Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index d1f82cf..7f9d521 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -5,17 +5,23 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
 
 # Written by Francois Fleuret <francois@fleuret.org>
 

-# torch.backends.cuda.matmul.allow_tf23
-# torch.autocast(torch.bfloat16)
-
-import math, sys, argparse, time, tqdm, os
+import math, sys, argparse, time, tqdm, os, datetime, warnings

 
 import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 
 import ffutils
 
 import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 
 import ffutils

-import mygpt, tasks
+import mygpt
+import sky, quizz_machine
+
+# world quizzes vs. culture quizzes
+
+######################################################################
+
+accuracy_to_make_c_quizzes = 0.975
+nb_new_c_quizzes_for_train = 1000
+nb_new_c_quizzes_for_test = 100

 
 ######################################################################
 
 
 ######################################################################
 


@@ -32,34 +38,31 @@ parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 

-parser.add_argument(
-    "--task",
-    type=str,
-    default="sandbox",
-    help="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world",
-)
-

 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
 
 parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
 
 parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 

-parser.add_argument("--nb_epochs", type=int, default=None)
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
+
+########################################
+
+parser.add_argument("--nb_epochs", type=int, default=10000)

 
 parser.add_argument("--batch_size", type=int, default=None)
 
 
 parser.add_argument("--batch_size", type=int, default=None)
 

+parser.add_argument("--physical_batch_size", type=int, default=None)
+

 parser.add_argument("--nb_train_samples", type=int, default=None)
 
 parser.add_argument("--nb_test_samples", type=int, default=None)
 
 parser.add_argument("--nb_train_samples", type=int, default=None)
 
 parser.add_argument("--nb_test_samples", type=int, default=None)
 

-parser.add_argument("--optim", type=str, default="adam")
-

 parser.add_argument("--learning_rate", type=float, default=1e-4)
 
 parser.add_argument("--learning_rate", type=float, default=1e-4)
 

-parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
+########################################

 
 

-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)
 
 
 parser.add_argument("--dim_model", type=int, default=None)
 


@@ -73,157 +76,40 @@ parser.add_argument("--nb_blocks", type=int, default=None)
 
 parser.add_argument("--dropout", type=float, default=0.1)
 
 
 parser.add_argument("--dropout", type=float, default=0.1)
 

-parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-
-parser.add_argument("--no_checkpoint", action="store_true", default=False)
-
-parser.add_argument("--overwrite_results", action="store_true", default=False)
-
-parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
-
-##############################
-# picoclvr options
-
-parser.add_argument("--sandbox_level", type=int, default=0)
-
-parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
-
-parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
-
-parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
-
-##############################
-# picoclvr options
-
-parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
-
-parser.add_argument("--picoclvr_height", type=int, default=12)
-
-parser.add_argument("--picoclvr_width", type=int, default=16)
-
-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_width", type=int, default=39)
-
-parser.add_argument("--maze_nb_walls", type=int, default=45)
-
-##############################
-# 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=5)
-
-parser.add_argument("--snake_length", type=int, default=200)
-
-##############################
-# Stack options
-
-parser.add_argument("--stack_nb_steps", type=int, default=100)
-
-parser.add_argument("--stack_nb_stacks", type=int, default=3)
-
-parser.add_argument("--stack_nb_digits", type=int, default=3)
-
-parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
+########################################

 
 

-##############################
-# Expr options
-
-parser.add_argument("--expr_nb_variables", type=int, default=5)
-
-parser.add_argument("--expr_sequence_length", type=int, default=40)
-
-parser.add_argument("--expr_operand_max", type=int, default=9)
-
-parser.add_argument("--expr_result_max", type=int, default=99)
-
-parser.add_argument("--expr_input_file", type=str, default=None)
+parser.add_argument("--deterministic_synthesis", action="store_true", default=False)

 
 

-##############################
-# World options
+parser.add_argument("--nb_gpts", type=int, default=5)

 
 

-parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
+parser.add_argument("--dirty_debug", action="store_true", default=False)

 
 ######################################################################
 
 args = parser.parse_args()
 
 
 ######################################################################
 
 args = parser.parse_args()
 

-assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
-

 if args.result_dir is None:
 if args.result_dir is None:

-    args.result_dir = f"results_{args.task}"
+    args.result_dir = f"results_culture"

 
 ######################################################################
 
 
 ######################################################################
 

-default_task_args = {
-    "sandbox": {
-        "nb_epochs": 50,
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 10000,
-    },
-    "picoclvr": {
-        "nb_epochs": 25,
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "mnist": {
-        "nb_epochs": 25,
-        "batch_size": 10,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "maze": {
-        "nb_epochs": 25,
-        "batch_size": 5,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "snake": {
-        "nb_epochs": 5,
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "stack": {
-        "nb_epochs": 5,
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 1000,
-    },
-    "expr": {
-        "nb_epochs": 40,
-        "batch_size": 25,
-        "nb_train_samples": 1000000,
-        "nb_test_samples": 10000,
-    },
-    "rpl": {
-        "nb_epochs": 40,
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 10000,
-    },
-    "world": {
-        "nb_epochs": 10,
-        "batch_size": 25,
-        "nb_train_samples": 25000,
-        "nb_test_samples": 1000,
-    },
+if args.dirty_debug:
+    accuracy_to_make_c_quizzes = 0.0
+    nb_new_c_quizzes_for_train = 100
+    nb_new_c_quizzes_for_test = 10
+
+######################################################################
+
+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)

 
 ######################################################################
 
 
 ######################################################################
 


@@ -235,6 +121,13 @@ default_model_args = {
         "nb_heads": 2,
         "nb_blocks": 2,
     },
         "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,
     "37M": {
         "dim_model": 512,
         "dim_keys": 64,


@@ -270,9 +163,8 @@ else:
 try:
     os.mkdir(args.result_dir)
 except FileExistsError:
 try:
     os.mkdir(args.result_dir)
 except FileExistsError:

-    if not args.overwrite_results:
-        print(f"result directory {args.result_dir} already exists")
-        exit(1)
+    print(f"result directory {args.result_dir} already exists")
+    exit(1)

 
 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 
 
 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 


@@ -298,363 +190,304 @@ def log_string(s):
     sys.stdout.flush()
 
 
     sys.stdout.flush()
 
 

+log_string(f"argv {' '.join(sys.argv)}")
+

 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
 
 ######################################################################
 
 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
 
 ######################################################################
 

+if args.dirty_debug:
+    args.nb_train_samples = 2500
+    args.nb_test_samples = 100

 
 

-def picoclvr_pruner_horizontal_green(p):
-    return not ("green" in p and ("left" in p or "right" in p))
+if args.physical_batch_size is None:
+    args.physical_batch_size = args.batch_size
+else:
+    assert args.batch_size % args.physical_batch_size == 0
+
+assert args.nb_train_samples % args.batch_size == 0
+assert args.nb_test_samples % args.batch_size == 0
+
+quizz_machine = quizz_machine.QuizzMachine(
+    problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
+    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,
+)

 
 

+######################################################################

 
 

-picoclvr_pruner_train = (
-    picoclvr_pruner_horizontal_green
-    if args.picocvlr_prune_properties in {"train+eval"}
-    else None
-)
+log_string(f"device {device}")

 
 

-picoclvr_pruner_eval = (
-    (lambda p: not picoclvr_pruner_horizontal_green(p))
-    if args.picocvlr_prune_properties in {"train+eval", "eval"}
-    else None
-)
+vocabulary_size = quizz_machine.vocabulary_size()
+
+log_string(f"vocabulary_size {vocabulary_size}")

 
 ######################################################################
 
 
 ######################################################################
 

-if args.task == "sandbox":
-    if args.sandbox_level == 0:
-        problem = tasks.ProblemLevel0(
-            nb_sentences=args.sandbox_levels_nb_items,
-            len_prompt=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
-        )
-    elif args.sandbox_level == 1:
-        problem = tasks.ProblemLevel1(
-            nb_operators=args.sandbox_levels_nb_items,
-            len_source=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
-        )
-    elif args.sandbox_level == 2:
-        problem = tasks.ProblemLevel2(
-            len_source=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
-        )
-    else:
-        raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
-
-    task = tasks.SandBox(
-        problem,
-        # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        logger=log_string,
-        device=device,
-    )
+# Compute the entropy of the training tokens

