Initial commit
authorFrançois Fleuret <francois@fleuret.org>
Sat, 11 Mar 2023 12:58:56 +0000 (13:58 +0100)
committerFrançois Fleuret <francois@fleuret.org>
Sat, 11 Mar 2023 12:58:56 +0000 (13:58 +0100)
beaver.py [new file with mode: 0755]
maze.py [new file with mode: 0755]
mygpt.py [new file with mode: 0755]

diff --git a/beaver.py b/beaver.py
new file mode 100755 (executable)
index 0000000..b0fa03c
--- /dev/null
+++ b/beaver.py
@@ -0,0 +1,631 @@
+#!/usr/bin/env python
+
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
+# torch.backends.cuda.matmul.allow_tf23
+# torch.autocast(torch.bfloat16)
+
+import math, sys, argparse, time, tqdm, itertools, os
+
+import torch, torchvision
+from torch import nn
+from torch.nn import functional as F
+
+import mygpt, tensorstack
+
+######################################################################
+
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+    torch.backends.cuda.matmul.allow_tf32 = True
+else:
+    device = torch.device("cpu")
+
+######################################################################
+
+parser = argparse.ArgumentParser(
+    description="An implementation of GPT with cache to solve a toy geometric reasoning task."
+)
+
+parser.add_argument("--log_filename", type=str, default="train.log")
+
+parser.add_argument("--result_dir", type=str, default="results_default")
+
+parser.add_argument("--seed", type=int, default=0)
+
+parser.add_argument("--nb_epochs", type=int, default=25)
+
+parser.add_argument("--batch_size", type=int, default=100)
+
+parser.add_argument("--data_size", type=int, default=-1)
+
+parser.add_argument("--optim", type=str, default="adam")
+
+parser.add_argument("--learning_rate", type=float, default=1e-3)
+
+parser.add_argument(
+    "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
+)
+
+parser.add_argument("--dim_model", type=int, default=512)
+
+parser.add_argument("--dim_keys", type=int, default=64)
+
+parser.add_argument("--dim_hidden", type=int, default=2048)
+
+parser.add_argument("--nb_heads", type=int, default=8)
+
+parser.add_argument("--nb_blocks", type=int, default=12)
+
+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("--overwrite_results", action="store_true", default=False)
+
+parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
+
+##############################
+# picoclvr options
+
+parser.add_argument("--world_height", type=int, default=23)
+
+parser.add_argument("--world_width", type=int, default=31)
+
+parser.add_argument("--world_nb_walls", type=int, default=15)
+
+######################################################################
+
+args = parser.parse_args()
+
+assert args.prune_properties in {"none", "train+eval", "eval"}
+
+try:
+    os.mkdir(args.result_dir)
+except FileExistsError:
+    if not args.overwrite_results:
+        print(f"result directory {args.result_dir} already exists")
+        exit(1)
+
+log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
+
+if args.seed >= 0:
+    # torch.backends.cudnn.deterministic = True
+    # torch.backends.cudnn.benchmark = False
+    # torch.use_deterministic_algorithms(True)
+    torch.manual_seed(args.seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(args.seed)
+
+######################################################################
+
+
+def log_string(s):
+    t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
+
+    if log_file is not None:
+        log_file.write(t + s + "\n")
+        log_file.flush()
+
+    print(t + s)
+    sys.stdout.flush()
+
+
+for n in vars(args):
+    log_string(f"args.{n} {getattr(args, n)}")
+
+######################################################################
+
+
+def masked_inplace_autoregression(
+    model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
+):
+
+    for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+        i = (ar_mask.sum(0) > 0).nonzero()
+        if i.min() > 0:
+            model(
+                mygpt.BracketedSequence(input, 0, i.min())
+            )  # Needed to initialize the model's cache
+        for s in range(i.min(), i.max() + 1):
+            output = model(mygpt.BracketedSequence(input, s, 1)).x
+            logits = output[:, s]
+            if forbidden_tokens is not None:
+                logits = logits.masked_fill(forbidden_tokens, float("-inf"))
+            if args.deterministic_synthesis:
+                t_next = logits.argmax(1)
+            else:
+                dist = torch.distributions.categorical.Categorical(logits=logits)
+                t_next = dist.sample()
+            input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
+
+
+######################################################################
+
+
+class Task:
+    def batches(self, split="train"):
+        pass
+
+    def vocabulary_size(self):
+        pass
+
+    def produce_results(self, n_epoch, model):
+        pass
+
+
+######################################################################
+
+import picoclvr
+
+
+class TaskPicoCLVR(Task):
+
+    # Make a tensor from a list of strings
+    def tensorize(self, descr):
+        token_descr = [s.strip().split(" ") for s in descr]
+        l = max([len(s) for s in token_descr])
+        token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
+        id_descr = [[self.token2id[u] for u in s] for s in token_descr]
