X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=1d52b6defc3cd22ddea610e1fbfc071358354a47;hb=cd3329fc206bacfd90a8e2cbe364244359568733;hp=9a3d34633bd88cff4ac1e05ee77989e4d725e7b1;hpb=960c93d7c0aea41d180814c46d3a05686a426764;p=picoclvr.git

diff --git a/main.py b/main.py
index 9a3d346..1d52b6d 100755
--- a/main.py
+++ b/main.py
@@ -32,8 +32,8 @@ parser = argparse.ArgumentParser(
 parser.add_argument(
     "--task",
     type=str,
-    default="sandbox",
-    help="byheart, learnop, guessop, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl",
+    default="twotargets",
+    help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
 )
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
@@ -46,7 +46,7 @@ parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
 
 ########################################
 
-parser.add_argument("--nb_epochs", type=int, default=None)
+parser.add_argument("--nb_epochs", type=int, default=25)
 
 parser.add_argument("--batch_size", type=int, default=None)
 
@@ -62,7 +62,7 @@ parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30:
 
 ########################################
 
-parser.add_argument("--model", type=str, default="37M")
+parser.add_argument("--model", type=str, default=None)
 
 parser.add_argument("--dim_model", type=int, default=None)
 
@@ -89,16 +89,21 @@ parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 ##############################
 # rpl options
 
-parser.add_argument("--rpl_nb_starting_values", type=int, default=5)
+parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 
 parser.add_argument("--rpl_max_input", type=int, default=9)
 
-parser.add_argument("--rpl_prog_len", type=int, default=10)
+parser.add_argument("--rpl_prog_len", type=int, default=8)
 
-parser.add_argument("--rpl_nb_runs", type=int, default=8)
+parser.add_argument("--rpl_nb_runs", type=int, default=5)
 
 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 
+##############################
+# grid options
+
+parser.add_argument("--grid_size", type=int, default=6)
+
 ##############################
 # picoclvr options
 
@@ -113,18 +118,18 @@ parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 ##############################
 # Maze options
 
-parser.add_argument("--maze_height", type=int, default=23)
+parser.add_argument("--maze_height", type=int, default=13)
 
-parser.add_argument("--maze_width", type=int, default=39)
+parser.add_argument("--maze_width", type=int, default=21)
 
-parser.add_argument("--maze_nb_walls", type=int, default=45)
+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_height", type=int, default=9)
 
-parser.add_argument("--snake_width", type=int, default=8)
+parser.add_argument("--snake_width", type=int, default=12)
 
 parser.add_argument("--snake_nb_colors", type=int, default=5)
 
@@ -155,9 +160,11 @@ parser.add_argument("--expr_result_max", type=int, default=99)
 parser.add_argument("--expr_input_file", type=str, default=None)
 
 ##############################
-# World options
+# Mixing
 
-parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
+parser.add_argument("--mixing_hard", action="store_true", default=False)
+
+parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 
 ######################################################################
 
@@ -171,59 +178,101 @@ if args.result_dir is None:
 ######################################################################
 
 default_task_args = {
-    "sandbox": {
-        "nb_epochs": 50,
+    "addition": {
+        "model": "352M",
         "batch_size": 25,
-        "nb_train_samples": 100000,
+        "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
-    "picoclvr": {
-        "nb_epochs": 25,
+    "byheart": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 50000,
+        "nb_test_samples": 10000,
+    },
+    "expr": {
+        "model": "352M",
+        "batch_size": 25,
+        "nb_train_samples": 2500000,
+        "nb_test_samples": 10000,
+    },
+    "grid": {
+        "model": "37M",
         "batch_size": 25,
         "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
-    "mnist": {
-        "nb_epochs": 25,
+    "qmlp": {
+        "model": "37M",
         "batch_size": 10,
-        "nb_train_samples": 250000,
+        "nb_train_samples": 100000,
+        "nb_test_samples": 1000,
+    },
+    "guessop": {
+        "model": "352M",
+        "batch_size": 25,
+        "nb_train_samples": 1000000,
+        "nb_test_samples": 10000,
+    },
+    "learnop": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 50000,
         "nb_test_samples": 10000,
     },
     "maze": {
-        "nb_epochs": 25,
+        "model": "37M",
         "batch_size": 5,
+        "nb_train_samples": 100000,
+        "nb_test_samples": 10000,
+    },
+    "picoclvr": {
+        "model": "37M",
+        "batch_size": 25,
         "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
+    "rpl": {
+        "model": "352M",
+        "batch_size": 5,
+        "nb_train_samples": 2500000,
+        "nb_test_samples": 10000,
+    },
     "snake": {
-        "nb_epochs": 5,
+        "model": "37M",
         "batch_size": 25,
         "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
     "stack": {
-        "nb_epochs": 5,
+        "model": "37M",
         "batch_size": 25,
         "nb_train_samples": 100000,
         "nb_test_samples": 1000,
     },
-    "expr": {
-        "nb_epochs": 40,
+    "twotargets": {
+        "model": "37M",
         "batch_size": 25,
-        "nb_train_samples": 1000000,
+        "nb_train_samples": 50000,
         "nb_test_samples": 10000,
     },
-    "rpl": {
-        "nb_epochs": 40,
+    "memory": {
+        "model": "4M",
+        "batch_size": 100,
+        "nb_train_samples": 5000,
+        "nb_test_samples": 1000,
+    },
+    "mixing": {
+        "model": "37M",
         "batch_size": 25,
-        "nb_train_samples": 100000,
+        "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
-    "world": {
-        "nb_epochs": 10,
-        "batch_size": 25,
-        "nb_train_samples": 25000,
-        "nb_test_samples": 1000,
+    "mnist": {
+        "model": "37M",
+        "batch_size": 10,
+        "nb_train_samples": 60000,
+        "nb_test_samples": 10000,
     },
 }
 
