Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index 0a266a8..63819f2 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -16,190 +16,14 @@ import ffutils
 import mygpt
 import sky, grids, quiz_machine
 
 import mygpt
 import sky, grids, quiz_machine
 

-import torch.multiprocessing as mp
+import threading

 
 

-# mp.set_start_method('spawn')
+import torch.multiprocessing as mp

 
 # world quizzes vs. culture quizzes
 
 ######################################################################
 
 
 # world quizzes vs. culture quizzes
 
 ######################################################################
 

-
-def log_string(s):
-    t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
-
-    if log_file is not None:
-        log_file.write(t + s + "\n")
-        log_file.flush()
-
-    print(t + s)
-    sys.stdout.flush()
-
-
-######################################################################
-
-
-def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
-    if local_device is None:
-        local_device = device
-
-    with torch.autograd.no_grad():
-        model.eval().to(local_device)
-
-        nb_test_samples, acc_test_loss = 0, 0.0
-        nb_samples_accumulated = 0
-
-        for input in quiz_machine.batches(model, split="test"):
-            input = input.to(local_device)
-
-            bs = model(mygpt.BracketedSequence(input))
-            output = bs.x
-
-            loss = F.cross_entropy(output.transpose(1, 2), input)
-
-            acc_test_loss += loss.item() * input.size(0)
-
-            nb_test_samples += input.size(0)
-
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
-
-        log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
-
-        model.main_test_accuracy = quiz_machine.produce_results(
-            n_epoch=n_epoch,
-            model=model,
-            result_dir=args.result_dir,
-            deterministic_synthesis=deterministic_synthesis,
-        )
-
-
-def one_epoch(model, quiz_machine, local_device=None):
-    if local_device is None:
-        local_device = device
-
-    optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
-
-    model.to(local_device).train()
-
-    nb_train_samples, acc_train_loss = 0, 0.0
-
-    for input in quiz_machine.batches(model, split="train"):
-        input = input.to(local_device)
-
-        if nb_train_samples % args.batch_size == 0:
-            optimizer.zero_grad()
-
-        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)
-
-        loss.backward()
-
-        if nb_train_samples % args.batch_size == 0:
-            optimizer.step()
-
-    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-
-    log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
-
-    run_tests(model, quiz_machine, deterministic_synthesis=False)
-
-
-######################################################################
-
-
-def standard_validity(logproba):
-    l = logproba.sort(dim=-1).values
-    return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
-    # warnings.warn("TEST!!!", RuntimeWarning)
-    # print(l.exp())
-    # return (l[:, 0] < math.log(0.99))
-
-
-def valid_c_quizzes(recorded, criteria):
-    result = [q[criteria(lp)] for q, lp in recorded]
-    return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
-
-
-######################################################################
-
-
-def create_c_quizzes(
-    models,
-    quiz_machine,
-    nb_for_train=1000,
-    nb_for_test=100,
-):
-    quizzes_and_logproba_records = []
-
-    nb_to_create = nb_for_train + nb_for_test
-
-    # ------------------------------------------------------------
-
-    file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
-
-    with open(file_name, "w") as logp_file:
-        while (
-            valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
-            < nb_to_create
-        ):
-            # Select a model at random to generate the new quizzes
-
-            model_for_generation = models[torch.randint(len(models), (1,))]
-
-            c_quizzes = quiz_machine.generate_quizzes(
-                nb_to_create,
-                model_for_generation=model_for_generation,
-                temperature=args.generation_temperature,
-            )
-
-            c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
-
-            if c_quizzes.size(0) > 0:
-                logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
-                for l in logproba:
-                    s = " ".join([str(x.item()) for x in l])
-                    logp_file.write(s + "\n")
-                quizzes_and_logproba_records.append((c_quizzes, logproba))
-
-            nb_validated = valid_c_quizzes(
-                quizzes_and_logproba_records, standard_validity
-            ).size(0)
-
-            log_string(
-                f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
-            )
-
-    # store the new c_quizzes which have been validated
-
-    new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
-
-    quiz_machine.reverse_random_half_in_place(new_c_quizzes)
-
-    quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
-    quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
-
-    # save a bunch of images to investigate what quizzes with a
-    # certain nb of correct predictions look like
-
-    q = new_c_quizzes[:72]
-
-    if q.size(0) > 0:
-        quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
-
-
-######################################################################
-
-if torch.cuda.is_available():
-    device = torch.device("cuda")
-    torch.backends.cuda.matmul.allow_tf32 = True
-else:
-    device = torch.device("cpu")
-
-######################################################################
-

 parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )


@@ -250,7 +74,7 @@ parser.add_argument("--problem", type=str, default="grids")
 
 parser.add_argument("--nb_threads", type=int, default=1)
 
 
 parser.add_argument("--nb_threads", type=int, default=1)
 

-parser.add_argument("--nb_gpus", type=int, default=1)
+parser.add_argument("--gpus", type=str, default="all")

 
 parser.add_argument("--nb_gpts", type=int, default=5)
 
 
 parser.add_argument("--nb_gpts", type=int, default=5)
 