+        return torch.tensor(id_descr, device=self.device)
+
+    # Make a list of strings from a tensor
+    def detensorize(self, x):
+        return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
+
+    # trim all the tensors in the tuple z to remove as much token from
+    # left and right in the first tensor. If z is a tuple, all its
+    # elements are trimed according to the triming for the first
+    def trim(self, z, token="<nul>"):
+        n = self.token2id[token]
+        if type(z) == tuple:
+            x = z[0]
+            i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return tuple([t[:, a:b] for t in z])
+        else:
+            i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
+            a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+            return z[:, a:b]
+
+    ######################
+    # Not the cleanest part of the code
+
+    # Extract the last image of each sequence, from the last <img>
+    # included, and set to <nul> all the tokens from the beginning of
+    # that image to the end
+    def excise_last_image(self, input):
+        t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
+        nb_img_tokens = self.height * self.width + 1
+
+        input = input.clone()
+        t = (input == t_img).long()
+        tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
+        i = (t * tail_masks).nonzero(as_tuple=True)
+        j = (
+            i[0][:, None],
+            i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
+        )
+        images = self.trim(input[j])
+        input[j] = t_nul
+        loss_masks = 1 - tail_masks
+        input, loss_masks = self.trim((input, loss_masks))
+        return input, loss_masks, images
+
+    def add_true_image(self, input, images, loss_masks):
+        t_nul = self.token2id["<nul>"]
+        nb_img_tokens = self.height * self.width + 1
+        input = F.pad(input, (0, nb_img_tokens), value=t_nul)
+        loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
+        t = (input == t_nul).long()
+        i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
+        j = (
+            i[0][:, None],
+            i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
+        )
+        input[j] = images
+        loss_masks[j] = 1
+        input, loss_masks = self.trim((input, loss_masks))
+        return input, loss_masks
+
+    def add_generated_image(self, input, loss_masks, model):
+        t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
+        nb_img_tokens = self.height * self.width + 1
+
+        input = F.pad(input, (0, nb_img_tokens), value=t_nul)
+        loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
+        t = (input == t_nul).long()
+        i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
+        input[i] = t_img
+
+        j = (
+            i[0][:, None],
+            i[1][:, None]
+            + 1
+            + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
+        )
+        ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
+        ar_masks[j] = 1
+        forbidden_tokens = (
+            torch.arange(self.vocabulary_size(), device=input.device) == t_nul
+        )
+        with torch.autograd.no_grad():
+            t = model.training
+            model.eval()
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                input,
+                ar_masks,
+                forbidden_tokens,
+                device=self.device,
+            )
+            model.train(t)
+
+        input, loss_masks = self.trim((input, loss_masks))
+
+        return input, loss_masks
+
+    ######################
+
+    def __init__(
+        self,
+        batch_size,
+        height,
+        width,
+        nb_colors=5,
+        device=torch.device("cpu"),
+        pruner_train=None,
+        pruner_eval=None,
+    ):
+        def generate_descr(nb, cache_suffix, pruner):
+            return picoclvr.generate(
+                nb,
+                height=self.height,
+                width=self.width,
+                nb_colors=nb_colors,
+                pruner=pruner,
+            )
+
+        self.height = height
+        self.width = width
+        self.batch_size = batch_size
+        self.device = device
+        nb = args.data_size if args.data_size > 0 else 250000
+        self.pruner_train = pruner_train
+        self.pruner_eval = pruner_eval
+
+        param = {
+            "nb": nb,
+            "height": height,
+            "width": width,
+            "nb_colors": nb_colors,
+            "batch_size": batch_size,
+            "rng_state": list(torch.get_rng_state()),
+        }
+
+        log_string(f"generating {nb} samples (can take some time)")
+        self.train_descr = generate_descr(
+            (nb * 4) // 5, "train", pruner=self.pruner_train
+        )
+        self.test_descr = generate_descr((nb * 1) // 5, "test", pruner=None)
+
+        # Build the tokenizer
+        tokens = {"<nul>", "<img>"}
+        for d in [self.train_descr, self.test_descr]:
+            for s in d:
+                for t in s.strip().split(" "):
+                    tokens.add(t)
+        # make this set a sorted list to get the same tensors given
+        # the same descr
+        tokens = list(tokens)
