projects / culture.git / blobdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | inline | side by side
Update.
[culture.git] / main.py
diff --git a/main.py b/main.py
index 549e7ea..a021a71 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -12,7 +12,7 @@ from torch import nn
 from torch.nn import functional as F
 
 import ffutils
 from torch.nn import functional as F
 
 import ffutils
-import mygpt, tasks, problems
+import mygpt, tasks
 
 ######################################################################
 
 
 ######################################################################
 
@@ -29,8 +29,6 @@ parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
-parser.add_argument("--task", type=str, default="world", help="world")
-
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
 
 parser.add_argument("--result_dir", type=str, default=None)
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
 
 parser.add_argument("--result_dir", type=str, default=None)
@@ -82,23 +80,20 @@ parser.add_argument("--check", action="store_true", default=False)
 args = parser.parse_args()
 
 if args.result_dir is None:
 args = parser.parse_args()
 
 if args.result_dir is None:
-    args.result_dir = f"results_{args.task}"
+    args.result_dir = f"results_culture"
 
 ######################################################################
 
 
 ######################################################################
 
-default_task_args = {
-    "world": {
-        "model": "37M",
-        "batch_size": 100,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
+default_args = {
+    "model": "37M",
+    "batch_size": 100,
+    "nb_train_samples": 250000,
+    "nb_test_samples": 10000,
 }
 
 }
 
-if args.task in default_task_args:
-    for k, v in default_task_args[args.task].items():
-        if getattr(args, k) is None:
-            setattr(args, k, v)
+for k, v in default_args.items():
+    if getattr(args, k) is None:
+        setattr(args, k, v)
 
 ######################################################################
 
 
 ######################################################################
 
@@ -188,7 +183,7 @@ for n in vars(args):
 ######################################################################
 
 if args.check:
 ######################################################################
 
 if args.check:
-    args.nb_train_samples = 500
+    args.nb_train_samples = 2500
     args.nb_test_samples = 100
 
 if args.physical_batch_size is None:
     args.nb_test_samples = 100
 
 if args.physical_batch_size is None:
@@ -199,229 +194,14 @@ else:
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
-if args.task == "file":
-    assert (
-        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_train_file,
-        args.filetask_test_file,
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        shuffle=True,
-        device=device,
-    )
-    args.max_percents_of_test_in_train = 0
-
-elif args.task == "byheart":
-    task = tasks.SandBox(
-        problem=problems.ProblemByHeart(separation=args.byheart_separation),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-    args.max_percents_of_test_in_train = -1
-
-elif args.task == "world":
-    task = tasks.World(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        result_dir=args.result_dir,
-        logger=log_string,
-        device=device,
-    )
-    args.max_percents_of_test_in_train = -1
-
-elif args.task == "learnop":
-    task = tasks.SandBox(
-        problem=problems.ProblemLearnOperator(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-
-elif args.task == "guessop":
-    task = tasks.SandBox(
-        problem=problems.ProblemGuessOperator(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-
-elif args.task == "twotargets":
-    task = tasks.SandBox(
-        problem=problems.ProblemTwoTargets(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        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.physical_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.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "addition":
-    task = tasks.SandBox(
-        problem=problems.ProblemAddition(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "picoclvr":
-    task = tasks.PicoCLVR(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        height=args.picoclvr_height,
-        width=args.picoclvr_width,
-        nb_colors=args.picoclvr_nb_colors,
-        logger=log_string,
-        device=device,
-        pruner_train=picoclvr_pruner_train,
-        pruner_eval=picoclvr_pruner_eval,
-    )
-
-elif args.task == "mnist":
-    task = tasks.MNIST(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        device=device,
-    )
-
-elif args.task == "maze":
-    task = tasks.Maze(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        height=args.maze_height,
-        width=args.maze_width,
-        nb_walls=args.maze_nb_walls,
-        device="cpu",
-    )
-
-elif args.task == "snake":
-    task = tasks.Snake(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        height=args.snake_height,
-        width=args.snake_width,
-        nb_colors=args.snake_nb_colors,
-        length=args.snake_length,
-        prompt_length=args.snake_length // 2,
-        device=device,
-    )
-
-elif args.task == "stack":
-    task = tasks.Stack(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        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,
-    )
-
-elif args.task == "expr":
-    task = tasks.Expr(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        nb_variables=args.expr_nb_variables,
-        sequence_length=args.expr_sequence_length,
-        operand_max=args.expr_operand_max,
-        result_max=args.expr_result_max,
-        batch_size=args.physical_batch_size,
-        device=device,
-    )
-
-elif args.task == "rpl":
-    task = tasks.RPL(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        nb_starting_values=args.rpl_nb_starting_values,
-        max_input=args.rpl_max_input,
-        prog_len=args.rpl_prog_len,
-        nb_runs=args.rpl_nb_runs,
-        no_prog=args.rpl_no_prog,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "grid":
-    task = tasks.Grid(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        size=args.grid_size,
-        fraction_play=args.grid_fraction_play,
-        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.physical_batch_size,
-        result_dir=args.result_dir,
-        logger=log_string,
-        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.physical_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}")
+task = tasks.World(
+    nb_train_samples=args.nb_train_samples,
+    nb_test_samples=args.nb_test_samples,
+    batch_size=args.physical_batch_size,
+    result_dir=args.result_dir,
+    logger=log_string,
+    device=device,
+)
 