@@ -258,7 +82,11 @@ parser.add_argument("--min_to_validate", type=int, default=None)
 
 parser.add_argument("--max_to_validate", type=int, default=None)
 
 
 parser.add_argument("--max_to_validate", type=int, default=None)
 

-parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.9)
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+
+parser.add_argument("--proba_understands", type=float, default=0.99)
+
+parser.add_argument("--proba_not_understands", type=float, default=0.5)

 
 parser.add_argument("--generation_temperature", type=float, default=2.0)
 
 
 parser.add_argument("--generation_temperature", type=float, default=2.0)
 


@@ -266,6 +94,19 @@ parser.add_argument("--dirty_debug", action="store_true", default=False)
 
 ######################################################################
 
 
 ######################################################################
 

+grids_tasks = ", ".join(
+    [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
+)
+
+parser.add_argument(
+    "--grids_tasks",
+    type=str,
+    default=None,
+    help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
+)
+
+######################################################################
+

 parser.add_argument("--sky_height", type=int, default=6)
 
 parser.add_argument("--sky_width", type=int, default=8)
 parser.add_argument("--sky_height", type=int, default=6)
 
 parser.add_argument("--sky_width", type=int, default=8)


@@ -359,11 +200,6 @@ except FileExistsError:
 
 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")
 

-log_string(f"argv {' '.join(sys.argv)}")
-
-for n in vars(args):
-    log_string(f"args.{n} {getattr(args, n)}")
-

 if args.seed >= 0:
     # torch.backends.cudnn.deterministic = True
     # torch.backends.cudnn.benchmark = False
 if args.seed >= 0:
     # torch.backends.cudnn.deterministic = True
     # torch.backends.cudnn.benchmark = False


@@ -374,6 +210,39 @@ if args.seed >= 0:
 
 ######################################################################
 
 
 ######################################################################
 

+
+def log_string(s):
+    t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
+
+    if log_file is not None:
+        log_file.write(t + s + "\n")
+        log_file.flush()
+
+    print(t + s)
+    sys.stdout.flush()
+
+
+log_string(f"argv {' '.join(sys.argv)}")
+
+for n in vars(args):
+    log_string(f"args.{n} {getattr(args, n)}")
+
+
+######################################################################
+
+if args.gpus == "all":
+    gpus_idx = range(torch.cuda.device_count())
+else:
+    gpus_idx = [int(k) for k in args.gpus.split(",")]
+
+gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
+
+if torch.cuda.is_available():
+    main_device = gpus[0]
+else:
+    assert len(gpus) == 0
+    main_device = torch.device("cpu")
+

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


@@ -393,16 +262,17 @@ if args.problem == "sky":
         nb_birds=args.sky_nb_birds,
         nb_iterations=args.sky_nb_iterations,
         speed=args.sky_speed,
         nb_birds=args.sky_nb_birds,
         nb_iterations=args.sky_nb_iterations,
         speed=args.sky_speed,

-        max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
+        max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,

         chunk_size=100,
         nb_threads=args.nb_threads,
     )
     back_accuracy = False
 elif args.problem == "grids":
     problem = grids.Grids(
         chunk_size=100,
         nb_threads=args.nb_threads,
     )
     back_accuracy = False
 elif args.problem == "grids":
     problem = grids.Grids(

-        max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
+        max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,

         chunk_size=100,
         nb_threads=args.nb_threads,
         chunk_size=100,
         nb_threads=args.nb_threads,

+        tasks=args.grids_tasks,

     )
     back_accuracy = True
 else:
     )
     back_accuracy = True
 else:


@@ -418,12 +288,12 @@ quiz_machine = quiz_machine.QuizMachine(
     batch_size=args.physical_batch_size,
     result_dir=args.result_dir,
     logger=log_string,
     batch_size=args.physical_batch_size,
     result_dir=args.result_dir,
     logger=log_string,

-    device=device,
+    device=main_device,

 )
 
 ######################################################################
 
 )
 
 ######################################################################
 

-log_string(f"device {device}")
+log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")

 
 vocabulary_size = quiz_machine.vocabulary_size()
 
 
 vocabulary_size = quiz_machine.vocabulary_size()
 


@@ -431,6 +301,153 @@ log_string(f"vocabulary_size {vocabulary_size}")
 
