X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=world.py;h=fa305cf1d4d03408924928e6b029f8f8d6a305cb;hb=2ac9d1299a84f96228f49fbdac02d5a7017445e5;hp=5c21fadcc95c5d21e328b5d9cbc37c2987e148dd;hpb=8e23dd068df00df61c690ffa89ecc8cb9db4b32d;p=picoclvr.git

diff --git a/world.py b/world.py
index 5c21fad..fa305cf 100755
--- a/world.py
+++ b/world.py
@@ -72,8 +72,12 @@ def train_encoder(
     lr_end=1e-4,
     nb_epochs=10,
     batch_size=25,
+    logger=None,
     device=torch.device("cpu"),
 ):
+    if logger is None:
+        logger = lambda s: print(s)
+
     mu, std = train_input.float().mean(), train_input.float().std()
 
     def encoder_core(depth, dim):
@@ -132,7 +136,7 @@ def train_encoder(
 
     nb_parameters = sum(p.numel() for p in model.parameters())
 
-    print(f"nb_parameters {nb_parameters}")
+    logger(f"nb_parameters {nb_parameters}")
 
     model.to(device)
 
@@ -145,6 +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)
@@ -164,6 +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)
@@ -179,7 +185,7 @@ def train_encoder(
         train_loss = acc_train_loss / train_input.size(0)
         test_loss = acc_test_loss / test_input.size(0)
 
-        print(f"train_ae {k} lr {lr} train_loss {train_loss} test_loss {test_loss}")
+        logger(f"train_ae {k} lr {lr} train_loss {train_loss} test_loss {test_loss}")
         sys.stdout.flush()
 
     return encoder, quantizer, decoder
@@ -326,7 +332,7 @@ def generate_episodes(nb, steps):
     for n in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world-data"):
         frames, actions = generate_episode(steps)
         all_frames += frames
-        all_actions += [actions]
+        all_actions += [actions[None, :]]
     return torch.cat(all_frames, 0).contiguous(), torch.cat(all_actions, 0)
 
 
@@ -337,6 +343,8 @@ 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"]
 
@@ -344,32 +352,33 @@ 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, device=device
+        train_input, test_input, nb_epochs=nb_epochs, logger=logger, device=device
     )
     encoder.train(False)
     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)
@@ -378,9 +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(