Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index efcc0dd..5dceefc 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -5,219 +5,121 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
 
 # Written by Francois Fleuret <francois@fleuret.org>
 

-# torch.backends.cuda.matmul.allow_tf23
-# torch.autocast(torch.bfloat16)
-
-import math, sys, argparse, time, tqdm, os
+import math, sys, argparse, time, tqdm, os, datetime, warnings, copy

 
 import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 
 
 import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 

-import ffutils
-import mygpt, tasks
+import ffutils, grids, attae

 
 

-######################################################################
+import threading, subprocess

 
 

-if torch.cuda.is_available():
-    device = torch.device("cuda")
-    torch.backends.cuda.matmul.allow_tf32 = True
-else:
-    device = torch.device("cpu")
+# import torch.multiprocessing as mp
+
+torch.set_float32_matmul_precision("high")
+
+# torch.set_default_dtype(torch.bfloat16)

 
 ######################################################################
 
 parser = argparse.ArgumentParser(
 
 ######################################################################
 
 parser = argparse.ArgumentParser(

-    description="An implementation of GPT with cache.",

     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 

-parser.add_argument(
-    "--task",
-    type=str,
-    default="sandbox",
-    help="sandbox, picoclvr, mnist, maze, snake, stack, expr, world",
-)
-
-parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
+parser.add_argument("--log_filename", type=str, default="train.log")

 
 parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
 
 parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 

-parser.add_argument("--nb_epochs", type=int, default=None)
-
-parser.add_argument("--batch_size", type=int, default=None)
+parser.add_argument("--resume", action="store_true", default=False)

 
 

-parser.add_argument("--nb_train_samples", type=int, default=None)
+# ----------------------------------

 
 

-parser.add_argument("--nb_test_samples", type=int, default=None)
+parser.add_argument("--nb_epochs", type=int, default=10000)

 
 

-parser.add_argument("--optim", type=str, default="adam")
+parser.add_argument("--batch_size", type=int, default=25)

 
 

-parser.add_argument("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--train_batch_size", type=int, default=None)

 
 

-parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
+parser.add_argument("--eval_batch_size", type=int, default=25)

 
 

-parser.add_argument("--model", type=str, default="37M")
+parser.add_argument("--nb_train_samples", type=int, default=50000)

 
 

-parser.add_argument("--dim_model", type=int, default=None)
+parser.add_argument("--nb_test_samples", type=int, default=2500)

 
 

-parser.add_argument("--dim_keys", type=int, default=None)
+parser.add_argument("--nb_c_quizzes", type=int, default=5000)

 
 

-parser.add_argument("--dim_hidden", type=int, default=None)
+parser.add_argument("--c_quiz_multiplier", type=int, default=1)

 
 

-parser.add_argument("--nb_heads", type=int, default=None)
+parser.add_argument("--learning_rate", type=float, default=5e-4)

 
 

-parser.add_argument("--nb_blocks", type=int, default=None)
+parser.add_argument("--nb_have_to_be_correct", type=int, default=3)

 
 

-parser.add_argument("--dropout", type=float, default=0.1)
+parser.add_argument("--nb_have_to_be_wrong", type=int, default=1)

 
 

-parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
+parser.add_argument("--nb_mistakes_to_be_wrong", type=int, default=5)

 
 

-parser.add_argument("--no_checkpoint", action="store_true", default=False)
+# ----------------------------------

 
 

-parser.add_argument("--overwrite_results", action="store_true", default=False)
+parser.add_argument("--model_type", type=str, default="standard")

 
 

-parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
-
-##############################
-# picoclvr options
-
-parser.add_argument("--sandbox_level", type=int, default=0)
-
-parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
-
-parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
-
-parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
-
-##############################
-# picoclvr options
-
-parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
-
-parser.add_argument("--picoclvr_height", type=int, default=12)
-
-parser.add_argument("--picoclvr_width", type=int, default=16)
-
-parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
-
-##############################
-# Maze options
-
-parser.add_argument("--maze_height", type=int, default=23)
+parser.add_argument("--model", type=str, default="37M")

 
 

-parser.add_argument("--maze_width", type=int, default=39)
+parser.add_argument("--dim_model", type=int, default=None)

 
 

-parser.add_argument("--maze_nb_walls", type=int, default=45)
+parser.add_argument("--dim_keys", type=int, default=None)

 
 

-##############################
-# Snake options
+parser.add_argument("--dim_hidden", type=int, default=None)

 
 

-parser.add_argument("--snake_height", type=int, default=6)
+parser.add_argument("--nb_heads", type=int, default=None)

 
 

-parser.add_argument("--snake_width", type=int, default=8)
+parser.add_argument("--nb_blocks", type=int, default=None)

 
 

-parser.add_argument("--snake_nb_colors", type=int, default=5)
+parser.add_argument("--dropout", type=float, default=0.5)

 
 

-parser.add_argument("--snake_length", type=int, default=200)
+# ----------------------------------

 
 

-##############################
-# Stack options
+parser.add_argument("--nb_threads", type=int, default=1)

 
 

-parser.add_argument("--stack_nb_steps", type=int, default=100)
+parser.add_argument("--gpus", type=str, default="all")