+        tokens.sort()
+        self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
+        self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
+
+        # Tokenize the train and test sets
+        self.train_input = self.tensorize(self.train_descr)
+        self.test_input = self.tensorize(self.test_descr)
+
+    def batches(self, split="train"):
+        assert split in {"train", "test"}
+        input = self.train_input if split == "train" else self.test_input
+        for batch in tqdm.tqdm(
+            input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
+        ):
+            yield self.trim(batch)
+
+    def vocabulary_size(self):
+        return len(self.token2id)
+
+    def compute_missing_properties(self, n_epoch, model, pruner=None):
+
+        acc_nb_requested_properties = []
+        acc_nb_missing_properties = []
+        acc_nb_results = 0
+
+        for input in tqdm.tqdm(
+            self.test_input.split(self.batch_size),
+            dynamic_ncols=True,
+            desc=f"test-properties",
+        ):
+            tape, loss_masks, _ = self.excise_last_image(input)
+            tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
+            result_descr = self.detensorize(tape)
+            np = picoclvr.nb_properties(
+                result_descr,
+                height=self.height,
+                width=self.width,
+                pruner=pruner,
+            )
+            nb_requested_properties, _, nb_missing_properties = zip(*np)
+            acc_nb_requested_properties += nb_requested_properties
+            acc_nb_missing_properties += nb_missing_properties
+            acc_nb_results += len(result_descr)
+
+        nb_requested_properties = sum(acc_nb_requested_properties)
+        nb_missing_properties = sum(acc_nb_missing_properties)
+
+        prefix = "" if pruner is None else "pruned_"
+        log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
+        log_string(
+            f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
+        )
+        log_string(
+            f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
+        )
+
+    ######################################################################
+
+    def produce_results(self, n_epoch, model):
+
+        self.compute_missing_properties(n_epoch, model)
+
+        if self.pruner_eval is not None:
+            self.compute_missing_properties(n_epoch, model, self.pruner_eval)
+
+        nb_tokens_to_generate = self.height * self.width + 3
+        result_descr = []
+        nb_per_primer = 8
+        primer = []
+
+        for primer_descr in [
+            "red above green <sep> green top <sep> blue right of red",
+            "there is red <sep> there is yellow <sep> there is blue",
+            "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
+            "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
+        ]:
+            primer += [primer_descr] * nb_per_primer
+
+        tape = self.tensorize(primer)
+        loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
+        tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
+        result_descr = self.detensorize(tape)
+
+        np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
+
+        acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
+        acc_nb_results = len(result_descr)
+
+        nb_requested_properties = sum(acc_nb_requested_properties)
+        nb_missing_properties = sum(acc_nb_missing_properties)
+
+        prefix = "demo_"
+        log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
+        log_string(
+            f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
+        )
+        log_string(
+            f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
+        )
+
+        img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
+
+        if img.dim() == 5:
+            if img.size(1) == 1:
+                img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
+            else:
+                img = torch.cat(
+                    [
+                        torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
+                        for x in img
+                    ],
+                    0,
+                )
+
+        image_name = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
+        torchvision.utils.save_image(
+            img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
+        )
+        log_string(f"wrote {image_name}")
+
+
+######################################################################
+
+log_string(f"device {device}")
+
+
+def pruner_horizontal_green(p):
+    return not ("green" in p and ("left" in p or "right" in p))
+
+
+task = TaskPicoCLVR(
+    batch_size=args.batch_size,
+    height=args.height,
+    width=args.width,
+    nb_colors=args.nb_colors,
+    device=device,
+    pruner_train=pruner_horizontal_green
+    if args.prune_properties in {"train+eval"}
+    else None,
+    pruner_eval=(lambda p: not pruner_horizontal_green(p))
+    if args.prune_properties in {"train+eval", "eval"}
+    else None,
+)
+
+vocabulary_size = task.vocabulary_size()
+
+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,
+)
+
+model.to(device)
+
+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.")
