X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=ca0d1524b04d740c8aaa38fa970b503d868bdba8;hb=d16410119a4e5c1117f7f0fbbe80e3e54f81f28b;hp=9198edc25ce5501faef8e5dd8b74d65e9a22836e;hpb=1eef58fd084437bbcd2041b946b468615e203dd8;p=culture.git

diff --git a/main.py b/main.py
index 9198edc..ca0d152 100755
--- a/main.py
+++ b/main.py
@@ -5,7 +5,7 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
-import math, sys, argparse, time, tqdm, os, datetime
+import math, sys, argparse, time, tqdm, os, datetime, warnings
 
 import torch, torchvision
 from torch import nn
@@ -33,7 +33,7 @@ parser.add_argument(
     "--task",
     type=str,
     default="twotargets",
-    help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
+    help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
 )
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
@@ -46,10 +46,12 @@ parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
 
 ########################################
 
-parser.add_argument("--nb_epochs", type=int, default=25)
+parser.add_argument("--nb_epochs", type=int, default=50)
 
 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)
@@ -82,7 +84,7 @@ parser.add_argument("--deterministic_synthesis", action="store_true", default=Fa
 
 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 
-parser.add_argument("--overwrite_results", action="store_true", default=False)
+parser.add_argument("--resume", action="store_true", default=False)
 
 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 
@@ -144,6 +146,11 @@ parser.add_argument("--snake_nb_colors", type=int, default=5)
 
 parser.add_argument("--snake_length", type=int, default=200)
 
+##############################
+# ByHeart options
+
+parser.add_argument("--byheart_separation", type=int, default=1)
+
 ##############################
 # Stack options
 
@@ -153,7 +160,7 @@ 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)
+parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 
 ##############################
 # Expr options
@@ -175,6 +182,19 @@ parser.add_argument("--mixing_hard", action="store_true", default=False)
 
 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 
+##############################
+# greed options
+
+parser.add_argument("--greed_height", type=int, default=5)
+
+parser.add_argument("--greed_width", type=int, default=7)
+
+parser.add_argument("--greed_T", type=int, default=25)
+
+parser.add_argument("--greed_nb_walls", type=int, default=5)
+
+parser.add_argument("--greed_nb_coins", type=int, default=2)
+
 ######################################################################
 
 args = parser.parse_args()
@@ -199,6 +219,12 @@ default_task_args = {
         "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
+    "world": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 50000,
+        "nb_test_samples": 10000,
+    },
     "byheart": {
         "model": "37M",
         "batch_size": 25,
@@ -289,6 +315,12 @@ default_task_args = {
         "nb_train_samples": 60000,
         "nb_test_samples": 10000,
     },
+    "greed": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 25000,
+        "nb_test_samples": 10000,
+    },
 }
 
 if args.task in default_task_args:
@@ -348,7 +380,7 @@ else:
 try:
     os.mkdir(args.result_dir)
 except FileExistsError:
-    if not args.overwrite_results:
+    if not args.resume:
         print(f"result directory {args.result_dir} already exists")
         exit(1)
 
@@ -403,6 +435,14 @@ picoclvr_pruner_eval = (
 
 ######################################################################
 
+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
+
 if args.task == "file":
     assert (
         args.filetask_train_file is not None and args.filetask_test_file is not None
@@ -412,17 +452,29 @@ if args.task == "file":
         args.filetask_test_file,
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
+        shuffle=True,
         device=device,
     )
     args.max_percents_of_test_in_train = 0
 
 elif args.task == "byheart":
     task = tasks.SandBox(
-        problem=problems.ProblemByHeart(),
+        problem=problems.ProblemByHeart(separation=args.byheart_separation),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
+        logger=log_string,
+        device=device,
+    )
+    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,
     )
@@ -433,7 +485,7 @@ elif args.task == "learnop":
         problem=problems.ProblemLearnOperator(),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         logger=log_string,
         device=device,
     )
@@ -444,7 +496,7 @@ elif args.task == "guessop":
         problem=problems.ProblemGuessOperator(),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         logger=log_string,
         device=device,
     )
@@ -455,7 +507,7 @@ elif args.task == "twotargets":
         problem=problems.ProblemTwoTargets(),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         logger=log_string,
         device=device,
     )
@@ -465,7 +517,7 @@ elif args.task == "memory":
         problem=problems.ProblemMemory(),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         logger=log_string,
         device=device,
     )
@@ -477,7 +529,7 @@ elif args.task == "mixing":
         ),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         logger=log_string,
         device=device,
     )
@@ -487,7 +539,7 @@ elif args.task == "addition":
         problem=problems.ProblemAddition(),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         logger=log_string,
         device=device,
     )
@@ -496,7 +548,7 @@ 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,
+        batch_size=args.physical_batch_size,
         height=args.picoclvr_height,
         width=args.picoclvr_width,
         nb_colors=args.picoclvr_nb_colors,
@@ -510,7 +562,7 @@ 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,
+        batch_size=args.physical_batch_size,
         device=device,
     )
 
@@ -518,18 +570,18 @@ 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,
+        batch_size=args.physical_batch_size,
         height=args.maze_height,
         width=args.maze_width,
         nb_walls=args.maze_nb_walls,
-        device=device,
+        device="cpu",
     )
 
