Update.

[picoclvr.git] / world.py

diff --git a/world.py b/world.py
index c33a584..fb8609d 100755 (executable)
--- a/world.py
+++ b/world.py
@@ -65,15 +65,19 @@ class SignSTE(nn.Module):
 def train_encoder(
     train_input,
     test_input,
-    depth=3,
+    depth=2,
     dim_hidden=48,
     nb_bits_per_token=8,
     lr_start=1e-3,
     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)
 
@@ -179,7 +183,7 @@ 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
@@ -322,20 +326,35 @@ def generate_episode(steps, size=64):
 
 
 def generate_episodes(nb, steps):
-    all_frames = []
+    all_frames, all_actions = [], []
     for n in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world-data"):
         frames, actions = generate_episode(steps)
         all_frames += frames
-    return torch.cat(all_frames, 0).contiguous()
+        all_actions += [actions[None, :]]
+    return torch.cat(all_frames, 0).contiguous(), torch.cat(all_actions, 0)
+
+
+def create_data_and_processors(
+    nb_train_samples,
+    nb_test_samples,
+    mode,
+    nb_steps,
+    nb_epochs=10,
+    device=torch.device("cpu"),
+    logger=None,
+):
+    assert mode in ["first_last"]
 
+    if mode == "first_last":
+        steps = [True] + [False] * (nb_steps + 1) + [True]
 
-def create_data_and_processors(nb_train_samples, nb_test_samples, nb_epochs=10):
-    steps = [True] + [False] * 30 + [True]
-    train_input = generate_episodes(nb_train_samples, steps)
-    test_input = generate_episodes(nb_test_samples, steps)
+    train_input, train_actions = generate_episodes(nb_train_samples, steps)
+    train_input, train_actions = train_input.to(device), train_actions.to(device)
+    test_input, test_actions = generate_episodes(nb_test_samples, steps)
+    test_input, test_actions = test_input.to(device), test_actions.to(device)
 
     encoder, quantizer, decoder = train_encoder(
-        train_input, test_input, nb_epochs=nb_epochs
+        train_input, test_input, nb_epochs=nb_epochs, logger=logger, device=device
     )
     encoder.train(False)
     quantizer.train(False)
@@ -380,15 +399,26 @@ def create_data_and_processors(nb_train_samples, nb_test_samples, nb_epochs=10):
 
         return torch.cat(frames, dim=0)
 
-    return train_input, test_input, frame2seq, seq2frame
+    return train_input, train_actions, test_input, test_actions, frame2seq, seq2frame
 
 
 ######################################################################
 
 if __name__ == "__main__":
-    train_input, test_input, frame2seq, seq2frame = create_data_and_processors(
+    (
+        train_input,
+        train_actions,
+        test_input,
+        test_actions,
+        frame2seq,
+        seq2frame,
+    ) = create_data_and_processors(
         # 10000, 1000,
-        100, 100, nb_epochs=2,
+        100,
+        100,
+        nb_epochs=2,
+        mode="first_last",
+        nb_steps=20,
     )
 
     input = test_input[:64]