+
+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)
+
+######################################################################
+
+nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
+
+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)
+
+##############################
+
+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))
+else:
+    u = {
+        int(k): float(v)
+        for k, v in [
+            tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
+        ]
+    }
+
+    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
+
+log_string(f"learning_rate_schedule {learning_rate_schedule}")
+
+##############################
+
+nb_samples_seen = 0
+
+if nb_epochs_finished >= nb_epochs:
+    task.produce_results(nb_epochs_finished, model)
+
+for n_epoch in range(nb_epochs_finished, nb_epochs):
+
+    learning_rate = learning_rate_schedule[n_epoch]
+
+    log_string(f"learning_rate {learning_rate}")
+
+    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}.")
+
+    model.train()
+
+    nb_train_samples, acc_train_loss = 0, 0.0
+
+    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)
+
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+    with torch.autograd.no_grad():
+
+        model.eval()
+
+        nb_test_samples, acc_test_loss = 0, 0.0
+
+        for input in task.batches(split="test"):
+            input = input.to(device)
+
+            # input, loss_masks, true_images = task.excise_last_image(input)
+            # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
+
+            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)
+
+        train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+
+        log_string(
+            f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
+        )
+
+        task.produce_results(n_epoch, model)
+
+    checkpoint = {
+        "nb_epochs_finished": n_epoch + 1,
+        "model_state": model.state_dict(),
+        "rng_state": torch.get_rng_state(),
+    }
+
+    if torch.cuda.is_available():
+        checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+
+    checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
+    torch.save(checkpoint, checkpoint_name)
+    log_string(f"saved checkpoint {checkpoint_name}")
+
+######################################################################
diff --git a/maze.py b/maze.py
new file mode 100755 (executable)
index 0000000..2c44319
--- /dev/null
+++ b/maze.py
@@ -0,0 +1,218 @@
+#!/usr/bin/env python
+
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
+import torch, torchvision
+
+######################################################################
+
+v_empty, v_wall, v_start, v_goal, v_path = 0, 1, 2, 3, 4
+
+
+def create_maze(h=11, w=17, nb_walls=8):
+    a, k = 0, 0
+
+    while k < nb_walls:
+        while True:
+            if a == 0:
+                m = torch.zeros(h, w, dtype=torch.int64)
+                m[0, :] = 1
+                m[-1, :] = 1
+                m[:, 0] = 1
+                m[:, -1] = 1
+
+            r = torch.rand(4)
+
+            if r[0] <= 0.5:
+                i1, i2, j = (
+                    int((r[1] * h).item()),
+                    int((r[2] * h).item()),
+                    int((r[3] * w).item()),
+                )
+                i1, i2, j = i1 - i1 % 2, i2 - i2 % 2, j - j % 2
+                i1, i2 = min(i1, i2), max(i1, i2)
+                if i2 - i1 > 1 and i2 - i1 <= h / 2 and m[i1 : i2 + 1, j].sum() <= 1:
+                    m[i1 : i2 + 1, j] = 1
+                    break
+            else:
+                i, j1, j2 = (
+                    int((r[1] * h).item()),
+                    int((r[2] * w).item()),
+                    int((r[3] * w).item()),
+                )
+                i, j1, j2 = i - i % 2, j1 - j1 % 2, j2 - j2 % 2
+                j1, j2 = min(j1, j2), max(j1, j2)
+                if j2 - j1 > 1 and j2 - j1 <= w / 2 and m[i, j1 : j2 + 1].sum() <= 1:
+                    m[i, j1 : j2 + 1] = 1
+                    break
+            a += 1
+
+            if a > 10 * nb_walls:
+                a, k = 0, 0
+
+        k += 1
+
+    return m
+
+
+######################################################################
+
+
+def compute_distance(walls, i, j):
+    max_length = walls.numel()
+    dist = torch.full_like(walls, max_length)
+
+    dist[i, j] = 0
+    pred_dist = torch.empty_like(dist)
+
+    while True:
+        pred_dist.copy_(dist)
+        d = (
