Update.

[picoclvr.git] / main.py

diff --git a/main.py b/main.py
index f8e451b..acecfdd 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -92,6 +92,17 @@ parser.add_argument("--maze_width", type=int, default=21)
 
 parser.add_argument("--maze_nb_walls", type=int, default=15)
 
+##############################
+# Snake options
+
+parser.add_argument("--snake_height", type=int, default=6)
+
+parser.add_argument("--snake_width", type=int, default=8)
+
+parser.add_argument("--snake_nb_colors", type=int, default=3)
+
+parser.add_argument("--snake_length", type=int, default=400)
+
 ######################################################################
 
 args = parser.parse_args()
@@ -488,7 +499,7 @@ class TaskMNIST(Task):
         masked_inplace_autoregression(
             model, self.batch_size, results, ar_mask, device=self.device
         )
-        image_name = os.path.join(args.result_dir, f"result_mnist_{n_epoch:04d}.png")
+        image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
         torchvision.utils.save_image(
             1 - results.reshape(-1, 1, 28, 28) / 255.0,
             image_name,
@@ -608,7 +619,7 @@ class TaskMaze(Task):
             mazes, paths = self.seq2map(input)
             _, predicted_paths = self.seq2map(result)
 
-            filename = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
+            filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
             maze.save_image(
                 filename,
                 mazes=mazes,
@@ -627,7 +638,7 @@ class TaskMaze(Task):
 def generate_snake_sequences(
     nb, height, width, nb_colors, length, device=torch.device("cpu")
 ):
-    world = torch.randint(nb_colors, (nb, height, width), device=device)
+    worlds = torch.randint(nb_colors, (nb, height, width), device=device)
     # nb x 2
     snake_position = torch.cat(
         (
@@ -636,17 +647,17 @@ def generate_snake_sequences(
         ),
         1,
     )
-    snake_direction = torch.randint(4, (nb, 1), device=device)
-    result = torch.empty(nb, 2*length, device=device, dtype=torch.int64)
-    count = torch.arange(nb, device=device)  # [:,None]
+    snake_direction = torch.randint(4, (nb,), device=device)
+    sequences = torch.empty(nb, 2 * length, device=device, dtype=torch.int64)
+    i = torch.arange(nb, device=device)  # [:,None]
 
     for l in range(length):
         # nb x 3
         snake_next_direction = torch.cat(
             (
-                (snake_direction - 1) % 4,
-                snake_direction,
-                (snake_direction + 1) % 4,
+                (snake_direction[:, None] - 1) % 4,
+                snake_direction[:, None],
+                (snake_direction[:, None] + 1) % 4,
             ),
             1,
         )
@@ -668,25 +679,26 @@ def generate_snake_sequences(
                 snake_next_position[:, :, 1] >= 0, snake_next_position[:, :, 1] < width
             ),
         ).float()
-        val = torch.rand_like(val) * val * torch.tensor([[1.,4.,1.]], device=device)
+        val = (
+            torch.rand_like(val) * val * torch.tensor([[1.0, 4.0, 1.0]], device=device)
+        )
 
         # nb
-        i = torch.arange(val.size(0), device=device)
         j = val.argmax(1)
+        snake_direction = snake_next_direction[i, j]
 
-        # nb x 1
-        snake_direction = snake_next_direction[i[:, None], j[:, None]]
-
-        result[:, 2*l] = world[count, snake_position[:, 0], snake_position[:, 1]]
-        result[:, 2*l+1] = snake_direction[:,0]
+        sequences[:, 2 * l] = worlds[i, snake_position[:, 0], snake_position[:, 1]] + 4
+        sequences[:, 2 * l + 1] = snake_direction
 
         # nb x 2
-        snake_position = snake_next_position[i[:, None], j[:, None]].squeeze(1)
+        snake_position = snake_next_position[i, j]
+
+    return sequences, worlds
 
-    return result
 
-generate_snake_sequences(nb=2, height=4, width=5, nb_colors=3, length=10)
-exit(0)
+# generate_snake_sequences(nb=1, height=4, width=6, nb_colors=3, length=20)
+# exit(0)
+
 
 class TaskSnake(Task):
     def __init__(
@@ -705,10 +717,10 @@ class TaskSnake(Task):
         self.width = width
         self.device = device
 
-        self.train_input = generate_snake_sequences(
+        self.train_input, self.train_worlds = generate_snake_sequences(
             nb_train_samples, height, width, nb_colors, length, self.device
         )
-        self.test_input = generate_snake_sequences(
+        self.test_input, self.test_worlds = generate_snake_sequences(
             nb_test_samples, height, width, nb_colors, length, self.device
         )
 
@@ -729,6 +741,44 @@ class TaskSnake(Task):
     def vocabulary_size(self):
         return self.nb_codes
 
+    def produce_results(self, n_epoch, model):
+        with torch.autograd.no_grad():
+            t = model.training
+            model.eval()
+
+            def compute_nb_correct(input):
+                result = input.clone()
+                i = torch.arange(result.size(1), device=result.device)
+                ar_mask = torch.logical_and(i >= i.size(0) // 2, i % 2 == 0)[
+                    None, :
+                ].long()
+                result *= 1 - ar_mask
+                masked_inplace_autoregression(
+                    model, self.batch_size, result, ar_mask, device=self.device
+                )
+
+                nb_total = ar_mask.sum() * input.size(0)
+                nb_correct = ((result == input).long() * ar_mask).sum()
+
+                # nb_total = result.size(0)
+                # nb_correct = ((result - input).abs().sum(1) == 0).sum()
+
+                return nb_total, nb_correct
+
+            train_nb_total, train_nb_correct = compute_nb_correct(self.train_input)
+
+            log_string(
+                f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
+            )
+
+            test_nb_total, test_nb_correct = compute_nb_correct(self.test_input)
+
+            log_string(
+                f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
+            )
+
+            model.train(t)
+
 
 ######################################################################
 
@@ -786,10 +836,10 @@ elif args.task == "snake":
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
         batch_size=args.batch_size,
-        height=6,
-        width=8,
-        nb_colors=5,
-        length=100,
+        height=args.snake_height,
+        width=args.snake_width,
+        nb_colors=args.snake_nb_colors,
+        length=args.snake_length,
         device=device,
     )