X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=0f2cb61f5645839cc2e205f1cfd3097fe8075697;hb=232299b8af7e66a02e64bb2e47b525e2f50b099d;hp=55f2c2f4c6f9347cc6b8d04d5311bab811e4057a;hpb=a3f5378c684fb58b0dc839638e768c3a4b8e8a83;p=picoclvr.git

diff --git a/main.py b/main.py
index 55f2c2f..0f2cb61 100755
--- a/main.py
+++ b/main.py
@@ -33,7 +33,7 @@ parser.add_argument(
     "--task",
     type=str,
     default="twotargets",
-    help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
+    help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
 )
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
@@ -89,7 +89,9 @@ parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 ##############################
 # filetask
 
-parser.add_argument("--filetask_file", type=str, default=None)
+parser.add_argument("--filetask_train_file", type=str, default=None)
+
+parser.add_argument("--filetask_test_file", type=str, default=None)
 
 ##############################
 # rpl options
@@ -173,6 +175,19 @@ parser.add_argument("--mixing_hard", action="store_true", default=False)
 
 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 
+##############################
+# greed options
+
+parser.add_argument("--greed_height", type=int, default=5)
+
+parser.add_argument("--greed_width", type=int, default=7)
+
+parser.add_argument("--greed_T", type=int, default=25)
+
+parser.add_argument("--greed_nb_walls", type=int, default=5)
+
+parser.add_argument("--greed_nb_coins", type=int, default=2)
+
 ######################################################################
 
 args = parser.parse_args()
@@ -287,6 +302,12 @@ default_task_args = {
         "nb_train_samples": 60000,
         "nb_test_samples": 10000,
     },
+    "greed": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 25000,
+        "nb_test_samples": 10000,
+    },
 }
 
 if args.task in default_task_args:
@@ -403,13 +424,15 @@ picoclvr_pruner_eval = (
 
 if args.task == "file":
     assert (
-        args.filetask_file is not None
-    ), "You have to specify the task file with --filetask_file <filename>"
+        args.filetask_train_file is not None and args.filetask_test_file is not None
+    ), "You have to specify the task train and test files"
     task = tasks.TaskFromFile(
-        args.filetask_file,
+        args.filetask_train_file,
+        args.filetask_test_file,
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
         batch_size=args.batch_size,
+        shuffle=True,
         device=device,
     )
     args.max_percents_of_test_in_train = 0
@@ -595,6 +618,20 @@ elif args.task == "qmlp":
         device=device,
     )
 
+elif args.task == "greed":
+    task = tasks.Greed(
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        height=args.greed_height,
+        width=args.greed_width,
+        T=args.greed_T,
+        nb_walls=args.greed_nb_walls,
+        nb_coins=args.greed_nb_coins,
+        logger=log_string,
+        device=device,
+    )
+
 else:
     raise ValueError(f"Unknown task {args.task}")
 
@@ -666,8 +703,6 @@ if args.task == "expr" and args.expr_input_file is not None:
 
 ######################################################################
 
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
-
 # Compute the entropy of the training tokens
 
 token_count = 0
@@ -736,7 +771,7 @@ log_string(f"learning_rate_schedule {learning_rate_schedule}")
 
 nb_samples_seen = 0
 
-if nb_epochs_finished >= nb_epochs:
+if nb_epochs_finished >= args.nb_epochs:
     task.produce_results(
         n_epoch=nb_epochs_finished,
         model=model,
@@ -747,7 +782,7 @@ if nb_epochs_finished >= nb_epochs:
 
 time_pred_result = None
 
-for n_epoch in range(nb_epochs_finished, nb_epochs):
+for n_epoch in range(nb_epochs_finished, args.nb_epochs):
     learning_rate = learning_rate_schedule[n_epoch]
 
     log_string(f"learning_rate {learning_rate}")