 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,
+        batch_size=args.physical_batch_size,
         height=args.snake_height,
         width=args.snake_width,
         nb_colors=args.snake_nb_colors,
@@ -542,7 +594,7 @@ 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,
+        batch_size=args.physical_batch_size,
         logger=log_string,
         nb_steps=args.stack_nb_steps,
         nb_stacks=args.stack_nb_stacks,
@@ -559,7 +611,7 @@ elif args.task == "expr":
         sequence_length=args.expr_sequence_length,
         operand_max=args.expr_operand_max,
         result_max=args.expr_result_max,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         device=device,
     )
 
@@ -567,7 +619,7 @@ 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,
+        batch_size=args.physical_batch_size,
         nb_starting_values=args.rpl_nb_starting_values,
         max_input=args.rpl_max_input,
         prog_len=args.rpl_prog_len,
@@ -581,7 +633,7 @@ elif args.task == "grid":
     task = tasks.Grid(
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         size=args.grid_size,
         fraction_play=args.grid_fraction_play,
         logger=log_string,
@@ -592,12 +644,26 @@ elif args.task == "qmlp":
     task = tasks.QMLP(
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
+        batch_size=args.physical_batch_size,
         result_dir=args.result_dir,
         logger=log_string,
         device=device,
     )
 
+elif args.task == "greed":
+    task = tasks.Greed(
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.physical_batch_size,
+        height=args.greed_height,
+        width=args.greed_width,
+        T=args.greed_T,
+        nb_walls=args.greed_nb_walls,
+        nb_coins=args.greed_nb_coins,
+        logger=log_string,
+        device=device,
+    )
+
 else:
     raise ValueError(f"Unknown task {args.task}")
 
@@ -669,12 +735,10 @@ if args.task == "expr" and args.expr_input_file is not None:
 
 ######################################################################
 
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
-
 # Compute the entropy of the training tokens
 
 token_count = 0
-for input in task.batches(split="train"):
+for input in task.batches(split="train", desc="train-entropy"):
     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()
@@ -696,9 +760,13 @@ if args.max_percents_of_test_in_train >= 0:
         yield s
 
     nb_test, nb_in_train = 0, 0
-    for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
+    for test_subset in subsets_as_tuples(
+        task.batches(split="test", desc="test-check"), 25000
+    ):
         in_train = set()
-        for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
+        for train_subset in subsets_as_tuples(
+            task.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)
@@ -737,9 +805,7 @@ log_string(f"learning_rate_schedule {learning_rate_schedule}")
 
 ##############################
 
-nb_samples_seen = 0
-
-if nb_epochs_finished >= nb_epochs:
+if nb_epochs_finished >= args.nb_epochs:
     task.produce_results(
         n_epoch=nb_epochs_finished,
         model=model,
@@ -750,9 +816,10 @@ if nb_epochs_finished >= nb_epochs:
 
 time_pred_result = None
 
-for n_epoch in range(nb_epochs_finished, 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":
@@ -770,50 +837,114 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
 
     for input in task.batches(split="train"):
         input = input.to(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)
-        nb_samples_seen += input.size(0)
 
-        optimizer.zero_grad()
         loss.backward()
-        optimizer.step()
 
+        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 = 0, 0.0
+        nb_samples_accumulated = 0
 
         for input in task.batches(split="test"):
             input = input.to(device)
 
-            output = model(mygpt.BracketedSequence(input)).x
+            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)
 
-        train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+            acc_test_loss += loss.item() * input.size(0)
 
-        log_string(
-            f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
-        )
+            nb_test_samples += input.size(0)
 
-        task.produce_results(
+        main_test_accuracy = task.produce_results(
             n_epoch=n_epoch,
             model=model,
             result_dir=args.result_dir,
             logger=log_string,
-            deterministic_synthesis=args.deterministic_synthesis,
+            deterministic_synthesis=deterministic_synthesis,
         )
 
-        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)}"
+        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+
+        log_string(f"test_perplexity {n_epoch} {test_perplexity}")
+
+    return main_test_accuracy
+
+
+######################################################################
+
+for n_epoch in range(nb_epochs_finished, args.nb_epochs):
+    learning_rate = learning_rate_schedule[n_epoch]
+
+    one_epoch(model, task, learning_rate)
+
+    test_accuracy = run_tests(model, task, deterministic_synthesis=False)
+
+    # --------------------------------------------
+
+    if test_accuracy >= 0.8:
+        nb_for_train, nb_for_test = 1000, 100
+        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,
+                result_dir=args.result_dir,
+                logger=log_string,
+                nb=nb_required,
+                model=model,
+                nb_runs=10,
             )
-        time_pred_result = time_current_result
+
+            to_keep = new_quizzes[torch.logical_and(nb_correct >= 8, nb_correct < 10)]
+            log_string(f"keep {to_keep.size(0)} quizzes")
+            kept.append(to_keep)
+
+        new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+
+        task.store_new_quizzes(new_quizzes[:nb_for_train], train=True)
+        task.store_new_quizzes(new_quizzes[nb_for_train:], train=False)
+
+        task.save_image(
+            new_quizzes[:96],
+            args.result_dir,
+            f"world_new_{n_epoch:04d}.png",
+            log_string,
+        )
+
+    # --------------------------------------------
+
+    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
+
+    # --------------------------------------------
 
     checkpoint = {
         "nb_epochs_finished": n_epoch + 1,