X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=6c4099f51513c5e1550b3820466d303639c6e9d4;hb=5b7022591f48382ec84b1dda17297b1ed15166d7;hp=a7338c7bdf5402cfa6daded72f594c8a5f88c62d;hpb=a86dff174205c38d8e90d0d89ea399a6afb36359;p=culture.git

diff --git a/main.py b/main.py
index a7338c7..6c4099f 100755
--- a/main.py
+++ b/main.py
@@ -20,16 +20,6 @@ import threading
 
 import torch.multiprocessing as mp
 
-# world quizzes vs. culture quizzes
-
-######################################################################
-
-if torch.cuda.is_available():
-    device = torch.device("cuda")
-    torch.backends.cuda.matmul.allow_tf32 = True
-else:
-    device = torch.device("cpu")
-
 ######################################################################
 
 parser = argparse.ArgumentParser(
@@ -42,6 +32,8 @@ parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
+parser.add_argument("--resume", action="store_true", default=False)
+
 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
 
 ########################################
@@ -82,15 +74,15 @@ parser.add_argument("--problem", type=str, default="grids")
 
 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("--min_to_validate", type=int, default=None)
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
 
-parser.add_argument("--max_to_validate", type=int, default=None)
+parser.add_argument("--proba_understands", type=float, default=0.99)
 
-parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+parser.add_argument("--proba_not_understands", type=float, default=0.5)
 
 parser.add_argument("--generation_temperature", type=float, default=2.0)
 
@@ -98,6 +90,19 @@ 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)
@@ -112,12 +117,6 @@ parser.add_argument("--sky_speed", type=int, default=3)
 
 args = parser.parse_args()
 
-if args.min_to_validate is None:
-    args.min_to_validate = args.nb_gpts - 1
-
-if args.max_to_validate is None:
-    args.max_to_validate = args.nb_gpts - 1
-
 if args.result_dir is None:
     args.result_dir = f"results_culture"
 
@@ -183,11 +182,15 @@ else:
 
 ######################################################################
 
-try:
-    os.mkdir(args.result_dir)
-except FileExistsError:
-    print(f"result directory {args.result_dir} already exists")
-    exit(1)
+if args.resume:
+    assert os.path.isdir(args.result_dir)
+
+else:
+    try:
+        os.mkdir(args.result_dir)
+    except FileExistsError:
+        print(f"result directory {args.result_dir} already exists")
+        exit(1)
 
 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 
@@ -213,6 +216,10 @@ def log_string(s):
     sys.stdout.flush()
 
 
+now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
+
+os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py")
+
 log_string(f"argv {' '.join(sys.argv)}")
 
 for n in vars(args):
@@ -221,6 +228,19 @@ for n in vars(args):
 
 ######################################################################
 
+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
@@ -240,21 +260,24 @@ if args.problem == "sky":
         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(
-        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,
+        tasks=args.grids_tasks,
     )
     back_accuracy = True
 else:
     raise ValueError
 
+problem.save_some_examples(args.result_dir)
+
 quiz_machine = quiz_machine.QuizMachine(
     problem=problem,
     nb_train_samples=args.nb_train_samples,
@@ -263,12 +286,12 @@ quiz_machine = quiz_machine.QuizMachine(
     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()
 
@@ -277,13 +300,7 @@ log_string(f"vocabulary_size {vocabulary_size}")
 ######################################################################
 
 
-######################################################################
-
-
-def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
-    if local_device is None:
-        local_device = device
-
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
     with torch.autograd.no_grad():
         model.eval().to(local_device)
 
@@ -304,7 +321,7 @@ def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
 
         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 
-        log_string(f"test_perplexity {n_epoch} {test_perplexity}")
+        log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
 
         model.main_test_accuracy = quiz_machine.produce_results(
             n_epoch=n_epoch,
@@ -314,14 +331,11 @@ def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
         )
 
 
-def one_epoch(model, quiz_machine, local_device=None):
-    if local_device is None:
-        local_device = device
+def one_epoch(model, quiz_machine, local_device=main_device):
+    model.to(local_device).train()
 
     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"):
@@ -343,11 +357,11 @@ def one_epoch(model, quiz_machine, local_device=None):
 
     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}")
+    log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
 
     run_tests(model, quiz_machine, deterministic_synthesis=False)
 
-    model.TRAINING_LOCK.release()
+    model.to(main_device)
 
 
 ######################################################################
@@ -355,79 +369,99 @@ def one_epoch(model, quiz_machine, local_device=None):
 
 def standard_validity(logproba):
     l = logproba.sort(dim=-1).values
-    return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
+    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 valid_quizzes_and_logprobas(recorded, criteria):
+    validated_quizzes, validated_logprobas = [], []
+    for q, lp in recorded:
+        validated_indices = criteria(lp)
+        validated_quizzes.append(q[validated_indices])
+        validated_logprobas.append(lp[validated_indices])
 
+    if len(validated_quizzes) > 0:
+        return torch.cat(validated_quizzes, dim=0), torch.cat(
+            validated_logprobas, dim=0
+        )
+    else:
+        return None, None
 
-######################################################################
 
+######################################################################
 
-def create_c_quizzes(
-    models,
-    quiz_machine,
-    nb_for_train=1000,
-    nb_for_test=100,
-):
-    quizzes_and_logproba_records = []
 