 
 

-elif args.task == "picoclvr":
-    task = tasks.PicoCLVR(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.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,
+token_count = 0
+for input in quizz_machine.batches(split="train", desc="train-entropy"):
+    token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
+        (0, 1)

     )
     )

+token_probas = token_count / token_count.sum()
+entropy = -torch.xlogy(token_probas, token_probas).sum()
+train_set_perplexity = math.exp(entropy)

 
 

-elif args.task == "mnist":
-    task = tasks.MNIST(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        device=device,
-    )
+######################################################################
+# A bit of paranoia never hurts

 
 

-elif args.task == "maze":
-    task = tasks.Maze(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        height=args.maze_height,
-        width=args.maze_width,
-        nb_walls=args.maze_nb_walls,
-        device=device,
+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(
+        quizz_machine.batches(split="test", desc="test-check"), 25000
+    ):
+        in_train = set()
+        for train_subset in subsets_as_tuples(
+            quizz_machine.batches(split="train", desc="train-check"), 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"

     )
 
     )
 

-elif args.task == "snake":
-    task = tasks.Snake(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.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,
-    )
+    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"

 
 

-elif args.task == "stack":
-    task = tasks.Stack(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.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.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.batch_size,
-        device=device,
-    )
+def one_epoch(model, quizz_machine):
+    optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)

 
 

-elif args.task == "world":
-    task = tasks.World(
-        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,
-        logger=log_string,
-        device=device,
-    )
+    model.train()

 
 

-else:
-    raise ValueError(f"Unknown task {args.task}")
+    nb_train_samples, acc_train_loss = 0, 0.0

 
 

-######################################################################
+    for input in quizz_machine.batches(split="train"):
+        input = input.to(device)

 
 

-log_string(f"device {device}")
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.zero_grad()

 
 

-vocabulary_size = task.vocabulary_size()
+        output = model(mygpt.BracketedSequence(input)).x
+        loss = F.cross_entropy(output.transpose(1, 2), input)
+        acc_train_loss += loss.item() * input.size(0)

 
 

-log_string(f"vocabulary_size {vocabulary_size}")
+        nb_train_samples += input.size(0)

 
 

-##############################
+        loss.backward()

 
 

-model = mygpt.MyGPT(
-    vocabulary_size=vocabulary_size,
-    dim_model=args.dim_model,
-    dim_keys=args.dim_keys,
-    dim_hidden=args.dim_hidden,
-    nb_heads=args.nb_heads,
-    nb_blocks=args.nb_blocks,
-    causal=True,
-    dropout=args.dropout,
-)
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.step()

 
 

-model.to(device)
+    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+
+    log_string(f"train_perplexity {n_epoch} {train_perplexity}")

 
 

-nb_parameters = sum(p.numel() for p in model.parameters())
-log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")

 
 ######################################################################
 
 
 ######################################################################
 

-nb_epochs_finished = 0

 
 

-if args.no_checkpoint:
-    log_string(f"not trying to load checkpoint.")
+def run_tests(model, quizz_machine, deterministic_synthesis):
+    with torch.autograd.no_grad():
+        model.eval()

 
 

-else:
-    try:
-        checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
-        checkpoint = torch.load(checkpoint_name)
-        nb_epochs_finished = checkpoint["nb_epochs_finished"]
-        model.load_state_dict(checkpoint["model_state"])
-        torch.set_rng_state(checkpoint["rng_state"])
-        if torch.cuda.is_available():
-            torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
+        nb_test_samples, acc_test_loss = 0, 0.0
+        nb_samples_accumulated = 0

 
 

-        log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
+        for input in quizz_machine.batches(split="test"):
+            input = input.to(device)

 
 

-    except FileNotFoundError:
-        log_string("starting from scratch.")
+            bs = model(mygpt.BracketedSequence(input))
+            output = bs.x

 
 

-    except:
-        log_string("error when loading the checkpoint.")
-        exit(1)
+            loss = F.cross_entropy(output.transpose(1, 2), input)

