Update.
authorFrançois Fleuret <francois@fleuret.org>
Sat, 15 Jul 2023 22:04:16 +0000 (00:04 +0200)
committerFrançois Fleuret <francois@fleuret.org>
Sat, 15 Jul 2023 22:04:16 +0000 (00:04 +0200)
main.py
tasks.py
world.py

diff --git a/main.py b/main.py
index 58e8046..305bd3c 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -136,7 +136,7 @@ parser.add_argument("--expr_input_file", type=str, default=None)
 ##############################
 # World options
 
-parser.add_argument("--world_vqae_nb_epochs", type=int, default=10)
+parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 
 ######################################################################
 
@@ -187,9 +187,9 @@ default_args = {
         "nb_test_samples": 10000,
     },
     "world": {
-        "nb_epochs": 5,
+        "nb_epochs": 10,
         "batch_size": 25,
-        "nb_train_samples": 10000,
+        "nb_train_samples": 125000,
         "nb_test_samples": 1000,
     },
 }
@@ -334,6 +334,7 @@ elif args.task == "world":
         nb_test_samples=args.nb_test_samples,
         batch_size=args.batch_size,
         vqae_nb_epochs=args.world_vqae_nb_epochs,
+        logger=log_string,
         device=device,
     )
 
index 96d0621..df3fd81 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -29,7 +29,7 @@ def masked_inplace_autoregression(
             batches,
             dynamic_ncols=True,
             desc=progress_bar_desc,
-            #total=input.size(0) // batch_size,
+            # total=input.size(0) // batch_size,
         )
 
     with torch.autograd.no_grad():
@@ -957,6 +957,7 @@ class World(Task):
         nb_test_samples,
         batch_size,
         vqae_nb_epochs,
+        logger=None,
         device=torch.device("cpu"),
     ):
         self.batch_size = batch_size
@@ -964,9 +965,9 @@ class World(Task):
 
         (
             train_frames,
-            self.train_actions,
+            train_action_seq,
             test_frames,
-            self.test_actions,
+            test_action_seq,
             self.frame2seq,
             self.seq2frame,
         ) = world.create_data_and_processors(
@@ -975,15 +976,33 @@ class World(Task):
             mode="first_last",
             nb_steps=30,
             nb_epochs=vqae_nb_epochs,
+            logger=logger,
             device=device,
         )
 
-        self.train_input = self.frame2seq(train_frames)
-        self.train_input = self.train_input.reshape(self.train_input.size(0) // 2, -1)
-        self.test_input = self.frame2seq(test_frames)
-        self.test_input = self.test_input.reshape(self.test_input.size(0) // 2, -1)
+        print(f"{train_action_seq.size()=}")
 
-        self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+        train_frame_seq = self.frame2seq(train_frames)
+        test_frame_seq = self.frame2seq(test_frames)
+
+        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
+
+        self.len_frame_seq = train_frame_seq.size(1)
+        self.len_action_seq = train_action_seq.size(1)
+        self.nb_codes = nb_frame_codes + nb_action_codes
+
+        train_frame_seq = train_frame_seq.reshape(train_frame_seq.size(0) // 2, 2, -1)
+        train_action_seq += nb_frame_codes
+        self.train_input = torch.cat(
+            (train_frame_seq[:, 0, :], train_action_seq, train_frame_seq[:, 1, :]), 1
+        )
+
+        test_frame_seq = test_frame_seq.reshape(test_frame_seq.size(0) // 2, 2, -1)
+        test_action_seq += nb_frame_codes
+        self.test_input = torch.cat(
+            (test_frame_seq[:, 0, :], test_action_seq, test_frame_seq[:, 1, :]), 1
+        )
 
     def batches(self, split="train", nb_to_use=-1, desc=None):
         assert split in {"train", "test"}
@@ -1003,11 +1022,16 @@ class World(Task):
     def produce_results(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis
     ):
-        l = self.train_input.size(1)
-        k = torch.arange(l, device=self.device)[None, :]
-        result = self.test_input[:64].clone()
+        k = torch.arange(
+            2 * self.len_frame_seq + self.len_action_seq, device=self.device
+        )[None, :]
+
+        input = self.test_input[:64]
+        result = input.clone()
 
-        ar_mask = (k >= l // 2).long().expand_as(result)
+        ar_mask = (
+            (k >= self.len_frame_seq + self.len_action_seq).long().expand_as(result)
+        )
         result *= 1 - ar_mask
 
         masked_inplace_autoregression(
@@ -1019,14 +1043,22 @@ class World(Task):
             device=self.device,
         )
 
-        result = result.reshape(result.size(0) * 2, -1)
+        seq_start = input[:, : self.len_frame_seq]
+        seq_end = input[:, self.len_frame_seq + self.len_action_seq :]
+        seq_predicted = result[:, self.len_frame_seq + self.len_action_seq :]
+
+        result = torch.cat(
+            (seq_start[:, None, :], seq_end[:, None, :], seq_predicted[:, None, :]), 1
+        )
+        result = result.reshape(-1, result.size(-1))
+        print(f"{result.size()=}")
 
         frames = self.seq2frame(result)
         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
         torchvision.utils.save_image(
             frames.float() / (world.Box.nb_rgb_levels - 1),
             image_name,
-            nrow=8,
+            nrow=12,
             padding=1,
             pad_value=0.0,
         )
index 5c21fad..fb8609d 100755 (executable)
--- 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)
 
@@ -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
@@ -326,7 +330,7 @@ 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 +341,7 @@ def create_data_and_processors(
     nb_steps,
     nb_epochs=10,
     device=torch.device("cpu"),
+    logger=None,
 ):
     assert mode in ["first_last"]
 
@@ -349,7 +354,7 @@ def create_data_and_processors(
     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, device=device
+        train_input, test_input, nb_epochs=nb_epochs, logger=logger, device=device
     )
     encoder.train(False)
     quantizer.train(False)