 
 

-parser.add_argument("--stack_nb_stacks", type=int, default=3)
+# ----------------------------------

 
 

-parser.add_argument("--stack_nb_digits", type=int, default=3)
+parser.add_argument("--nb_models", type=int, default=5)

 
 

-parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
+parser.add_argument("--diffusion_nb_iterations", type=int, default=25)

 
 

-##############################
-# Expr options
+parser.add_argument("--diffusion_proba_corruption", type=float, default=0.05)

 
 

-parser.add_argument("--expr_nb_variables", type=int, default=5)
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.95)

 
 

-parser.add_argument("--expr_sequence_length", type=int, default=40)
+parser.add_argument("--proba_prompt_noise", type=float, default=0.05)

 
 

-parser.add_argument("--expr_operand_max", type=int, default=9)
+parser.add_argument("--proba_hint", type=float, default=0.25)

 
 

-parser.add_argument("--expr_result_max", type=int, default=99)
+parser.add_argument("--quizzes", type=str, default=None)

 
 

-parser.add_argument("--expr_input_file", type=str, default=None)
+######################################################################

 
 

-##############################
-# World options
+grids_tasks = ", ".join(
+    [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
+)

 
 

-parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
+parser.add_argument(
+    "--grids_world_tasks",
+    type=str,
+    default="replace_color,translate,grow,frame",
+    help="A comma-separated subset of: " + grids_tasks + ".",
+)

 
 ######################################################################
 
 args = parser.parse_args()
 
 
 ######################################################################
 
 args = parser.parse_args()
 

-assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
-

 if args.result_dir is None:
 if args.result_dir is None:

-    args.result_dir = f"results_{args.task}"
-
-######################################################################
-
-default_task_args = {
-    "sandbox": {
-        "nb_epochs": 50,
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 10000,
-    },
-    "picoclvr": {
-        "nb_epochs": 25,
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "mnist": {
-        "nb_epochs": 25,
-        "batch_size": 10,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "maze": {
-        "nb_epochs": 25,
-        "batch_size": 5,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "snake": {
-        "nb_epochs": 5,
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "stack": {
-        "nb_epochs": 5,
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 1000,
-    },
-    "expr": {
-        "nb_epochs": 40,
-        "batch_size": 25,
-        "nb_train_samples": 1000000,
-        "nb_test_samples": 10000,
-    },
-    "world": {
-        "nb_epochs": 10,
-        "batch_size": 25,
-        "nb_train_samples": 25000,
-        "nb_test_samples": 1000,
-    },
-}
-
-if args.task in default_task_args:
-    for k, v in default_task_args[args.task].items():
-        if getattr(args, k) is None:
-            setattr(args, k, v)
+    args.result_dir = f"results_culture"

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


@@ -229,6 +131,13 @@ default_model_args = {
         "nb_heads": 2,
         "nb_blocks": 2,
     },
         "nb_heads": 2,
         "nb_blocks": 2,
     },

+    "4M": {
+        "dim_model": 256,
+        "dim_keys": 32,
+        "dim_hidden": 1024,
+        "nb_heads": 4,
+        "nb_blocks": 6,
+    },

     "37M": {
         "dim_model": 512,
         "dim_keys": 64,
     "37M": {
         "dim_model": 512,
         "dim_keys": 64,


@@ -261,10 +170,14 @@ else:
 
 ######################################################################
 
 
 ######################################################################
 

-try:
-    os.mkdir(args.result_dir)
-except FileExistsError:
-    if not args.overwrite_results:
+if args.resume:
+    if not os.path.isdir(args.result_dir):
+        print(f"Trying to resume from a non-existing result dir {args.result_dir}.")
+        exit(1)
+else:
+    try:
+        os.mkdir(args.result_dir)
+    except FileExistsError:

         print(f"result directory {args.result_dir} already exists")
         exit(1)
 
         print(f"result directory {args.result_dir} already exists")
         exit(1)
 


@@ -282,6 +195,9 @@ if args.seed >= 0:
 
 
 def log_string(s):
 
 
 def log_string(s):

+    """print the given string prefixed with a time stamps, and log it
+    into log_file is not None"""
+

     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 
     if log_file is not None:
     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 
     if log_file is not None:


@@ -292,355 +208,793 @@ def log_string(s):
     sys.stdout.flush()
 
 
     sys.stdout.flush()
 
 

+######################################################################
+# Create a time-stamped archive of the source code
+
+with open("this_run.sh", "w") as f:
+    f.write(f"{' '.join(sys.argv)}\n")
+
+now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
+
+os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
+
+######################################################################
+
+log_string(f"argv {' '.join(sys.argv)}")
+

 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
 
 ######################################################################
 
 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(",")]

 
 

-def picoclvr_pruner_horizontal_green(p):
-    return not ("green" in p and ("left" in p or "right" in p))
+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")

 
 

-picoclvr_pruner_train = (
-    picoclvr_pruner_horizontal_green
-    if args.picocvlr_prune_properties in {"train+eval"}
-    else None
-)
+if args.train_batch_size is None:
+    args.train_batch_size = args.batch_size
+else:
+    assert args.batch_size % args.train_batch_size == 0