 ######################################################################
 
 
 ######################################################################
 
@@ -555,23 +335,30 @@ def create_quizzes(
     task,
     nb_for_train=1000,
     nb_for_test=100,
     task,
     nb_for_train=1000,
     nb_for_test=100,
+    desired_average_logits=None,
 ):
     kept = []
 ):
     kept = []
+    nb_generated_tokens, sum_logits = 0, 0
 
     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 
     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
-        new_quizzes, nb_correct = task.create_new_quizzes(
+        nb_to_generate = 4 * (nb_for_train + nb_for_test)
+        new_quizzes, nb_correct, average_logits = task.create_new_quizzes(
             n_epoch=n_epoch,
             result_dir=args.result_dir,
             logger=log_string,
             n_epoch=n_epoch,
             result_dir=args.result_dir,
             logger=log_string,
-            nb=4 * (nb_for_train + nb_for_test),
+            nb=nb_to_generate,
             model=model,
             other_models=other_models,
             model=model,
             other_models=other_models,
+            desired_average_logits=desired_average_logits,
         )
 
         )
 
-        print(nb_correct)
+        nb_generated_tokens += new_quizzes.numel()
+        sum_logits += average_logits * new_quizzes.numel()
 
         to_keep = new_quizzes[nb_correct == len(other_models) - 1]
 
         to_keep = new_quizzes[nb_correct == len(other_models) - 1]
-        log_string(f"keep {to_keep.size(0)} quizzes")
+        log_string(
+            f"keep {to_keep.size(0)}/{new_quizzes.size(0)} quizzes ({to_keep.size(0)*100/new_quizzes.size(0):.02f}%)"
+        )
         kept.append(to_keep)
 
     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
         kept.append(to_keep)
 
     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
@@ -580,12 +367,14 @@ def create_quizzes(
     task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
 
     task.save_image(
     task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
 
     task.save_image(
-        new_quizzes[:96],
+        new_quizzes[:72],
         args.result_dir,
         f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
         log_string,
     )
 
         args.result_dir,
         f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
         log_string,
     )
 
+    return sum_logits / nb_generated_tokens
+
 
 ######################################################################
 
 
 ######################################################################
 
@@ -623,7 +412,13 @@ if args.check:
     nb_new_quizzes_for_train = 10
     nb_new_quizzes_for_test = 10
 
     nb_new_quizzes_for_train = 10
     nb_new_quizzes_for_test = 10
 
+desired_average_logits = None
+
 for n_epoch in range(args.nb_epochs):
 for n_epoch in range(args.nb_epochs):
+    a = [(model.id, float(model.main_test_accuracy)) for model in models]
+    a.sort(key=lambda p: p[0])
+    log_string(f"current accuracies {a}")
+
     # select the model with lowest accuracy
     models.sort(key=lambda model: model.main_test_accuracy)
     model = models[0]
     # select the model with lowest accuracy
     models.sort(key=lambda model: model.main_test_accuracy)
     model = models[0]
@@ -635,6 +430,8 @@ for n_epoch in range(args.nb_epochs):
     # improve it
     one_epoch(model, task)
 
     # improve it
     one_epoch(model, task)
 
+    task.renew_samples(args.nb_train_samples // args.nb_gpts)
+
     log_string(
         f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
     )
     log_string(
         f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
     )
@@ -642,17 +439,34 @@ for n_epoch in range(args.nb_epochs):
     # test it
     run_tests(model, task, deterministic_synthesis=False)
 
     # test it
     run_tests(model, task, deterministic_synthesis=False)
 
-    if model.main_test_accuracy >= accuracy_to_make_quizzes:
+    log_string(
+        f"test_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
+    )
+
+    if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_quizzes:
         other_models = models.copy()
         other_models.remove(model)
 
         other_models = models.copy()
         other_models.remove(model)
 
-        create_quizzes(
+        average_logits = create_quizzes(
             model,
             other_models,
             task,
             nb_for_train=nb_new_quizzes_for_train,
             nb_for_test=nb_new_quizzes_for_test,
             model,
             other_models,
             task,
             nb_for_train=nb_new_quizzes_for_train,
             nb_for_test=nb_new_quizzes_for_test,
+            desired_average_logits=desired_average_logits,
         )
 
         )
 
+        # We keep the first average logits as a reference
+        if desired_average_logits is None:
+            desired_average_logits = average_logits
+        else:
+            log_string(
+                f"desired_average_logits {desired_average_logits} average_logits {average_logits}"
+            )
+
+        # We update everyone
+        for model in models:
+            run_tests(model, task, deterministic_synthesis=False)
+
 
 ######################################################################
 
 ######################################################################
An attempt at having GPTs general their own quizzes.