+            torch.cat(
+                (
+                    dist[None, 1:-1, 0:-2],
+                    dist[None, 2:, 1:-1],
+                    dist[None, 1:-1, 2:],
+                    dist[None, 0:-2, 1:-1],
+                ),
+                0,
+            ).min(dim=0)[0]
+            + 1
+        )
+
+        dist[1:-1, 1:-1] = torch.min(dist[1:-1, 1:-1], d)
+        dist = walls * max_length + (1 - walls) * dist
+
+        if dist.equal(pred_dist):
+            return dist * (1 - walls)
+
+
+######################################################################
+
+
+def compute_policy(walls, i, j):
+    distance = compute_distance(walls, i, j)
+    distance = distance + walls.numel() * walls
+
+    value = distance.new_full((4,) + distance.size(), walls.numel())
+    value[0, :, 1:] = distance[:, :-1]
+    value[1, :, :-1] = distance[:, 1:]
+    value[2, 1:, :] = distance[:-1, :]
+    value[3, :-1, :] = distance[1:, :]
+
+    proba = (value.min(dim=0)[0][None] == value).float()
+    proba = proba / proba.sum(dim=0)[None]
+    proba = proba * (1 - walls) + walls.float() / 4
+
+    return proba
+
+
+######################################################################
+
+
+def mark_path(walls, i, j, goal_i, goal_j):
+    policy = compute_policy(walls, goal_i, goal_j)
+    action = torch.distributions.categorical.Categorical(
+        policy.permute(1, 2, 0)
+    ).sample()
+    walls[i, j] = 4
+    n, nmax = 0, walls.numel()
+    while i != goal_i or j != goal_j:
+        di, dj = [(0, -1), (0, 1), (-1, 0), (1, 0)][action[i, j]]
+        i, j = i + di, j + dj
+        assert walls[i, j] == 0
+        walls[i, j] = 4
+        n += 1
+        assert n < nmax
+
+
+def valid_paths(mazes, paths):
+    still_ok = (mazes - (paths * (paths < 4))).view(mazes.size(0), -1).abs().sum(1) == 0
+    reached = still_ok.new_zeros(still_ok.size())
+    current, pred_current = paths.clone(), paths.new_zeros(paths.size())
+    goal = (mazes == v_goal).long()
+    while not pred_current.equal(current):
+        # print(current)
+        # print(f'{still_ok=} {reached=}')
+        pred_current.copy_(current)
+        u = (current == v_start).long()
+        possible_next = (
+            u[:, 2:, 1:-1] + u[:, 0:-2, 1:-1] + u[:, 1:-1, 2:] + u[:, 1:-1, 0:-2] > 0
+        ).long()
+        u = u[:, 1:-1, 1:-1]
+        reached += ((goal[:, 1:-1, 1:-1] * possible_next).sum((1, 2)) == 1) * (
+            (current == v_path).sum((1, 2)) == 0
+        )
+        current[:, 1:-1, 1:-1] = (1 - u) * current[:, 1:-1, 1:-1] + (
+            v_start - v_path
+        ) * (possible_next * (current[:, 1:-1, 1:-1] == v_path))
+        still_ok *= (current == v_start).sum((1, 2)) <= 1
+
+    return still_ok * reached
+
+
+######################################################################
+
+
+def create_maze_data(nb, h=11, w=17, nb_walls=8, dist_min=-1):
+    mazes = torch.empty(nb, h, w, dtype=torch.int64)
+    paths = torch.empty(nb, h, w, dtype=torch.int64)
+
+    for n in range(nb):
+        maze = create_maze(h, w, nb_walls)
+        i = (1 - maze).nonzero()
+        while True:
+            start, goal = i[torch.randperm(i.size(0))[:2]]
+            if (start - goal).abs().sum() >= dist_min:
+                break
+
+        path = maze.clone()
+        mark_path(path, start[0], start[1], goal[0], goal[1])
+        maze[start[0], start[1]] = v_start
+        maze[goal[0], goal[1]] = v_goal
+        path[start[0], start[1]] = v_start
+        path[goal[0], goal[1]] = v_goal
+
+        mazes[n] = maze
+        paths[n] = path
+
+    return mazes, paths
+
+
+######################################################################
+
+
+def save_image(name, mazes, paths):
+    mazes, paths = mazes.cpu(), paths.cpu()
+
+    colors = torch.tensor(
+        [
+            [255, 255, 255],  # empty
+            [0, 0, 0],  # wall
+            [0, 255, 0],  # start
+            [0, 0, 255],  # goal
+            [255, 0, 0],  # path
+        ]
+    )
+
+    mazes = colors[mazes.reshape(-1)].reshape(mazes.size() + (-1,)).permute(0, 3, 1, 2)
+    paths = colors[paths.reshape(-1)].reshape(paths.size() + (-1,)).permute(0, 3, 1, 2)
+
+    img = torch.cat((mazes.unsqueeze(1), paths.unsqueeze(1)), 1)
+    img = img.reshape((-1,) + img.size()[2:]).float() / 255.0
+
+    torchvision.utils.save_image(img, name, padding=1, pad_value=0.5, nrow=8)
+
+
+######################################################################
+