 
 

-picoclvr_pruner_eval = (
-    (lambda p: not picoclvr_pruner_horizontal_green(p))
-    if args.picocvlr_prune_properties in {"train+eval", "eval"}
-    else None
-)
+assert args.nb_train_samples % args.batch_size == 0
+assert args.nb_test_samples % args.batch_size == 0

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

-if args.task == "sandbox":
-    if args.sandbox_level == 0:
-        problem = tasks.ProblemLevel0(
-            nb_sentences=args.sandbox_levels_nb_items,
-            len_prompt=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
-        )
-    elif args.sandbox_level == 1:
-        problem = tasks.ProblemLevel1(
-            nb_operators=args.sandbox_levels_nb_items,
-            len_source=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
-        )
-    elif args.sandbox_level == 2:
-        problem = tasks.ProblemLevel2(
-            len_source=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
-        )
+
+def optimizer_to(optim, device):
+    """Move the optimizer optim to the device"""
+    for param in optim.state.values():
+        # Not sure there are any global tensors in the state dict
+        if isinstance(param, torch.Tensor):
+            param.data = param.data.to(device)
+            if param._grad is not None:
+                param._grad.data = param._grad.data.to(device)
+        elif isinstance(param, dict):
+            for subparam in param.values():
+                if isinstance(subparam, torch.Tensor):
+                    subparam.data = subparam.data.to(device)
+                    if subparam._grad is not None:
+                        subparam._grad.data = subparam._grad.data.to(device)
+
+
+######################################################################
+
+
+def generate_quiz_set(nb_samples, c_quizzes, c_quiz_multiplier=1):
+    if c_quizzes is None:
+        quizzes = problem.generate_w_quizzes(nb_samples)
+        nb_w_quizzes = quizzes.size(0)
+        nb_c_quizzes = 0

     else:
     else:

-        raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
-
-    task = tasks.SandBox(
-        problem,
-        # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        logger=log_string,
-        device=device,
-    )
+        if c_quiz_multiplier > 1:
+            n = min(c_quiz_multiplier, (nb_samples // 2) // c_quizzes.size(0))
+            body = c_quizzes.repeat(n, 1)
+            if n < c_quiz_multiplier:
+                tail = c_quizzes[
+                    torch.randperm(c_quizzes.size(0))[: nb_samples // 2 - body.size(0)]
+                ]
+                c_quizzes = torch.cat([body, tail], dim=0)
+            else:
+                c_quizzes = body
+
+        if c_quizzes.size(0) > nb_samples // 2:
+            i = torch.randperm(c_quizzes.size(0))[: nb_samples // 2]
+            c_quizzes = c_quizzes[i]
+
+        w_quizzes = problem.generate_w_quizzes(nb_samples - c_quizzes.size(0))
+
+        quizzes = torch.cat([w_quizzes, c_quizzes], dim=0)
+        nb_w_quizzes = w_quizzes.size(0)
+        nb_c_quizzes = c_quizzes.size(0)
+
+    i = torch.randperm(quizzes.size(0), device=quizzes.device)
+    quizzes = quizzes[i].contiguous()
+
+    log_string(f"quiz_set nb_w_quizzes {nb_w_quizzes} nb_c_quizzes {nb_c_quizzes}")
+
+    return quizzes
+
+
+######################################################################
+
+
+def add_hints_imt(imt_set):
+    """Set every component of the mask to zero with probability
+    args.proba_hint, and for each component set to zero, copy the
+    corresponding value from the target into the input
+
+    """
+    input, masks, targets = imt_set.unbind(dim=1)
+    # h = torch.rand(masks.size(), device=masks.device) - masks
+    # t = h.sort(dim=1).values[:, args.nb_hints, None]
+    # mask_hints = (h < t).long()
+    mask_hints = (
+        torch.rand(input.size(), device=input.device) < args.proba_hint
+    ).long() * masks
+    masks = (1 - mask_hints) * masks
+    input = (1 - mask_hints) * input + mask_hints * targets
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
+
+
+def add_noise_imt(imt_set):
+    """Replace every component of the input by a random value with
+    probability args.proba_prompt_noise."""
+    input, masks, targets = imt_set.unbind(dim=1)
+    noise = problem.pure_noise(input.size(0), input.device)
+    change = (1 - masks) * (
+        torch.rand(input.size(), device=input.device) < args.proba_prompt_noise
+    ).long()
+    input = (1 - change) * input + change * noise
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
+

 
 

-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,
-        height=args.picoclvr_height,
-        width=args.picoclvr_width,
-        nb_colors=args.picoclvr_nb_colors,
-        logger=log_string,
-        device=device,
-        pruner_train=picoclvr_pruner_train,
-        pruner_eval=picoclvr_pruner_eval,
+######################################################################
+# Prediction
+
+
+def samples_for_prediction_imt(input):
+    nb = input.size(0)
+    masks = input.new_zeros(input.size())
+    u = F.one_hot(torch.randint(4, (nb,), device=masks.device), num_classes=4)
+    masks.view(nb, 4, -1)[...] = u[:, :, None]
+    targets = input
+    input = (1 - masks) * targets
+
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
+
+
+def ae_predict(model, imt_set, local_device=main_device):
+    model.eval().to(local_device)
+
+    record = []
+
+    src = tqdm.tqdm(
+        imt_set.split(args.eval_batch_size),
+        dynamic_ncols=True,
+        desc="predict",
+        total=imt_set.size(0) // args.eval_batch_size,
+        delay=10,

