Update.

[picoclvr.git] / main.py

diff --git a/main.py b/main.py
index 314a961..c1f4dc7 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -37,7 +37,7 @@ parser.add_argument(
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
 
-parser.add_argument("--result_dir", type=str, default="results_default")
+parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
@@ -113,7 +113,9 @@ parser.add_argument("--stack_nb_steps", type=int, default=100)
 
 parser.add_argument("--stack_nb_stacks", type=int, default=1)
 
-parser.add_argument("--stack_nb_digits", type=int, default=1)
+parser.add_argument("--stack_nb_digits", type=int, default=3)
+
+parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 
 ######################################################################
 
@@ -121,22 +123,8 @@ args = parser.parse_args()
 
 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 
-try:
-    os.mkdir(args.result_dir)
-except FileExistsError:
-    if not args.overwrite_results:
-        print(f"result directory {args.result_dir} already exists")
-        exit(1)
-
-log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
-
-if args.seed >= 0:
-    # torch.backends.cudnn.deterministic = True
-    # torch.backends.cudnn.benchmark = False
-    # torch.use_deterministic_algorithms(True)
-    torch.manual_seed(args.seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed_all(args.seed)
+if args.result_dir is None:
+    args.result_dir = f"results_{args.task}"
 
 ######################################################################
 
@@ -180,6 +168,25 @@ if args.task in default_args:
 
 ######################################################################
 
+try:
+    os.mkdir(args.result_dir)
+except FileExistsError:
+    if not args.overwrite_results:
+        print(f"result directory {args.result_dir} already exists")
+        exit(1)
+
+log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
+
+if args.seed >= 0:
+    # torch.backends.cudnn.deterministic = True
+    # torch.backends.cudnn.benchmark = False
+    # torch.use_deterministic_algorithms(True)
+    torch.manual_seed(args.seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(args.seed)
+
+######################################################################
+
 
 def log_string(s):
     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
@@ -702,14 +709,14 @@ class TaskMaze(Task):
                 model, "train", nb_to_use=1000
             )
             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}%"
+                f"accuracy_train {n_epoch} 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, count = self.compute_error(
                 model, "test", nb_to_use=1000
             )
             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}%"
+                f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
             )
 
             if count is not None:
@@ -855,7 +862,7 @@ class TaskSnake(Task):
             )
 
             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}%"
+                f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
             )
 
             model.train(t)
@@ -876,6 +883,7 @@ class TaskStack(Task):
         nb_steps,
         nb_stacks,
         nb_digits,
+        fraction_values_for_train=None,
         device=torch.device("cpu"),
     ):
         self.batch_size = batch_size
@@ -884,20 +892,37 @@ class TaskStack(Task):
         self.nb_digits = nb_digits
         self.device = device
 
+        if fraction_values_for_train is None:
+            values_for_train = None
+            values_for_test = None
+        else:
+            all = torch.randperm(10**nb_digits)
+            nb_for_train = int(all.size(0) * fraction_values_for_train)
+            values_for_train = all[:nb_for_train]
+            values_for_test = all[nb_for_train:]
+
         self.train_input, self.train_stack_counts = stack.generate_sequences(
-            nb_train_samples, nb_steps, nb_stacks, nb_digits, self.device
+            nb_train_samples,
+            nb_steps,
+            nb_stacks,
+            nb_digits,
+            values_for_train,
+            self.device,
         )
 
         self.test_input, self.test_stack_counts = stack.generate_sequences(
-            nb_test_samples, nb_steps, nb_stacks, nb_digits, self.device
+            nb_test_samples,
+            nb_steps,
+            nb_stacks,
+            nb_digits,
+            values_for_test,
+            self.device,
         )
 
-        mask = self.test_input.clone()
-        stack.remove_popped_values(mask, self.nb_stacks, self.nb_digits)
-        mask = mask != self.test_input
-        counts = self.test_stack_counts.flatten()[mask.flatten()]
+        i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
+        counts = self.test_stack_counts.flatten()[i.flatten()]
         counts = F.one_hot(counts).sum(0)
-        log_string(f"stack_count {counts}")
+        log_string(f"pop_stack_counts {counts}")
 
         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 
@@ -942,11 +967,12 @@ class TaskStack(Task):
             test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 
             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}%"
+                f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
             )
 
-            #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
-            input = self.test_input[:10, :20]
+            ##############################################################
+            # Log a few generated sequences
+            input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
             result = input.clone()
             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
             ar_mask = (result != input).long()
@@ -961,7 +987,7 @@ class TaskStack(Task):
                 log_string(
                     f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
                 )
-            #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+            ##############################################################
 
             model.train(t)
 
@@ -1038,6 +1064,7 @@ elif args.task == "stack":
         nb_steps=args.stack_nb_steps,
         nb_stacks=args.stack_nb_stacks,
         nb_digits=args.stack_nb_digits,
+        fraction_values_for_train=args.stack_fraction_values_for_train,
         device=device,
     )