+if __name__ == "__main__":
+
+    mazes, paths = create_maze_data(32, dist_min=10)
+    save_image("test.png", mazes, paths)
+    print(valid_paths(mazes, paths))
+
+######################################################################
diff --git a/mygpt.py b/mygpt.py
new file mode 100755 (executable)
index 0000000..5ea4668
--- /dev/null
+++ b/mygpt.py
@@ -0,0 +1,290 @@
+#!/usr/bin/env python
+
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
+import math
+
+import torch
+
+from torch import nn
+from torch.nn import functional as F
+
+######################################################################
+
+
+class WithResidual(nn.Module):
+    def __init__(self, *f):
+        super().__init__()
+        self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
+
+    def forward(self, bs):
+        bs.x = bs.x + self.f(bs).x
+        return bs
+
+
+######################################################################
+
+# A BracketedSequence is a BxTx... tensor with a first and a nb time
+# steps to compute.
+
+# Modules able to process it expect that they will have to process a
+# first bracket starting at t=0, followed by a succession of brackets
+# that move forward in time, do not overlap, and cover the axis T with
+# no holes.
+#
+# Although it is more general, for a classical prompt-conditioned
+# auto-regressive process it will be a first bracket starting at 0 and
+# of arbitrary length for the "prompt", followed by brackets of length
+# 1 for the successive tokens.
+#
+# Modules able to process brackets may implement a cache that is
+# resetted when the input bracket starts at t=0
+
+
+class BracketedSequence:
+    def __init__(self, x, first=None, nb=None):
+        self.x = x
+        self.first = 0 if first is None else first
+        self.nb = x.size(1) if nb is None else nb
+
+    def slice(self):
+        return self.x[:, self.first : self.first + self.nb]
+
+
+######################################################################
+
+
+class CacheWrapper(nn.Module):
+    def __init__(self, *f):
+        super().__init__()
+        self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
+
+    def forward(self, bs):
+        if bs.first == 0:
+            y = self.f(bs.slice())
+            self.cache_y = y.new(*((y.size(0), bs.x.size(1)) + y.size()[2:]))
+            self.cache_y[:, bs.first : bs.first + bs.nb] = y
+        else:
+            self.cache_y[:, bs.first : bs.first + bs.nb] = self.f(bs.slice())
+
+        bs.x = self.cache_y
+
+        return bs
+
+
+##############################
+
+
+class AddPositionalEncoding(nn.Module):
+    def __init__(self, len_max):
+        super().__init__()
+        self.len_max = len_max
+
+    # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
+
+    def forward(self, bs):
+        if bs.first == 0:
+            t = torch.arange(bs.x.size(1), dtype=bs.x.dtype, device=bs.x.device)[
+                :, None
+            ]
+            j = torch.arange(bs.x.size(2), dtype=bs.x.dtype, device=bs.x.device)[
+                None, :
+            ]
+            k = j % 2
+            self.pe = torch.sin(
+                t / (self.len_max ** ((j - k) / bs.x.size(2))) + math.pi / 2 * k
+            )
+            self.cache_y = bs.x.new(bs.x.size())
+
+        self.cache_y[:, bs.first : bs.first + bs.nb] = (
+            bs.slice() + self.pe[bs.first : bs.first + bs.nb]
+        )
+
+        bs.x = self.cache_y
+
+        return bs
+
+
+##############################
+
+
+class QKVAttention(nn.Module):
+    def __init__(
+        self, dim_in, dim_qk, dim_v, nb_heads=1, causal=False, attention_dropout=0.0
+    ):
+        super().__init__()
+
+        def randw(*d):
+            return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
+
+        self.causal = causal
+        self.attention_dropout = attention_dropout
+
+        self.w_q = randw(nb_heads, dim_qk, dim_in)
+        self.w_k = randw(nb_heads, dim_qk, dim_in)
+        self.w_v = randw(nb_heads, dim_v, dim_in)
+        self.w_o = randw(dim_v * nb_heads, dim_in)
+
+    def forward(self, bs_q):
+        x_q = bs_q.x
+
+        if bs_q.first == 0:
+            self.cache_k = x_q.new_zeros(
+                x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
+            )
+            self.cache_v = x_q.new_zeros(
+                x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
+            )
+            self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