     )
 
     )
 

-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,
-        device=device,
+    for imt in src:
+        # some paranoia
+        imt = imt.clone()
+        imt[:, 0] = imt[:, 0] * (1 - imt[:, 1])
+
+        with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+            logits = model(imt[:, 0] * 2 + imt[:, 1])
+        dist = torch.distributions.categorical.Categorical(logits=logits)
+        result = (1 - imt[:, 1]) * imt[:, 0] + imt[:, 1] * dist.sample()
+        record.append(result)
+
+    return torch.cat(record)
+
+
+def predict_the_four_grids(
+    model, input, with_noise=False, with_hints=False, local_device=main_device
+):
+    input = input[:, None, :].expand(-1, 4, -1).reshape(-1, input.size(1))
+    nb = input.size(0)
+    masks = input.new_zeros(input.size())
+    u = F.one_hot(torch.arange(nb, device=masks.device) % 4, num_classes=4)
+    masks.view(nb, 4, -1)[...] = u[:, :, None]
+    targets = input
+    input = (1 - masks) * targets
+    imt_set = torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
+
+    if with_hints:
+        imt_set = add_hints_imt(imt_set)
+
+    if with_noise:
+        imt_set = add_noise_imt(imt_set)
+
+    result = ae_predict(model, imt_set, local_device=local_device)
+    result = (result * masks).reshape(-1, 4, result.size(1)).sum(dim=1)
+
+    return result
+
+
+######################################################################
+
+
+def samples_for_generation_imt(input):
+    nb = input.size(0)
+    probs_iterations = 0.1 ** torch.linspace(
+        0, 1, args.diffusion_nb_iterations, device=input.device

     )
     )

+    probs_iterations = probs_iterations[None, :] / probs_iterations.sum()
+    probs_iterations = probs_iterations.expand(nb, -1)
+    dist = torch.distributions.categorical.Categorical(probs=probs_iterations)
+    t = dist.sample() + 1
+    r = torch.rand(input.size(), device=input.device)
+    proba_erased = 1 - (1 - args.diffusion_proba_corruption) ** t
+    mask_erased = (r <= proba_erased[:, None]).long()
+
+    noise = problem.pure_noise(nb, input.device)
+    targets = input
+    input = (1 - mask_erased) * input + mask_erased * noise
+    masks = input.new_full(input.size(), 1)
+
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
+
+
+def prioritized_rand(low):
+    x = torch.rand(low.size(), device=low.device).sort(dim=1, descending=True).values
+    k = torch.rand(low.size(), device=low.device) + low.long()
+    k = k.sort(dim=1).indices
+    y = x.new(x.size())
+    y.scatter_(dim=1, index=k, src=x)
+    return y
+
+
+def ae_generate(model, nb, local_device=main_device):
+    model.eval().to(local_device)
+
+    # We loop through the iterations first and through the
+    # mini-batches second so that we keep only the samples that have
+    # not stabilized
+
+    all_input = problem.pure_noise(nb, local_device)
+    all_masks = all_input.new_full(all_input.size(), 1)
+    all_changed = torch.full((all_input.size(0),), True, device=all_input.device)
+
+    for it in range(args.diffusion_nb_iterations):
+        # log_string(f"nb_changed {all_changed.long().sum().item()}")
+
+        if not all_changed.any():
+            break
+
+        sub_input = all_input[all_changed].clone()
+        sub_masks = all_masks[all_changed].clone()
+        sub_changed = all_changed[all_changed].clone()

 
 

-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,
-        height=args.maze_height,
-        width=args.maze_width,
-        nb_walls=args.maze_nb_walls,
-        device=device,
+        src = zip(
+            sub_input.split(args.eval_batch_size),
+            sub_masks.split(args.eval_batch_size),
+            sub_changed.split(args.eval_batch_size),
+        )
+
+        for input, masks, changed in src:
+            with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+                logits = model(input * 2 + masks)
+            dist = torch.distributions.categorical.Categorical(logits=logits)
+            output = dist.sample()
+            r = prioritized_rand(input != output)
+            mask_changes = (r <= args.diffusion_proba_corruption).long() * masks
+            update = (1 - mask_changes) * input + mask_changes * output
+            changed[...] = changed & (update != input).max(dim=1).values
+            input[...] = update
+
+        a = all_changed.clone()
+        all_input[a] = sub_input
+        all_masks[a] = sub_masks
+        all_changed[a] = sub_changed
+
+    return all_input
+
+
+######################################################################
+
+
+def one_epoch(model, n_epoch, c_quizzes, train=True, local_device=main_device):
+    quizzes = generate_quiz_set(
+        args.nb_train_samples if train else args.nb_test_samples,
+        c_quizzes,
+        args.c_quiz_multiplier,