 
 

-######################################################################
+            acc_test_loss += loss.item() * input.size(0)

 
 

-if args.task == "expr" and args.expr_input_file is not None:
-    task.produce_results(
-        nb_epochs_finished,
-        model,
-        args.result_dir,
-        log_string,
-        args.deterministic_synthesis,
-        args.expr_input_file,
-    )
+            nb_test_samples += input.size(0)
+
+        main_test_accuracy = quizz_machine.produce_results(
+            n_epoch=n_epoch,
+            model=model,
+            result_dir=args.result_dir,
+            logger=log_string,
+            deterministic_synthesis=deterministic_synthesis,
+        )
+
+        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+
+        log_string(f"test_perplexity {n_epoch} {test_perplexity}")
+
+    model.main_test_accuracy = main_test_accuracy

 
 

-    exit(0)

 
 ######################################################################
 
 
 ######################################################################
 

-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default

 
 

-# Compute the entropy of the training tokens
+def create_c_quizzes(
+    models,
+    quizz_machine,
+    nb_for_train=1000,
+    nb_for_test=100,
+    min_ave_seq_logproba=None,
+):
+    kept = []
+    model_indexes = []
+    sum_logits, sum_nb_c_quizzes = 0, 0
+
+    while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
+        nb_to_generate = nb_for_train + nb_for_test
+
+        if len(model_indexes) == 0:
+            model_indexes = [i.item() for i in torch.randperm(len(models))]
+
+        model = models[model_indexes.pop()]
+
+        new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
+            nb=nb_to_generate,
+            model_for_generation=model,
+            models_for_validation=models,
+            min_ave_seq_logproba=min_ave_seq_logproba,
+            n_epoch=n_epoch,
+            result_dir=args.result_dir,
+            logger=log_string,
+        )

 
 

-token_count = 0
-for input in task.batches(split="train"):
-    token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
-token_probas = token_count / token_count.sum()
-entropy = -torch.xlogy(token_probas, token_probas).sum()
-train_set_perplexity = math.exp(entropy)
+        sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
+        sum_nb_c_quizzes += new_c_quizzes.size(0)

 
 

-##############################
+        to_keep = new_c_quizzes[nb_correct == len(models) - 1]

 
 

-# A bit of paranoia never hurts
+        if args.dirty_debug:
+            to_keep = new_c_quizzes[
+                torch.randint(3, (new_c_quizzes.size(0),), device=new_c_quizzes.device)
+                == 0
+            ]

 
 

-train_examples = {}
+        kept.append(to_keep)

 
 

+        log_string(
+            f"keep c_quizzes {to_keep.size(0)}/{new_c_quizzes.size(0)} ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%) total {sum([ x.size(0) for x in kept])}/{nb_to_generate}"
+        )

 
 

-for input in task.batches(split="train"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        train_examples[x.sum().item()] = x
+    new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]

 
 

-nb_total, nb_collisions = 0, 0
-for input in task.batches(split="test"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        nb_total += 1
-        y = train_examples.get(x.sum().item())
-        if y is not None:
-            if x.size() == y.size() and (x - y).abs().sum() == 0:
-                nb_collisions += 1
+    quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
+    quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)

 
 

-del train_examples
+    quizz_machine.problem.save_quizzes(
+        new_c_quizzes[:72],
+        args.result_dir,
+        f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
+    )

 
 

-log_string(
-    f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
-)
+    return sum_logits / sum_nb_c_quizzes

 
 

-##############################

 
 

-if args.learning_rate_schedule == "cos":
-    learning_rate_schedule = {}
-    for n_epoch in range(args.nb_epochs):
-        u = n_epoch / args.nb_epochs * math.pi
-        learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
-else:
-    u = {
-        int(k): float(v)
-        for k, v in [
-            tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
-        ]
-    }
-
-    learning_rate_schedule = {}
-    learning_rate = args.learning_rate
-    for n_epoch in range(args.nb_epochs):
-        if n_epoch in u:
-            learning_rate = u[n_epoch]
-        learning_rate_schedule[n_epoch] = learning_rate
-
-log_string(f"learning_rate_schedule {learning_rate_schedule}")
+######################################################################

