X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=main.py;h=4a1207d4571af75798390087c248ebed56a069a9;hb=6917d3d52a4b473d31121a471ab98fa114bdb1a6;hp=dace5f2e2276ff4ba2472b90677b253dea58a46e;hpb=798d9526e726b644979cf1124e714f705fdd5966;p=culture.git

diff --git a/main.py b/main.py
index dace5f2..4a1207d 100755
--- a/main.py
+++ b/main.py
@@ -82,12 +82,6 @@ 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("--resume", action="store_true", default=False)
-
-parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
-
 ##############################
 # filetask
 
@@ -207,6 +201,12 @@ if args.result_dir is None:
 ######################################################################
 
 default_task_args = {
+    "world": {
+        "model": "37M",
+        "batch_size": 100,
+        "nb_train_samples": 250000,
+        "nb_test_samples": 10000,
+    },
     "file": {
         "model": "37M",
         "batch_size": 25,
@@ -463,6 +463,17 @@ elif args.task == "byheart":
     )
     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(),
@@ -660,64 +671,28 @@ log_string(f"vocabulary_size {vocabulary_size}")
 
 ##############################
 
-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,
-)
+models = []
+
+for k in range(2):
+    models.append(
+        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)
+    )
 
-model.to(device)
 
-nb_parameters = sum(p.numel() for p in model.parameters())
+nb_parameters = sum(p.numel() for p in models[0].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.")
-
-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"])
-
-        log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
-
-    except FileNotFoundError:
-        log_string("starting from scratch.")
-
-    except:
-        log_string("error when loading the checkpoint.")
-        exit(1)
-
-######################################################################
-
-if args.task == "expr" and args.expr_input_file is not None:
-    task.produce_results(
-        n_epoch=nb_epochs_finished,
-        model=model,
-        result_dir=args.result_dir,
-        logger=log_string,
-        deterministic_synthesis=args.deterministic_synthesis,
-        input_file=args.expr_input_file,
-    )
-
-    exit(0)
-
-######################################################################
-
 # Compute the entropy of the training tokens
 
 token_count = 0
@@ -786,22 +761,12 @@ else:
 
 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 
-##############################
-
-if nb_epochs_finished >= args.nb_epochs:
-    task.produce_results(
-        n_epoch=nb_epochs_finished,
-        model=model,
-        result_dir=args.result_dir,
-        logger=log_string,
-        deterministic_synthesis=args.deterministic_synthesis,
-    )
-
 time_pred_result = None
 
-for n_epoch in range(nb_epochs_finished, args.nb_epochs):
-    learning_rate = learning_rate_schedule[n_epoch]
+######################################################################
+
 
+def one_epoch(model, task, learning_rate):
     log_string(f"learning_rate {learning_rate}")
 
     if args.optim == "sgd":
@@ -815,7 +780,7 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 
     model.train()
 
-    nb_train_samples, acc_train_loss_ar, acc_train_loss_ae = 0, 0.0, 0.0
+    nb_train_samples, acc_train_loss = 0, 0.0
 
     for input in task.batches(split="train"):
         input = input.to(device)
@@ -823,95 +788,129 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
         if nb_train_samples % args.batch_size == 0:
             optimizer.zero_grad()
 
-        if args.autoencoder_weight > 0:
-            bs_ar, bs_ae = model(mygpt.BracketedSequence(input), autoencoder=True)
-            output_ar, output_ae = bs_ar.x, bs_ae.x
-            loss_ar = F.cross_entropy(output_ar.transpose(1, 2), input)
-            loss_ae = F.cross_entropy(output_ae[:, 1:].transpose(1, 2), input[:, :-1])
-        else:
-            output = model(mygpt.BracketedSequence(input)).x
-            loss_ar = F.cross_entropy(output.transpose(1, 2), input)
-            loss_ae = loss_ar.new_full((1,), 0.0)
-
-        acc_train_loss_ar += loss_ar.item() * input.size(0)
-        acc_train_loss_ae += loss_ae.item() * input.size(0)
+        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_ar + args.autoencoder_weight * loss_ae).backward()
+        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} {train_perplexity}")
+
+
+######################################################################
+
+
+def run_tests(model, task, deterministic_synthesis):
     with torch.autograd.no_grad():
         model.eval()
 
-        nb_test_samples, acc_test_loss_ar, acc_test_loss_ae = 0, 0.0, 0.0
+        nb_test_samples, acc_test_loss = 0, 0.0
         nb_samples_accumulated = 0
 
         for input in task.batches(split="test"):
             input = input.to(device)
 
