Update.
authorFrançois Fleuret <francois@fleuret.org>
Sun, 16 Jul 2023 13:31:38 +0000 (15:31 +0200)
committerFrançois Fleuret <francois@fleuret.org>
Sun, 16 Jul 2023 13:31:38 +0000 (15:31 +0200)
tasks.py
world.py

index df3fd81..f8fb9b9 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -959,6 +959,7 @@ class World(Task):
         vqae_nb_epochs,
         logger=None,
         device=torch.device("cpu"),
+        device_storage=torch.device("cpu"),
     ):
         self.batch_size = batch_size
         self.device = device
@@ -978,12 +979,13 @@ class World(Task):
             nb_epochs=vqae_nb_epochs,
             logger=logger,
             device=device,
+            device_storage=device_storage,
         )
 
         print(f"{train_action_seq.size()=}")
 
-        train_frame_seq = self.frame2seq(train_frames)
-        test_frame_seq = self.frame2seq(test_frames)
+        train_frame_seq = self.frame2seq(train_frames).to(device_storage)
+        test_frame_seq = self.frame2seq(test_frames).to(device_storage)
 
         nb_frame_codes = max(train_frame_seq.max(), test_frame_seq.max()) + 1
         nb_action_codes = max(train_action_seq.max(), test_action_seq.max()) + 1
@@ -993,6 +995,7 @@ class World(Task):
         self.nb_codes = nb_frame_codes + nb_action_codes
 
         train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
+        print(f"{train_action_seq.device=} {nb_frame_codes.device=}")
         train_action_seq += nb_frame_codes
         self.train_input = torch.cat(
             (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
@@ -1014,7 +1017,7 @@ class World(Task):
         for batch in tqdm.tqdm(
             input.split(self.batch_size), dynamic_ncols=True, desc=desc
         ):
-            yield batch
+            yield batch.to(self.device)
 
     def vocabulary_size(self):
         return self.nb_codes
@@ -1026,7 +1029,7 @@ class World(Task):
             2 * self.len_frame_seq + self.len_action_seq, device=self.device
         )[None, :]
 
-        input = self.test_input[:64]
+        input = self.test_input[:64].to(self.device)
         result = input.clone()
 
         ar_mask = (
index fb8609d..fa305cf 100755 (executable)
--- a/world.py
+++ b/world.py
@@ -149,6 +149,7 @@ def train_encoder(
         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 = z if k < 2 else quantizer(z)
             output = decoder(zq)
@@ -168,6 +169,7 @@ def train_encoder(
         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 = z if k < 1 else quantizer(z)
             output = decoder(zq)
@@ -341,6 +343,7 @@ def create_data_and_processors(
     nb_steps,
     nb_epochs=10,
     device=torch.device("cpu"),
+    device_storage=torch.device("cpu"),
     logger=None,
 ):
     assert mode in ["first_last"]
@@ -349,9 +352,9 @@ def create_data_and_processors(
         steps = [True] + [False] * (nb_steps + 1) + [True]
 
     train_input, train_actions = generate_episodes(nb_train_samples, steps)
-    train_input, train_actions = train_input.to(device), train_actions.to(device)
+    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), test_actions.to(device)
+    test_input, test_actions = test_input.to(device_storage), test_actions.to(device_storage)
 
     encoder, quantizer, decoder = train_encoder(
         train_input, test_input, nb_epochs=nb_epochs, logger=logger, device=device
@@ -360,21 +363,22 @@ def create_data_and_processors(
     quantizer.train(False)
     decoder.train(False)
 
-    z = encoder(train_input[:1])
-    pow2 = (2 ** torch.arange(z.size(1), device=z.device))[None, None, :]
+    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)
                 )
-                * pow2
+                * p
             ).sum(-1)
 
             seq.append(output)
@@ -383,9 +387,10 @@ def create_data_and_processors(
 
     def seq2frame(input, batch_size=25, T=1e-2):
         frames = []
-
+        p = pow2.to(device)
         for seq in input.split(batch_size):
-            zd_bool = (seq[:, :, None] // pow2) % 2
+            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(