X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=main.py;h=1ef01e9c510ba9fbb2df269c19809f1f30d639bb;hb=050976a525fee2d3b824350a3058ab7299a2bd3d;hp=6b46fa0f23b68ca2c88466be37b60c069758bce6;hpb=e0ab20005e2578edff27d4246c6904cf1047ed22;p=culture.git

diff --git a/main.py b/main.py
index 6b46fa0..1ef01e9 100755
--- a/main.py
+++ b/main.py
@@ -12,15 +12,13 @@ from torch import nn
 from torch.nn import functional as F
 
 import ffutils
+
 import mygpt
 import sky, grids, quiz_machine
 
-# world quizzes vs. culture quizzes
-
-######################################################################
+import threading
 
-nb_new_c_quizzes_for_train = 1000
-nb_new_c_quizzes_for_test = 100
+# world quizzes vs. culture quizzes
 
 ######################################################################
 
@@ -33,7 +31,6 @@ else:
 ######################################################################
 
 parser = argparse.ArgumentParser(
-    description="An implementation of GPT with cache.",
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
@@ -43,7 +40,7 @@ parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
-parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
 
 ########################################
 
@@ -81,6 +78,10 @@ parser.add_argument("--deterministic_synthesis", action="store_true", default=Fa
 
 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("--nb_gpts", type=int, default=5)
 
 parser.add_argument("--min_to_validate", type=int, default=None)
@@ -124,17 +125,9 @@ if args.result_dir is None:
 
 ######################################################################
 
-if args.dirty_debug:
-    args.accuracy_to_make_c_quizzes = 0.0
-    args.nb_gpts = 2
-    nb_new_c_quizzes_for_train = 100
-    nb_new_c_quizzes_for_test = 10
-
-######################################################################
-
 default_args = {
     "model": "37M",
-    "batch_size": 100,
+    "batch_size": 25,
     "nb_train_samples": 100000,
     "nb_test_samples": 10000,
 }
@@ -249,10 +242,17 @@ 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,
+        chunk_size=100,
+        nb_threads=args.nb_threads,
     )
     back_accuracy = False
 elif args.problem == "grids":
-    problem = grids.Grids(device=device)
+    problem = grids.Grids(
+        max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
+        chunk_size=100,
+        nb_threads=args.nb_threads,
+    )
     back_accuracy = True
 else:
     raise ValueError
@@ -278,64 +278,56 @@ log_string(f"vocabulary_size {vocabulary_size}")
 
 ######################################################################
 
-# Compute the entropy of the training tokens
-
-token_count = 0
-for input in quiz_machine.batches(split="train", desc="train-entropy"):
-    token_count += F.one_hot(input, num_classes=quiz_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)
 
 ######################################################################
-# A bit of paranoia never hurts
 
-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
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
+    if local_device is None:
+        local_device = device
 
-    nb_test, nb_in_train = 0, 0
-    for test_subset in subsets_as_tuples(
-        quiz_machine.batches(split="test", desc="test-check"), 25000
-    ):
-        in_train = set()
-        for train_subset in subsets_as_tuples(
-            quiz_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)
+    with torch.autograd.no_grad():
+        model.eval().to(local_device)
 
-    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"
-    )
+        nb_test_samples, acc_test_loss = 0, 0.0
+        nb_samples_accumulated = 0
 
-    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"
+        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=None):
+    if local_device is None:
+        local_device = device
 
-def one_epoch(model, quiz_machine):
     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 
-    model.train()
+    model.to(local_device).train()
 
     nb_train_samples, acc_train_loss = 0, 0.0
 
-    for input in quiz_machine.batches(split="train"):
-        input = input.to(device)
+    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()
@@ -355,53 +347,98 @@ def one_epoch(model, quiz_machine):
 
     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 
+    run_tests(model, quiz_machine, deterministic_synthesis=False)
+
+    model.TRAINING_LOCK.release()
+
 
 ######################################################################
 
 
-def run_tests(model, quiz_machine, deterministic_synthesis):
-    with torch.autograd.no_grad():
-        model.eval()
+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))
 
-        nb_test_samples, acc_test_loss = 0, 0.0
-        nb_samples_accumulated = 0
 
-        for input in quiz_machine.batches(split="test"):
-            input = input.to(device)
+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([])
 
-            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)
+def create_c_quizzes(
+    models,
+    quiz_machine,
+    nb_for_train=1000,
+    nb_for_test=100,
+):
+    quizzes_and_logproba_records = []
 
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+    nb_to_create = nb_for_train + nb_for_test
 
-        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,
-        )
+    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)]
 
-def valid_c_quizzes(recorded, criteria):
-    result = [q[criteria(c)] for q, c in recorded]
-    return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
+            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)
 
 
 ######################################################################
 
 
-def create_c_quizzes(
+def create_c_quizzes_(
     models,
     quiz_machine,
     nb_for_train=1000,
@@ -413,8 +450,8 @@ def create_c_quizzes(
 