-            if args.autoencoder_weight > 0:
-                bs_ar, bs_ae = model(mygpt.BracketedSequence(input), autoencoder=True)
-                output_ar, output_ae = bs_ar.x, bs_ae.x
-                loss_ae = F.cross_entropy(
-                    output_ae[:, 1:].transpose(1, 2), input[:, :-1]
-                )
-                acc_test_loss_ae += loss_ae.item() * input.size(0)
-            else:
-                bs_ar = model(mygpt.BracketedSequence(input))
-                output_ar = bs_ar.x
+            bs = model(mygpt.BracketedSequence(input))
+            output = bs.x
 
-            loss_ar = F.cross_entropy(output_ar.transpose(1, 2), input)
+            loss = F.cross_entropy(output.transpose(1, 2), input)
 
-            acc_test_loss_ar += loss_ar.item() * input.size(0)
+            acc_test_loss += loss.item() * input.size(0)
 
             nb_test_samples += input.size(0)
 
-        train_ar_perplexity = math.exp(min(100, acc_train_loss_ar / nb_train_samples))
-        test_ar_perplexity = math.exp(min(100, acc_test_loss_ar / nb_test_samples))
-
-        log_string(
-            f"perplexity_ar {n_epoch} train_set {train_set_perplexity} train_prediction {train_ar_perplexity} test_prediction {test_ar_perplexity}"
+        main_test_accuracy = task.produce_results(
+            n_epoch=n_epoch,
+            model=model,
+            result_dir=args.result_dir,
+            logger=log_string,
+            deterministic_synthesis=deterministic_synthesis,
         )
 
-        if args.autoencoder_weight > 0:
-            train_ae_perplexity = math.exp(
-                min(100, acc_train_loss_ae / nb_train_samples)
-            )
-            test_ae_perplexity = math.exp(min(100, acc_test_loss_ae / nb_test_samples))
+        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 
-            log_string(
-                f"perplexity_ae {n_epoch} train_set {train_set_perplexity} train_prediction {train_ae_perplexity} test_prediction {test_ae_perplexity}"
-            )
+        log_string(f"test_perplexity {n_epoch} {test_perplexity}")
+
+    return main_test_accuracy
+
+
+######################################################################
 
-        task.produce_results(
+
+def create_quizzes(
+    model,
+    other_models,
+    task,
+    nb_for_train=1000,
+    nb_for_test=100,
+    nb_runs=10,
+    nb_min_correct=9,
+    nb_max_correct=9,
+):
+    kept = []
+
+    while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
+        new_quizzes, nb_correct = task.create_new_quizzes(
             n_epoch=n_epoch,
-            model=model,
             result_dir=args.result_dir,
             logger=log_string,
-            deterministic_synthesis=args.deterministic_synthesis,
+            nb=4 * (nb_for_train + nb_for_test),
+            model=model,
+            other_models=other_models,
+            nb_runs=nb_runs,
         )
 
-        time_current_result = datetime.datetime.now()
-        if time_pred_result is not None:
-            log_string(
-                f"next_result {time_current_result + (time_current_result - time_pred_result)}"
+        to_keep = new_quizzes[
+            torch.logical_and(
+                nb_correct >= nb_min_correct, nb_correct <= nb_max_correct
             )
-        time_pred_result = time_current_result
+        ]
+        log_string(f"keep {to_keep.size(0)} quizzes")
+        kept.append(to_keep)
 
-    checkpoint = {
-        "nb_epochs_finished": n_epoch + 1,
-        "model_state": model.state_dict(),
-        "rng_state": torch.get_rng_state(),
-    }
+    new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 
-    if torch.cuda.is_available():
-        checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+    task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
+    task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
+
+    task.save_image(
+        new_quizzes[:96],
+        args.result_dir,
+        f"world_new_{n_epoch:04d}.png",
+        log_string,
+    )
+
+
+######################################################################
+
+accuracy_to_make_quizzes = 0.95
+
+for n_epoch in range(nb_epochs_finished, args.nb_epochs):
+    learning_rate = learning_rate_schedule[n_epoch]
 
-    checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
-    torch.save(checkpoint, checkpoint_name)
-    log_string(f"saved checkpoint {checkpoint_name}")
+    for m in models:
+        one_epoch(m, task, learning_rate)
+        test_accuracy = run_tests(m, task, deterministic_synthesis=False)
+
+    if test_accuracy >= accuracy_to_make_quizzes:
+        other_models = models.copy()
+        other_models.remove(model)
+        create_quizzes(other_models, task)
+
+    # --------------------------------------------
+
+    time_current_result = datetime.datetime.now()
+    if time_pred_result is not None:
+        log_string(
+            f"next_result {time_current_result + (time_current_result - time_pred_result)}"
+        )
+    time_pred_result = time_current_result
 
 ######################################################################