if xh < 0 or xh > 1 or yh < 0 or yh > 1:
141 frames.append(scene2tensor(xh, yh, scene))
144 frames.append(scene2tensor(xh, yh, scene))
149 return frames, actions
152 ######################################################################
155 # ||x_i - c_j||^2 = ||x_i||^2 + ||c_j||^2 - 2<x_i, c_j>
159 return nx[:, None] + nc[None, :] - 2 * x @ c.t()
162 def update_centroids(x, c, nb_min=1):
163 _, b = sq2matrix(x, c).min(1)
167 for k in range(0, c.size(0)):
168 i = b.eq(k).nonzero(as_tuple=False).squeeze()
169 if i.numel() >= nb_min:
170 c[k] = x.index_select(0, i).mean(0)
172 n = torch.randint(x.size(0), (1,))
176 return c, b, nb_resets
179 def kmeans(x, nb_centroids, nb_min=1):
180 if x.size(0) < nb_centroids * nb_min:
181 print("Not enough points!")
184 c = x[torch.randperm(x.size(0))[:nb_centroids]]
185 t = torch.full((x.size(0),), -1)
189 c, u, nb_resets = update_centroids(x, c, nb_min)
191 nb_changes = (u - t).sign().abs().sum() + nb_resets
199 ######################################################################
202 def patchify(x, factor, invert_size=None):
203 if invert_size is None:
209 x.size(2) // factor, # 3
211 x.size(3) // factor, # 5
213 .permute(0, 2, 4, 1, 3, 5)
214 .reshape(-1, x.size(1), x.size(2) // factor, x.size(3) // factor)
223 invert_size[2] // factor, # 4
224 invert_size[3] // factor, # 5
226 .permute(0, 3, 1, 4, 2, 5)
227 .reshape(invert_size)
231 def train_encoder(input, device=torch.device("cpu")):
232 class SomeLeNet(nn.Module):
235 self.conv1 = nn.Conv2d(1, 32, kernel_size=5)
236 self.conv2 = nn.Conv2d(32, 64, kernel_size=5)
237 self.fc1 = nn.Linear(256, 200)
238 self.fc2 = nn.Linear(200, 10)
240 def forward(self, x):
241 x = F.relu(F.max_pool2d(self.conv1(x), kernel_size=3))
242 x = F.relu(F.max_pool2d(self.conv2(x), kernel_size=2))
243 x = x.view(x.size(0), -1)
244 x = F.relu(self.fc1(x))
248 ######################################################################
252 nb_parameters = sum(p.numel() for p in model.parameters())
254 print(f"nb_parameters {nb_parameters}")
256 optimizer = torch.optim.SGD(model.parameters(), lr=lr)
257 criterion = nn.CrossEntropyLoss()
262 train_input, train_targets = train_input.to(device), train_targets.to(device)
263 test_input, test_targets = test_input.to(device), test_targets.to(device)
265 mu, std = train_input.mean(), train_input.std()
266 train_input.sub_(mu).div_(std)
267 test_input.sub_(mu).div_(std)
269 start_time = time.perf_counter()
271 for k in range(nb_epochs):
274 for input, targets in zip(
275 train_input.split(batch_size), train_targets.split(batch_size)
277 output = model(input)
278 loss = criterion(output, targets)
279 acc_loss += loss.item()
281 optimizer.zero_grad()
286 for input, targets in zip(
287 test_input.split(batch_size), test_targets.split(batch_size)
289 wta = model(input).argmax(1)
290 nb_test_errors += (wta != targets).long().sum()
291 test_error = nb_test_errors / test_input.size(0)
292 duration = time.perf_counter() - start_time
294 print(f"loss {k} {duration:.02f}s {acc_loss:.02f} {test_error*100:.02f}%")
297 ######################################################################
299 if __name__ == "__main__":
304 start_time = time.perf_counter()
306 frames, actions = sequence(nb_steps=31)
308 end_time = time.perf_counter()
309 print(f"{nb / (end_time - start_time):.02f} samples per second")
311 input = torch.cat(all_frames, 0)
313 # x = patchify(input, 8)
314 # y = x.reshape(x.size(0), -1)
315 # print(f"{x.size()=} {y.size()=}")
316 # centroids, t = kmeans(y, 4096)
317 # results = centroids[t]
318 # results = results.reshape(x.size())
319 # results = patchify(results, 8, input.size())
321 print(f"{input.size()=} {results.size()=}")
323 torchvision.utils.save_image(input[:64], "orig.png", nrow=8)
324 torchvision.utils.save_image(results[:64], "qtiz.png", nrow=8)
326 # frames, actions = sequence(nb_steps=31, all_frames=True)
327 # frames = torch.cat(frames, 0)
328 # torchvision.utils.save_image(frames, "seq.png", nrow=8)