     # ------------------------------------------------------------
 
-    standard_validity = lambda nb_correct: torch.logical_and(
-        nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
+    standard_validity = lambda nb_correct: (nb_correct >= args.min_to_validate) & (
+        nb_correct <= args.max_to_validate
     )
 
     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
@@ -434,23 +471,34 @@ def create_c_quizzes(
                 temperature=args.generation_temperature,
             )
 
-            nb_correct, seq_logproba = quiz_machine.compute_correctness(
-                c_quizzes,
-                models,
-                bidirectional_validation=args.bidirectional_validation,
-                deterministic_validation=args.deterministic_validation,
-            )
+            # if args.prediction_correctness:
+
+            # else:
+            # logproba = quiz_machine.new(quiz_machine.size(0), len(models))
+            # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)):
+            # for model in models:
+            # l[...] = F.cross_entropy(model(q))
 
-            for n, l in zip(nb_correct, seq_logproba):
-                s = " ".join([str(x.item()) for x in l])
-                logp_file.write(f"{n} {s}\n")
+            c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 
-            if args.dirty_debug:
-                nb_correct = torch.randint(
-                    len(models) + 1, nb_correct.size(), device=c_quizzes.device
+            if c_quizzes.size(0) > 0:
+                nb_correct, seq_logproba = quiz_machine.compute_correctness(
+                    c_quizzes,
+                    models,
+                    bidirectional_validation=args.bidirectional_validation,
+                    deterministic_validation=args.deterministic_validation,
                 )
 
-            quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
+                for n, l in zip(nb_correct, seq_logproba):
+                    s = " ".join([str(x.item()) for x in l])
+                    logp_file.write(f"{n} {s}\n")
+
+                if args.dirty_debug:
+                    nb_correct = torch.randint(
+                        len(models) + 1, nb_correct.size(), device=c_quizzes.device
+                    )
+
+                quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 
             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
             nv = " ".join([str(x.item()) for x in nv])
@@ -499,6 +547,7 @@ def create_c_quizzes(
 models = []
 
 for k in range(args.nb_gpts):
+    log_string(f"creating model {k} and its w_quizzes")
     model = mygpt.MyGPT(
         vocabulary_size=vocabulary_size,
         dim_model=args.dim_model,
@@ -512,6 +561,16 @@ for k in range(args.nb_gpts):
 
     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
+    ).to(device)
+    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)
+    quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 
     models.append(model)
 
@@ -521,36 +580,112 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
+# Compute the entropy of the training tokens
+
+token_count = 0
+for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
+    token_count += F.one_hot(input, num_classes=quiz_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)
+
+######################################################################
+# A bit of paranoia never hurts
+
+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(
+        quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
+    ):
+        in_train = set()
+        for train_subset in subsets_as_tuples(
+            quiz_machine.batches(models[0], 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"
+    )
+
+    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"
+
+######################################################################
+
+nb_new_c_quizzes_for_train = args.nb_train_samples // 50
+nb_new_c_quizzes_for_test = args.nb_test_samples // 50
+
+log_string(
+    f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}"
+)
+
+######################################################################
+
+if args.dirty_debug:
+    args.accuracy_to_make_c_quizzes = 0.0
+    args.nb_gpts = 2
+    nb_new_c_quizzes_for_train = 100
+    nb_new_c_quizzes_for_test = 10
+
+######################################################################
+
 for n_epoch in range(args.nb_epochs):
     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 
     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
     log_string(f"current_test_accuracies {cta}")
 
+    ##################################################
     # Select, improve, and eval the worst model
 
-    weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
+    ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 
-    log_string(
-        f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
-    )
+    weakest_models = ranked_models[: args.nb_gpus]
 
-    one_epoch(weakest_model, quiz_machine)
+    for gpu_id, model in enumerate(weakest_models):
+        model.TRAINING_LOCK.acquire()
 
-    log_string(
-        f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
-    )
+        log_string(
+            f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+        )
 
-    run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
+        threading.Thread(
+            target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
+        ).start()
+
+    for model in weakest_models:
+        model.TRAINING_LOCK.acquire()
+        model.TRAINING_LOCK.release()
+
+    ##################################################
+    # Replace a fraction of the w_quizzes with fresh ones
 
     log_string(
-        f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
+        f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
     )
 
-    # Replace a fraction of the w_quizzes with fresh ones
+    # Renew entirely the train set
 
-    quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+    for model in weakest_models:
+        quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 
+    ##################################################
     # If all the models are good enough, generate new quizzes and
     # re-compute the test errors
 
@@ -562,8 +697,4 @@ for n_epoch in range(args.nb_epochs):
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 
-        for model in models:
-            run_tests(model, quiz_machine, deterministic_synthesis=False)
-
-
 ######################################################################