     )
 
     )
 

-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,
-        height=args.snake_height,
-        width=args.snake_width,
-        nb_colors=args.snake_nb_colors,
-        length=args.snake_length,
-        prompt_length=args.snake_length // 2,
-        device=device,
+    q_p, q_g = quizzes.to(local_device).chunk(2)
+
+    # Half of the samples train the prediction, and we inject noise in
+    # all, and hints in half
+    b_p = samples_for_prediction_imt(q_p)
+    b_p = add_noise_imt(b_p)
+    half = torch.rand(b_p.size(0)) < 0.5
+    b_p[half] = add_hints_imt(b_p[half])
+
+    # The other half are denoising examples for the generation
+    b_g = samples_for_generation_imt(q_g)
+
+    imt_set = torch.cat([b_p, b_g])
+    imt_set = imt_set[torch.randperm(imt_set.size(0), device=imt_set.device)]
+
+    if train:
+        label = "train"
+        model.train().to(local_device)
+        optimizer_to(model.optimizer, local_device)
+        batch_size = args.train_batch_size
+    else:
+        label = "test"
+        model.eval().to(local_device)
+        batch_size = args.eval_batch_size
+
+    nb_samples, acc_loss = 0, 0.0
+
+    for imt in tqdm.tqdm(
+        imt_set.split(batch_size),
+        dynamic_ncols=True,
+        desc=label,
+        total=quizzes.size(0) // batch_size,
+        delay=10,
+    ):
+        input, masks, targets = imt.unbind(dim=1)
+        if train and nb_samples % args.batch_size == 0:
+            model.optimizer.zero_grad()
+
+        with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+            logits = model(input * 2 + masks)
+
+        loss_per_token = F.cross_entropy(
+            logits.transpose(1, 2), targets, reduction="none"
+        )
+        loss = (loss_per_token * masks).mean()
+        acc_loss += loss.item() * imt.size(0)
+        nb_samples += imt.size(0)
+
+        if train:
+            loss.backward()
+
+            if nb_samples % args.batch_size == 0:
+                model.optimizer.step()
+
+    log_string(f"{label}_loss {n_epoch} model {model.id} {acc_loss/nb_samples}")
+
+
+######################################################################
+
+
+def save_inference_images(model, n_epoch, c_quizzes, c_quiz_multiplier, local_device):
+    # Save some images of the prediction results
+
+    quizzes = generate_quiz_set(150, c_quizzes, args.c_quiz_multiplier)
+    imt_set = samples_for_prediction_imt(quizzes.to(local_device))
+    result = ae_predict(model, imt_set, local_device=local_device).to("cpu")
+    masks = imt_set[:, 1].to("cpu")
+
+    correct = (quizzes == result).min(dim=1).values.long()
+    correct_parts = (2 * correct - 1)[:, None] * masks.reshape(masks.size(0), 4, -1)[
+        :, :, 1
+    ]
+    predicted_parts = correct_parts.abs()
+
+    problem.save_quizzes_as_image(
+        args.result_dir,
+        f"culture_prediction_{n_epoch}_{model.id}.png",
+        quizzes=result[:128],
+        predicted_parts=predicted_parts[:128],
+        correct_parts=correct_parts[:128],

     )
 
     )
 

-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,
-        logger=log_string,
-        nb_steps=args.stack_nb_steps,
-        nb_stacks=args.stack_nb_stacks,
-        nb_digits=args.stack_nb_digits,
-        fraction_values_for_train=args.stack_fraction_values_for_train,
-        device=device,
+    # Save some images of the ex nihilo generation of the four grids
+
+    result = ae_generate(model, 150, local_device=local_device).to("cpu")
+    problem.save_quizzes_as_image(
+        args.result_dir,
+        f"culture_generation_{n_epoch}_{model.id}.png",
+        quizzes=result[:128],

     )
 
     )
 

-elif args.task == "expr":
-    task = tasks.Expr(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        nb_variables=args.expr_nb_variables,
-        sequence_length=args.expr_sequence_length,
-        operand_max=args.expr_operand_max,
-        result_max=args.expr_result_max,
-        batch_size=args.batch_size,
-        device=device,
+
+######################################################################
+
+
+def one_complete_epoch(
+    model, n_epoch, train_c_quizzes, test_c_quizzes, local_device=main_device
+):
+    one_epoch(model, n_epoch, train_c_quizzes, train=True, local_device=local_device)
+
+    one_epoch(model, n_epoch, test_c_quizzes, train=False, local_device=local_device)
+
+    # Compute the test accuracy
+
+    quizzes = generate_quiz_set(args.nb_test_samples, c_quizzes, args.c_quiz_multiplier)
+    imt_set = samples_for_prediction_imt(quizzes.to(local_device))
+    result = ae_predict(model, imt_set, local_device=local_device).to("cpu")
+    correct = (quizzes == result).min(dim=1).values.long()
+
+    nb_correct, nb_total = correct.sum().item(), quizzes.size(0)
+    model.test_accuracy = nb_correct / nb_total
+
+    log_string(
+        f"test_accuracy {n_epoch} model {model.id} nb_correct {nb_correct} / {nb_total} ({model.test_accuracy*100:.02f}%)"