@@ -242,6 +291,13 @@ default_model_args = {
         "nb_heads": 2,
         "nb_blocks": 2,
     },
+    "4M": {
+        "dim_model": 256,
+        "dim_keys": 32,
+        "dim_hidden": 1024,
+        "nb_heads": 4,
+        "nb_blocks": 6,
+    },
     "37M": {
         "dim_model": 512,
         "dim_keys": 64,
@@ -305,6 +361,8 @@ def log_string(s):
     sys.stdout.flush()
 
 
+log_string(f"argv {' '.join(sys.argv)}")
+
 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
@@ -339,7 +397,7 @@ if args.task == "byheart":
         logger=log_string,
         device=device,
     )
-
+    args.max_percents_of_test_in_train = -1
 
 elif args.task == "learnop":
     task = tasks.SandBox(
@@ -373,6 +431,28 @@ elif args.task == "twotargets":
         device=device,
     )
 
+elif args.task == "memory":
+    task = tasks.SandBox(
+        problem=problems.ProblemMemory(),
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        logger=log_string,
+        device=device,
+    )
+
+elif args.task == "mixing":
+    task = tasks.SandBox(
+        problem=problems.ProblemMixing(
+            hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
+        ),
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        logger=log_string,
+        device=device,
+    )
+
 elif args.task == "addition":
     task = tasks.SandBox(
         problem=problems.ProblemAddition(),
@@ -468,12 +548,22 @@ elif args.task == "rpl":
         device=device,
     )
 
-elif args.task == "world":
-    task = tasks.World(
+elif args.task == "grid":
+    task = tasks.Grid(
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
         batch_size=args.batch_size,
-        vqae_nb_epochs=args.world_vqae_nb_epochs,
+        size=args.grid_size,
+        logger=log_string,
+        device=device,
+    )
+
+elif args.task == "qmlp":
+    task = tasks.QMLP(
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        result_dir=args.result_dir,
         logger=log_string,
         device=device,
     )
@@ -563,33 +653,33 @@ train_set_perplexity = math.exp(entropy)
 ######################################################################
 # A bit of paranoia never hurts
 
+if args.max_percents_of_test_in_train >= 0:
+
+    def subsets_as_tuples(batches, cs):
+        s = set()
+        for batch in batches:
+            for x in batch:
+                s.add(tuple([v.item() for v in x]))
+                if len(s) == cs:
+                    yield s
+                    s = set()
+        yield s
+
+    nb_test, nb_in_train = 0, 0
+    for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
+        in_train = set()
+        for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
+            in_train.update(test_subset.intersection(train_subset))
+        nb_in_train += len(in_train)
+        nb_test += len(test_subset)
+
+    log_string(
+        f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
+    )
 
-def subsets_as_tuples(batches, cs):
-    s = set()
-    for batch in batches:
-        for x in batch:
-            s.add(tuple([v.item() for v in x]))
-            if len(s) == cs:
-                yield s
-                s = set()
-    yield s
-
-
-nb_test, nb_in_train = 0, 0
-for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
-    in_train = set()
-    for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
-        in_train.update(test_subset.intersection(train_subset))
-    nb_in_train += len(in_train)
-    nb_test += len(test_subset)
-
-log_string(
-    f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
-)
-
-assert (
-    nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
-), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
+    assert (
+        nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
+    ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 
 ##############################