 
 

-##############################
+models = []

 
 

-nb_samples_seen = 0
+for k in range(args.nb_gpts):
+    model = mygpt.MyGPT(
+        vocabulary_size=vocabulary_size,
+        dim_model=args.dim_model,
+        dim_keys=args.dim_keys,
+        dim_hidden=args.dim_hidden,
+        nb_heads=args.nb_heads,
+        nb_blocks=args.nb_blocks,
+        causal=True,
+        dropout=args.dropout,
+    ).to(device)

 
 

-if nb_epochs_finished >= nb_epochs:
-    task.produce_results(
-        nb_epochs_finished,
-        model,
-        args.result_dir,
-        log_string,
-        args.deterministic_synthesis,
-    )
+    model.main_test_accuracy = 0.0
+    model.id = k

 
 

-for n_epoch in range(nb_epochs_finished, nb_epochs):
-    learning_rate = learning_rate_schedule[n_epoch]
+    models.append(model)

 
 

-    log_string(f"learning_rate {learning_rate}")

 
 

-    if args.optim == "sgd":
-        optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
-    elif args.optim == "adam":
-        optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
-    elif args.optim == "adamw":
-        optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
-    else:
-        raise ValueError(f"Unknown optimizer {args.optim}.")
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")

 
 

-    model.train()
+######################################################################

 
 

-    nb_train_samples, acc_train_loss = 0, 0.0
+min_ave_seq_logproba = None

 
 

-    for input in task.batches(split="train"):
-        input = input.to(device)
-        output = model(mygpt.BracketedSequence(input)).x
-        loss = F.cross_entropy(output.transpose(1, 2), input)
-        acc_train_loss += loss.item() * input.size(0)
-        nb_train_samples += input.size(0)
-        nb_samples_seen += input.size(0)
+for n_epoch in range(args.nb_epochs):
+    log_string(f"--- epoch {n_epoch} ----------------------------------------")

 
 

-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
+    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}")

 
 

-    with torch.autograd.no_grad():
-        model.eval()
+    # select the model with lowest accuracy
+    models.sort(key=lambda model: model.main_test_accuracy)
+    model = models[0]

 
 

-        nb_test_samples, acc_test_loss = 0, 0.0
+    log_string(
+        f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+    )

 
 

-        for input in task.batches(split="test"):
-            input = input.to(device)
+    # improve it
+    one_epoch(model, quizz_machine)

 
 

-            output = model(mygpt.BracketedSequence(input)).x
-            loss = F.cross_entropy(output.transpose(1, 2), input)
-            acc_test_loss += loss.item() * input.size(0)
-            nb_test_samples += input.size(0)
+    quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)

 
 

-        train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+    log_string(
+        f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
+    )

 
 

-        log_string(
-            f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
-        )
+    # test it
+    run_tests(model, quizz_machine, deterministic_synthesis=False)

 
 

-        task.produce_results(
-            n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
+    log_string(
+        f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
+    )
+
+    if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
+        ave_seq_logproba = create_c_quizzes(
+            models,
+            quizz_machine,
+            nb_for_train=nb_new_c_quizzes_for_train,
+            nb_for_test=nb_new_c_quizzes_for_test,
+            min_ave_seq_logproba=min_ave_seq_logproba,

         )
 
         )
 

-    checkpoint = {
-        "nb_epochs_finished": n_epoch + 1,
-        "model_state": model.state_dict(),
-        "rng_state": torch.get_rng_state(),
-    }
+        # We keep the first average logits as a reference
+        if min_ave_seq_logproba is None:
+            min_ave_seq_logproba = ave_seq_logproba
+        else:
+            log_string(
+                f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
+            )

 
 

-    if torch.cuda.is_available():
-        checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+        # We update everyone
+        for model in models:
+            run_tests(model, quizz_machine, deterministic_synthesis=False)

 
 

-    checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
-    torch.save(checkpoint, checkpoint_name)
-    log_string(f"saved checkpoint {checkpoint_name}")

 
 ######################################################################
 
 ######################################################################