+
+        q = torch.einsum(
+            "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_q
+        )
+        self.cache_k[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
+            "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_k
+        )
+        self.cache_v[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
+            "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_v
+        )
+
+        a = torch.einsum(
+            "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
+        ) / math.sqrt(self.w_q.size(1))
+
+        if self.causal:
+            if bs_q.first == 0:
+                self.cache_attzero = (
+                    torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
+                    < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
+                )
+            a = a.masked_fill(
+                self.cache_attzero[
+                    :, :, bs_q.first : bs_q.first + bs_q.nb, : bs_q.first + bs_q.nb
+                ],
+                float("-inf"),
+            )
+
+        a = a.softmax(dim=3)
+        a = F.dropout(a, self.attention_dropout, self.training)
+
+        y = torch.einsum(
+            "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs_q.first + bs_q.nb]
+        ).flatten(2)
+
+        self.cache_y[:, bs_q.first : bs_q.first + bs_q.nb] = y @ self.w_o
+
+        bs_q.x = self.cache_y
+
+        return bs_q
+
+
+##############################
+
+
+class MyGPT(nn.Module):
+    def __init__(
+        self,
+        vocabulary_size,
+        dim_model,
+        dim_keys,
+        dim_hidden,
+        nb_heads,
+        nb_blocks,
+        causal=False,
+        dropout=0.0,
+        len_max=1e5,
+    ):
+
+        super().__init__()
+
+        assert dim_model % nb_heads == 0
+
+        self.embedding = nn.Sequential(
+            CacheWrapper(nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)),
+            AddPositionalEncoding(len_max),
+        )
+
+        trunk_blocks = []
+
+        for b in range(nb_blocks):
+            trunk_blocks += [
+                WithResidual(
+                    CacheWrapper(nn.LayerNorm((dim_model,))),
+                    QKVAttention(
+                        dim_in=dim_model,
+                        dim_qk=dim_keys,
+                        dim_v=dim_model // nb_heads,
+                        nb_heads=nb_heads,
+                        causal=causal,
+                        attention_dropout=dropout,
+                    ),
+                ),
+                WithResidual(
+                    CacheWrapper(
+                        nn.LayerNorm((dim_model,)),
+                        nn.Linear(in_features=dim_model, out_features=dim_hidden),
+                        nn.ReLU(),
+                        nn.Linear(in_features=dim_hidden, out_features=dim_model),
+                        nn.Dropout(dropout),
+                    ),
+                ),
+            ]
+
+        self.trunk = nn.Sequential(*trunk_blocks)
+
+        self.readout = CacheWrapper(
+            nn.Linear(in_features=dim_model, out_features=vocabulary_size)
+        )
+
+        with torch.no_grad():
+            for m in self.modules():
+                if isinstance(m, nn.Embedding):
+                    m.weight.normal_(mean=0, std=2e-2)
+                elif isinstance(m, nn.LayerNorm):
+                    m.bias.zero_()
+                    m.weight.fill_(1.0)
+
+    def forward(self, bs):
+        bs.x = F.pad(bs.x, (1, -1))
+        bs = self.embedding(bs)
+        bs = self.trunk(bs)
+        bs = self.readout(bs)
+        return bs
+
+
+######################################################################
+
+if __name__ == "__main__":
+
+    print("Basic check.")
+
+    vocabulary_size = 10
+    x = torch.randint(vocabulary_size, (9, 7))
+
+    model = MyGPT(
+        vocabulary_size=vocabulary_size,
+        dim_model=18,
+        dim_keys=50,
+        dim_hidden=100,
+        nb_heads=2,
+        nb_blocks=1,
+        dropout=0.1,
+    )
+
+    model.eval()
+
+    y1 = model(BracketedSequence(x)).x
+
+    y2 = torch.randn_like(y1)
+    for s in range(x.size(1)):
+        z = model(BracketedSequence(x, s, 1))
+        y2[:, s] = z.x[:, s]
+
+    # print(y1.max(dim = 2).values)
+    # print(y2.max(dim = 2).values)
+    print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
+
+######################################################################