     )
 
     )
 

-elif args.task == "world":
-    task = tasks.World(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        vqae_nb_epochs=args.world_vqae_nb_epochs,
-        logger=log_string,
-        device=device,
+    save_inference_images(
+        model, n_epoch, c_quizzes, args.c_quiz_multiplier, local_device=local_device

     )
 
     )
 

-else:
-    raise ValueError(f"Unknown task {args.task}")

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

-log_string(f"device {device}")

 
 

-vocabulary_size = task.vocabulary_size()
+def max_nb_mistakes_on_one_grid(quizzes, prediction):
+    return (
+        (prediction != quizzes)
+        .long()
+        .reshape(quizzes.size(0), 4, -1)
+        .sum(dim=2)
+        .max(dim=1)
+        .values
+    )

 
 

-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,
-)
+def evaluate_quizzes(quizzes, models, with_hints, local_device):
+    nb_correct, nb_wrong = 0, 0
+
+    for model in models:
+        model = copy.deepcopy(model).to(local_device).eval()
+        predicted = predict_the_four_grids(
+            model=model,
+            input=quizzes,
+            with_noise=False,
+            with_hints=with_hints,
+            local_device=local_device,
+        )
+        nb_mistakes = max_nb_mistakes_on_one_grid(quizzes, predicted)
+        nb_correct += (nb_mistakes == 0).long()
+        nb_wrong += (nb_mistakes >= args.nb_mistakes_to_be_wrong).long()

 
 

-model.to(device)
+    # print("\n\n", nb_correct, nb_wrong)
+
+    return nb_correct, nb_wrong

 
 

-nb_parameters = sum(p.numel() for p in model.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.")
+def identity_quizzes(quizzes):
+    quizzes = quizzes.reshape(quizzes.size(0), 4, -1)
+    return (quizzes[:, 0] == quizzes[:, 1]).min(dim=1).values | (
+        quizzes[:, 2] == quizzes[:, 3]
+    ).min(dim=1).values

 
 

-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.")
+def generate_c_quizzes(models, nb_to_generate, local_device=main_device):
+    record = []
+    nb_validated = 0

 
 

-    except FileNotFoundError:
-        log_string("starting from scratch.")
+    start_time = time.perf_counter()
+    last_log = -1
+
+    while nb_validated < nb_to_generate:
+        # Generate new quizzes
+
+        model = models[torch.randint(len(models), (1,)).item()]
+        model = copy.deepcopy(model).to(local_device).eval()
+        generator_id = model.id
+
+        c_quizzes = ae_generate(
+            model=model, nb=args.eval_batch_size * 10, local_device=local_device
+        )
+
+        c_quizzes = c_quizzes[identity_quizzes(c_quizzes) == False]
+
+        if c_quizzes.size(0) > 0:
+            # Select the ones that are solved properly by some models and
+            # not understood by others
+
+            nb_correct, nb_wrong = evaluate_quizzes(
+                quizzes=c_quizzes,
+                models=models,
+                with_hints=True,
+                local_device=local_device,
+            )
+
+            to_keep = (nb_correct >= args.nb_have_to_be_correct) & (
+                nb_wrong >= args.nb_have_to_be_wrong
+            )
+
+            nb_validated += to_keep.long().sum().item()
+            record.append(c_quizzes[to_keep])
+
+        #####################
+
+        duration = time.perf_counter() - start_time
+
+        if last_log < 0 or duration > last_log + 10:
+            last_log = duration
+            if nb_validated > 0:
+                if nb_validated < nb_to_generate:
+                    d = (nb_to_generate - nb_validated) * duration / nb_validated
+                    e = (
+                        datetime.datetime.now() + datetime.timedelta(seconds=d)
+                    ).strftime("%a %H:%M")
+                else:
+                    e = "now!"
+            else:
+                e = "???"
+
+            log_string(
+                f"nb_validated {nb_validated} model {generator_id} (finishes {e} -- {int((nb_validated * 3600)/duration)}/h)"
+            )
+
+        #####################
+
+    duration = time.perf_counter() - start_time
+
+    log_string(f"generate_c_quizz_speed {int(3600 * nb_validated / duration)}/h")
+
+    return torch.cat(record).to("cpu")

 
 

-    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(
-        nb_epochs_finished,
-        model,
-        args.result_dir,
-        log_string,
-        args.deterministic_synthesis,
-        args.expr_input_file,
-    )

 
 

-    exit(0)
+def multithread_execution(fun, arguments):
+    # Single instance, no thread
+    if len(arguments) == 1:
+        return fun(*(arguments[0]))
+
+    records, threads = [], []
+
+    def threadable_fun(*args):
+        r = fun(*args)
+        if type(r) is not tuple:
+            r = (r,)
+        records.append(r)
+
+    for args in arguments:
+        # To get a different sequence between threads
+        log_string(f"dummy_rand {torch.rand(1)}")
+        # torch.rand(1)
+        t = threading.Thread(target=threadable_fun, daemon=True, args=args)
+        threads.append(t)
+        t.start()
+
+    for t in threads:
+        t.join()
+
+    if records[0] == (None,):
+        return
+    else:
+        return [
+            torch.cat([x[k] for x in records], dim=0) for k in range(len(records[0]))
+        ]
+