 ######################################################################
 
 
 ######################################################################
 

+
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
+    with torch.autograd.no_grad():
+        model.eval().to(local_device)
+
+        nb_test_samples, acc_test_loss = 0, 0.0
+        nb_samples_accumulated = 0
+
+        for input in quiz_machine.batches(model, split="test"):
+            input = input.to(local_device)
+
+            bs = model(mygpt.BracketedSequence(input))
+            output = bs.x
+
+            loss = F.cross_entropy(output.transpose(1, 2), input)
+
+            acc_test_loss += loss.item() * input.size(0)
+
+            nb_test_samples += input.size(0)
+
+        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 = quiz_machine.produce_results(
+            n_epoch=n_epoch,
+            model=model,
+            result_dir=args.result_dir,
+            deterministic_synthesis=deterministic_synthesis,
+        )
+
+
+def one_epoch(model, quiz_machine, local_device=main_device):
+    optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
+
+    model.to(local_device).train()
+
+    nb_train_samples, acc_train_loss = 0, 0.0
+
+    for input in quiz_machine.batches(model, split="train"):
+        input = input.to(local_device)
+
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.zero_grad()
+
+        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)
+
+        loss.backward()
+
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.step()
+
+    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+
+    log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
+
+    run_tests(model, quiz_machine, deterministic_synthesis=False)
+
+
+######################################################################
+
+
+def standard_validity(logproba):
+    l = logproba.sort(dim=-1).values
+    return (l[:, 0] < math.log(args.proba_not_understands)) & (
+        l[:, 1] > math.log(args.proba_understands)
+    )
+
+
+def valid_c_quizzes(recorded, criteria):
+    result = [q[criteria(lp)] for q, lp in recorded]
+    return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
+
+
+######################################################################
+
+
+def create_c_quizzes(
+    models,
+    quiz_machine,
+    nb_for_train=1000,
+    nb_for_test=100,
+):
+    quizzes_and_logproba_records = []
+
+    nb_to_create = nb_for_train + nb_for_test
+
+    # ------------------------------------------------------------
+
+    file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
+
+    with open(file_name, "w") as logp_file:
+        while (
+            valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
+            < nb_to_create
+        ):
+            # Select a model at random to generate the new quizzes
+
+            model_for_generation = models[torch.randint(len(models), (1,))]
+
+            c_quizzes = quiz_machine.generate_quizzes(
+                nb_to_create,
+                model_for_generation=model_for_generation,
+                temperature=args.generation_temperature,
+            )
+
+            c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
+
+            if c_quizzes.size(0) > 0:
+                logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
+                for l in logproba:
+                    s = " ".join([str(x.item()) for x in l])
+                    logp_file.write(s + "\n")
+                quizzes_and_logproba_records.append((c_quizzes, logproba))
+
+            nb_validated = valid_c_quizzes(
+                quizzes_and_logproba_records, standard_validity
+            ).size(0)
+
+            log_string(
+                f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
+            )
+
+    # store the new c_quizzes which have been validated
+
+    new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
+
+    quiz_machine.reverse_random_half_in_place(new_c_quizzes)
+
+    quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
+    quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
+
+    # save a bunch of images to investigate what quizzes with a
+    # certain nb of correct predictions look like
+
+    q = new_c_quizzes[:72]
+
+    if q.size(0) > 0:
+        quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
+
+
+######################################################################
+

 models = []
 
 for k in range(args.nb_gpts):
 models = []
 
 for k in range(args.nb_gpts):


@@ -444,18 +461,14 @@ for k in range(args.nb_gpts):
         nb_blocks=args.nb_blocks,
         causal=True,
         dropout=args.dropout,
         nb_blocks=args.nb_blocks,
         causal=True,
         dropout=args.dropout,

-    ).to(device)
+    ).to(main_device)

 
     model.main_test_accuracy = 0.0
     model.id = k
 
 
     model.main_test_accuracy = 0.0
     model.id = k
 

-    model.train_w_quizzes = quiz_machine.generate_token_sequences(
-        args.nb_train_samples
-    ).to(device)
+    model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)

     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)

-    model.test_w_quizzes = quiz_machine.generate_token_sequences(
-        args.nb_test_samples
-    ).to(device)
+    model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)

     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 
     models.append(model)
     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 
     models.append(model)


@@ -531,7 +544,7 @@ if args.dirty_debug:
 
     def standard_validity(logproba):
         l = logproba.sort(dim=-1).values
 
     def standard_validity(logproba):
         l = logproba.sort(dim=-1).values

-        return l[:, 0] < math.log(0.99)
+        return l[:, 0] < math.log(0.5)

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


@@ -543,38 +556,37 @@ for n_epoch in range(args.nb_epochs):
     log_string(f"current_test_accuracies {cta}")
 
     ##################################################
     log_string(f"current_test_accuracies {cta}")
 
     ##################################################

-    # Select, improve, and eval the worst models
+    # Select, improve, and eval the worst model

 
     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 
 
     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 

-    weakest_models = ranked_models[: args.nb_gpus]
+    weakest_models = ranked_models[: len(gpus)]

 
 

-    processes = []
+    threads = []

 
 

-    for gpu_id, model in enumerate(weakest_models):
-        log_string(
-            f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
-        )
+    for gpu, model in zip(gpus, weakest_models):
+        log_string(f"training model {model.id}")

 
 

-        process = mp.Process(
-            target=one_epoch, args=(model, quiz_machine, f"cuda:{gpu_id}")
+        t = threading.Thread(
+            target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)

         )
 
         )
 

-        processes.append(process)
+        threads.append(t)

 
 

-    for process in processes:
-        process.start()
+        t.start()

 
 

-    for process in processes:
-        process.join()
+    for t in threads:
+        t.join()

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

-    # Renew the train sets
+    # Replace a fraction of the w_quizzes with fresh ones

 
     log_string(
         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
     )
 
 
     log_string(
         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
     )
 

+    # Renew entirely the train set
+

     for model in weakest_models:
         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 
     for model in weakest_models:
         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)