+def create_c_quizzes(models, quiz_machine, nb_for_train=1000, nb_for_test=100):
     nb_to_create = nb_for_train + nb_for_test
 
-    # ------------------------------------------------------------
+    recorded_quizzes_logprobas = []
 
-    file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
+    nb_validated = 0
 
-    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
+    while nb_validated < nb_to_create:
+        model_for_generation = models[torch.randint(len(models), (1,))]
 
-            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 = 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)]
 
-            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)
+            recorded_quizzes_logprobas.append((c_quizzes, logproba))
 
-            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))
+            validated_quizzes, validated_logprobas = valid_quizzes_and_logprobas(
+                recorded_quizzes_logprobas, standard_validity
+            )
 
-            nb_validated = valid_c_quizzes(
-                quizzes_and_logproba_records, standard_validity
-            ).size(0)
+            if validated_quizzes is not None:
+                nb_validated = validated_quizzes.size(0)
 
-            log_string(
-                f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
-            )
+        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(validated_quizzes)
+    quiz_machine.store_c_quizzes(validated_quizzes[:nb_for_train], for_train=True)
+    quiz_machine.store_c_quizzes(
+        validated_quizzes[nb_for_train:nb_to_create], for_train=False
+    )
+
+    ######################################################################
+    # save the log probas
 
-    quiz_machine.reverse_random_half_in_place(new_c_quizzes)
+    file_name = os.path.join(
+        args.result_dir, f"culture_c_quiz_all_{n_epoch:04d}_logp.dat"
+    )
 
-    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)
+    with open(file_name, "w") as logp_file:
+        for _, ll in recorded_quizzes_logprobas:
+            for l in ll:
+                s = " ".join([str(x.item()) for x in l])
+                logp_file.write(s + "\n")
 
-    # save a bunch of images to investigate what quizzes with a
-    # certain nb of correct predictions look like
+    ######################################################################
+    # save images with their logprobas
 
-    q = new_c_quizzes[:72]
+    vq = validated_quizzes[:72]
+    vl = validated_logprobas[:72]
 
-    if q.size(0) > 0:
-        quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
+    if vq.size(0) > 0:
+        prefix = f"culture_c_quiz_{n_epoch:04d}"
+
+        file_name = os.path.join(args.result_dir, prefix + "_logp.dat")
+        with open(file_name, "w") as logp_file:
+            for l in vl:
+                s = " ".join([str(x.item()) for x in l])
+                logp_file.write(s + "\n")
+
+        quiz_machine.save_quiz_illustrations(args.result_dir, prefix, vq)
 
 
 ######################################################################
@@ -445,11 +479,10 @@ for k in range(args.nb_gpts):
         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.TRAINING_LOCK = threading.Lock()
 
     model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
@@ -458,6 +491,35 @@ for k in range(args.nb_gpts):
 
     models.append(model)
 
+######################################################################
+
+if args.resume:
+    try:
+        for model in models:
+            filename = f"gpt_{model.id:03d}.pth"
+
+            try:
+                d = torch.load(os.path.join(args.result_dir, filename))
+                model.load_state_dict(d[0])
+                model.main_test_accuracy = d[1]
+                log_string(f"successfully loaded {filename}")
+            except FileNotFoundError:
+                log_string(f"cannot find {filename}")
+                pass
+
+        try:
+            filename = "c_quizzes.pth"
+            quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
+            log_string(f"successfully loaded {filename}")
+        except FileNotFoundError:
+            log_string(f"cannot find {filename}")
+            pass
+
+    except:
+        log_string(f"error when loading {filename}.")
+        exit(1)
+
+######################################################################
 
 nb_parameters = sum(p.numel() for p in models[0].parameters())
 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
@@ -545,31 +607,35 @@ for n_epoch in range(args.nb_epochs):
 
     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)]
 
-    for gpu_id, model in enumerate(weakest_models):
-        model.TRAINING_LOCK.acquire()
+    threads = []
 
-        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}")
+
+        t = threading.Thread(
+            target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
         )
 
-        threading.Thread(
-            target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
-        ).start()
+        threads.append(t)
 
-    for model in weakest_models:
-        model.TRAINING_LOCK.acquire()
-        model.TRAINING_LOCK.release()
+        t.start()
 
-    ##################################################
-    # Replace a fraction of the w_quizzes with fresh ones
+    for t in threads:
+        t.join()
 
-    log_string(
-        f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
-    )
+    # Save the models to disk
 
-    # Renew entirely the train set
+    for model in weakest_models:
+        filename = f"gpt_{model.id:03d}.pth"
+        torch.save(
+            (model.state_dict(), model.main_test_accuracy),
+            os.path.join(args.result_dir, filename),
+        )
+        log_string(f"wrote {filename}")
+
+    # Renew the training samples
 
     for model in weakest_models:
         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
@@ -586,4 +652,8 @@ for n_epoch in range(args.nb_epochs):
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 
+        filename = "c_quizzes.pth"
+        quiz_machine.save_c_quizzes(os.path.join(args.result_dir, filename))
+        log_string(f"wrote {filename}")
+
 ######################################################################