Update.
authorFrançois Fleuret <francois@fleuret.org>
Tue, 11 Jul 2023 06:08:24 +0000 (08:08 +0200)
committerFrançois Fleuret <francois@fleuret.org>
Tue, 11 Jul 2023 06:08:24 +0000 (08:08 +0200)
tasks.py
world.py

index 82d965b..75781ab 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -880,17 +880,22 @@ class Expr(Task):
             values_input = expr.extract_results([self.seq2str(s) for s in input])
             values_result = expr.extract_results([self.seq2str(s) for s in result])
 
-            for i, r in zip(values_input, values_result):
-                for n, vi in i.items():
-                    vr = r.get(n)
-                    if vr is None or vr < 0:
-                        nb_missed += 1
-                    else:
-                        d = abs(vr - vi)
-                        if d >= nb_delta.size(0):
+            filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
+
+            with open(filename, "w") as f:
+                for i, r in zip(values_input, values_result):
+                    for n, vi in i.items():
+                        vr = r.get(n)
+                        f.write(f"{vi} {-1 if vr is None else vr}\n")
+
+                        if vr is None or vr < 0:
                             nb_missed += 1
                         else:
-                            nb_delta[d] += 1
+                            d = abs(vr - vi)
+                            if d >= nb_delta.size(0):
+                                nb_missed += 1
+                            else:
+                                nb_delta[d] += 1
 
             ######################################################################
 
index e76c07f..a43eff9 100755 (executable)
--- a/world.py
+++ b/world.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 
-import math
+import math, sys
 
 import torch, torchvision
 
@@ -30,7 +30,7 @@ class Box:
         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()
@@ -85,7 +85,7 @@ def random_scene():
     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),
@@ -105,7 +105,7 @@ def sequence(nb_steps=10, all_frames=False):
         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
@@ -138,10 +138,10 @@ def sequence(nb_steps=10, all_frames=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
@@ -228,70 +228,109 @@ def patchify(x, factor, invert_size=None):
         )
 
 
-def train_encoder(input, device=torch.device("cpu")):
-    class SomeLeNet(nn.Module):
-        def __init__(self):
-            super().__init__()
-            self.conv1 = nn.Conv2d(1, 32, kernel_size=5)
-            self.conv2 = nn.Conv2d(32, 64, kernel_size=5)
-            self.fc1 = nn.Linear(256, 200)
-            self.fc2 = nn.Linear(200, 10)
+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):
-            x = F.relu(F.max_pool2d(self.conv1(x), kernel_size=3))
-            x = F.relu(F.max_pool2d(self.conv2(x), kernel_size=2))
-            x = x.view(x.size(0), -1)
-            x = F.relu(self.fc1(x))
-            x = self.fc2(x)
-            return x
+    def forward(self, x):
+        return (x-self.mu)/torch.exp(self.log_var/2.0)
 
-    ######################################################################
+class SignSTE(nn.Module):
+    def __init__(self):
+        super().__init__()
 
-    model = SomeLeNet()
-
-    nb_parameters = sum(p.numel() for p in model.parameters())
+    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=50,
+    batch_size=25,
+    device=torch.device("cpu"),
+):
+    mu, std = train_input.mean(), train_input.std()
 
-    print(f"nb_parameters {nb_parameters}")
+    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(),
+    )
 
-    optimizer = torch.optim.SGD(model.parameters(), lr=lr)
-    criterion = nn.CrossEntropyLoss()
+    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.to(device)
-    criterion.to(device)
+    model = nn.Sequential(encoder, decoder)
 
-    train_input, train_targets = train_input.to(device), train_targets.to(device)
-    test_input, test_targets = test_input.to(device), test_targets.to(device)
+    nb_parameters = sum(p.numel() for p in model.parameters())
 
-    mu, std = train_input.mean(), train_input.std()
-    train_input.sub_(mu).div_(std)
-    test_input.sub_(mu).div_(std)
+    print(f"nb_parameters {nb_parameters}")
 
-    start_time = time.perf_counter()
+    model.to(device)
 
     for k in range(nb_epochs):
-        acc_loss = 0.0
+        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, targets in zip(
-            train_input.split(batch_size), train_targets.split(batch_size)
-        ):
+        for input in train_input.split(batch_size):
             output = model(input)
-            loss = criterion(output, targets)
-            acc_loss += loss.item()
+            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()
 
-        nb_test_errors = 0
-        for input, targets in zip(
-            test_input.split(batch_size), test_targets.split(batch_size)
-        ):
-            wta = model(input).argmax(1)
-            nb_test_errors += (wta != targets).long().sum()
-        test_error = nb_test_errors / test_input.size(0)
-        duration = time.perf_counter() - start_time
+        print(f"loss {k} {acc_loss/nb_samples}")
+        sys.stdout.flush()
 
-        print(f"loss {k} {duration:.02f}s {acc_loss:.02f} {test_error*100:.02f}%")
+    return encoder, decoder
 
 
 ######################################################################
@@ -300,15 +339,16 @@ 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)
+        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)
@@ -318,11 +358,15 @@ if __name__ == "__main__":
     # results = results.reshape(x.size())
     # results = patchify(results, 8, input.size())
 
-    print(f"{input.size()=} {results.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)