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

-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default

 
 

-# Compute the entropy of the training tokens
+def save_models(models, suffix=""):
+    if suffix != "":
+        suffix = "_" + suffix
+
+    for model in models:
+        filename = f"ae_{model.id:03d}{suffix}.pth"
+        torch.save(
+            {
+                "state_dict": model.state_dict(),
+                "optimizer_state_dict": model.optimizer.state_dict(),
+                "test_accuracy": model.test_accuracy,
+            },
+            os.path.join(args.result_dir, filename),
+        )

 
 

-token_count = 0
-for input in task.batches(split="train"):
-    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()
-train_set_perplexity = math.exp(entropy)
+    log_string(f"wrote ae_*{suffix}.pth")

 
 

-##############################

 
 

-# A bit of paranoia never hurts
+######################################################################

 
 

-train_examples = {}

 
 

+def save_quiz_image(models, c_quizzes, filename, local_device=main_device):
+    c_quizzes = c_quizzes.to(local_device)

 
 

-for input in task.batches(split="train"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        train_examples[x.sum().item()] = x
+    nb_correct, nb_wrong = evaluate_quizzes(
+        quizzes=c_quizzes,
+        models=models,
+        with_hints=False,
+        local_device=local_device,
+    )

 
 

-nb_total, nb_collisions = 0, 0
-for input in task.batches(split="test"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        nb_total += 1
-        y = train_examples.get(x.sum().item())
-        if y is not None:
-            if x.size() == y.size() and (x - y).abs().sum() == 0:
-                nb_collisions += 1
+    comments = [f"nb_correct {c} nb_wrong {w}" for c, w in zip(nb_correct, nb_wrong)]

 
 

-del train_examples
+    problem.save_quizzes_as_image(
+        args.result_dir,
+        filename,
+        quizzes=c_quizzes,
+        comments=comments,
+        delta=True,
+        nrow=8,
+    )

 
 

-log_string(
-    f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
+    log_string(f"wrote {filename}")
+
+
+######################################################################
+
+problem = grids.Grids(
+    max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
+    chunk_size=100,
+    nb_threads=args.nb_threads,
+    tasks=args.grids_world_tasks,

 )
 
 )
 

-##############################
+if not args.resume:
+    problem.save_some_examples(args.result_dir)
+
+
+log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
+
+vocabulary_size = problem.vocabulary_size()
+
+log_string(f"vocabulary_size {vocabulary_size}")
+
+######################################################################
+
+models = []

 
 

-if args.learning_rate_schedule == "cos":
-    learning_rate_schedule = {}
-    for n_epoch in range(args.nb_epochs):
-        u = n_epoch / args.nb_epochs * math.pi
-        learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
+if args.model_type == "standard":
+    model_constructor = attae.AttentionAE
+elif args.model_type == "functional":
+    model_constructor = attae.FunctionalAttentionAE

 else:
 else:

-    u = {
-        int(k): float(v)
-        for k, v in [
-            tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
-        ]
-    }
+    raise ValueError(f"Unknown model type {args.model_type}")
+
+
+for i in range(args.nb_models):
+    model = model_constructor(
+        vocabulary_size=vocabulary_size * 2,
+        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,
+        dropout=args.dropout,
+    )

 
 

-    learning_rate_schedule = {}
-    learning_rate = args.learning_rate
-    for n_epoch in range(args.nb_epochs):
-        if n_epoch in u:
-            learning_rate = u[n_epoch]
-        learning_rate_schedule[n_epoch] = learning_rate
+    # model = torch.compile(model)

 
 

-log_string(f"learning_rate_schedule {learning_rate_schedule}")
+    model.id = i
+    model.test_accuracy = 0.0
+    model.optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)

 
 

-##############################
+    models.append(model)

 
 

-nb_samples_seen = 0
+######################################################################

 
 

-if nb_epochs_finished >= nb_epochs:
-    task.produce_results(
-        nb_epochs_finished,
-        model,
-        args.result_dir,
-        log_string,
-        args.deterministic_synthesis,
+current_epoch = 0
+
+if args.resume:
+    for model in models:
+        filename = f"ae_{model.id:03d}.pth"
+
+        d = torch.load(
+            os.path.join(args.result_dir, filename),
+            map_location="cpu",
+            weights_only=False,
+        )
+        model.load_state_dict(d["state_dict"])
+        model.optimizer.load_state_dict(d["optimizer_state_dict"])
+        model.test_accuracy = d["test_accuracy"]
+        log_string(f"successfully loaded {filename}")
+
+    filename = "state.pth"
+    state = torch.load(
+        os.path.join(args.result_dir, filename),
+        map_location="cpu",
+        weights_only=False,

     )
 
     )
 

-for n_epoch in range(nb_epochs_finished, nb_epochs):
-    learning_rate = learning_rate_schedule[n_epoch]
+    log_string(f"successfully loaded {filename}")

 
 

