Update.

[picoclvr.git] / world.py

diff --git a/world.py b/world.py
index da7de75..b35a08e 100755 (executable)
--- a/world.py
+++ b/world.py
@@ -61,6 +61,19 @@ class SignSTE(nn.Module):
         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)
@@ -72,9 +85,9 @@ def loss_H(binary_logits, h_threshold=1):
 def train_encoder(
     train_input,
     test_input,
-    depth=2,
+    depth,
+    nb_bits_per_token,
     dim_hidden=48,
-    nb_bits_per_token=8,
     lambda_entropy=0.0,
     lr_start=1e-3,
     lr_end=1e-4,
@@ -159,7 +172,7 @@ def train_encoder(
         for input in tqdm.tqdm(train_input.split(batch_size), desc="vqae-train"):
             input = input.to(device)
             z = encoder(input)
-            zq = z if k < 2 else quantizer(z)
+            zq = quantizer(z)
             output = decoder(zq)
 
             output = output.reshape(
@@ -169,7 +182,7 @@ def train_encoder(
             train_loss = F.cross_entropy(output, input)
 
             if lambda_entropy > 0:
-                loss = loss + lambda_entropy * loss_H(z, h_threshold=0.5)
+                train_loss = train_loss + lambda_entropy * loss_H(z, h_threshold=0.5)
 
             acc_train_loss += train_loss.item() * input.size(0)
 
@@ -182,7 +195,7 @@ def train_encoder(
         for input in tqdm.tqdm(test_input.split(batch_size), desc="vqae-test"):
             input = input.to(device)
             z = encoder(input)
-            zq = z if k < 1 else quantizer(z)
+            zq = quantizer(z)
             output = decoder(zq)
 
             output = output.reshape(
@@ -353,6 +366,8 @@ def create_data_and_processors(
     nb_test_samples,
     mode,
     nb_steps,
+    depth=3,
+    nb_bits_per_token=8,
     nb_epochs=10,
     device=torch.device("cpu"),
     device_storage=torch.device("cpu"),
@@ -375,6 +390,8 @@ def create_data_and_processors(
     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,
@@ -439,26 +456,21 @@ if __name__ == "__main__":
         frame2seq,
         seq2frame,
     ) = create_data_and_processors(
-        # 10000, 1000,
-        100,
-        100,
-        nb_epochs=2,
+        25000, 1000,
+        nb_epochs=5,
         mode="first_last",
         nb_steps=20,
     )
 
-    input = test_input[:64]
+    input = test_input[:256]
 
     seq = frame2seq(input)
-
-    print(f"{seq.size()=} {seq.dtype=} {seq.min()=} {seq.max()=}")
-
     output = seq2frame(seq)
 
     torchvision.utils.save_image(
-        input.float() / (Box.nb_rgb_levels - 1), "orig.png", nrow=8
+        input.float() / (Box.nb_rgb_levels - 1), "orig.png", nrow=16
     )
 
     torchvision.utils.save_image(
-        output.float() / (Box.nb_rgb_levels - 1), "qtiz.png", nrow=8
+        output.float() / (Box.nb_rgb_levels - 1), "qtiz.png", nrow=16
     )