-    log_string(f"learning_rate {learning_rate}")
+    current_epoch = state["current_epoch"]
+    train_c_quizzes = state["train_c_quizzes"]
+    test_c_quizzes = state["test_c_quizzes"]

 
 

-    if args.optim == "sgd":
-        optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
-    elif args.optim == "adam":
-        optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
-    elif args.optim == "adamw":
-        optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
-    else:
-        raise ValueError(f"Unknown optimizer {args.optim}.")
+######################################################################
+
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")

 
 

-    model.train()

 
 

-    nb_train_samples, acc_train_loss = 0, 0.0
+######################################################################
+
+train_c_quizzes, test_c_quizzes = None, None
+
+######################################################################

 
 

-    for input in task.batches(split="train"):
-        input = input.to(device)
-        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)
+for n_epoch in range(current_epoch, args.nb_epochs):
+    start_time = time.perf_counter()

 
 

-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
+    state = {
+        "current_epoch": n_epoch,
+        "train_c_quizzes": train_c_quizzes,
+        "test_c_quizzes": test_c_quizzes,
+    }
+
+    filename = "state.pth"
+    torch.save(state, os.path.join(args.result_dir, filename))
+    log_string(f"wrote {filename}")

 
 

-    with torch.autograd.no_grad():
-        model.eval()
+    log_string(f"--- epoch {n_epoch} ----------------------------------------")

 
 

-        nb_test_samples, acc_test_loss = 0, 0.0
+    cta = " ".join([f"{float(m.test_accuracy):.04f}" for m in models])
+    log_string(f"current_test_accuracies {cta}")

 
 

-        for input in task.batches(split="test"):
-            input = input.to(device)
+    # --------------------------------------------------------------------

 
 

-            output = model(mygpt.BracketedSequence(input)).x
-            loss = F.cross_entropy(output.transpose(1, 2), input)
-            acc_test_loss += loss.item() * input.size(0)
-            nb_test_samples += input.size(0)
+    lowest_test_accuracy = min([float(m.test_accuracy) for m in models])

 
 

-        train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+    if lowest_test_accuracy >= args.accuracy_to_make_c_quizzes:
+        if train_c_quizzes is None:
+            save_models(models, "naive")

 
 

-        log_string(
-            f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
+        nb_gpus = len(gpus)
+        nb_c_quizzes_to_generate = (args.nb_c_quizzes + nb_gpus - 1) // nb_gpus
+
+        (new_c_quizzes,) = multithread_execution(
+            generate_c_quizzes,
+            [(models, nb_c_quizzes_to_generate, gpu) for gpu in gpus],

         )
 
         )
 

-        task.produce_results(
-            n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
+        save_quiz_image(
+            models, new_c_quizzes[:256], f"culture_c_quiz_{n_epoch:04d}.png"

         )
 
         )
 

-    checkpoint = {
-        "nb_epochs_finished": n_epoch + 1,
-        "model_state": model.state_dict(),
-        "rng_state": torch.get_rng_state(),
-    }
+        log_string(f"generated_c_quizzes {new_c_quizzes.size()}")

 
 

-    if torch.cuda.is_available():
-        checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+        train_c_quizzes = (
+            new_c_quizzes
+            if train_c_quizzes is None
+            else torch.cat([train_c_quizzes, new_c_quizzes])
+        )
+        train_c_quizzes = train_c_quizzes[-args.nb_train_samples :]

 
 

-    checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
-    torch.save(checkpoint, checkpoint_name)
-    log_string(f"saved checkpoint {checkpoint_name}")
+        nb_correct, _ = evaluate_quizzes(
+            quizzes=train_c_quizzes,
+            models=models,
+            with_hints=False,
+            local_device=local_device,
+        )

 
 

-######################################################################
+        test_c_quizzes = train_c_quizzes[nb_correct >= args.nb_have_to_be_correct]
+
+        for model in models:
+            model.test_accuracy = 0
+
+    if train_c_quizzes is None:
+        log_string("no_c_quiz")
+    else:
+        log_string(f"nb_c_quizzes {train_c_quizzes.size(0)}")
+
+    # --------------------------------------------------------------------
+
+    ranked_models = sorted(models, key=lambda m: float(m.test_accuracy))
+    weakest_models = ranked_models[: len(gpus)]
+
+    log_string(
+        f"weakest_accuracies {[model.test_accuracy for model in weakest_models]}"
+    )
+
+    multithread_execution(
+        one_complete_epoch,
+        [
+            (model, n_epoch, train_c_quizzes, test_c_quizzes, gpu)
+            for model, gpu in zip(weakest_models, gpus)
+        ],
+    )
+
+    save_models(models)
+
+    # --------------------------------------------------------------------
+
+    duration = time.perf_counter() - start_time
+    str_duration = ""
+    if duration >= 60:
+        str_duration += f"{int(duration)//60}min"
+    str_duration += f"{int(duration)%60}s"
+    str_next = (
+        datetime.datetime.now() + datetime.timedelta(seconds=duration)
+    ).strftime("%H:%M:%S")
+    log_string(f"epoch_duration {str_duration} next_finish {str_next}")