X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=5abe39b767c13299d8dcffcc3369682bfbfff69f;hb=HEAD;hp=920a446f920e6cd2beb67ac0df96457bfac55225;hpb=61cd7a140e44ccb966bad941fa31e395e51e50e2;p=beaver.git

diff --git a/beaver.py b/beaver.py
index 920a446..5abe39b 100755
--- a/beaver.py
+++ b/beaver.py
@@ -26,9 +26,7 @@ else:
 
 ######################################################################
 
-parser = argparse.ArgumentParser(
-    description="An implementation of GPT with cache to solve a toy geometric reasoning task."
-)
+parser = argparse.ArgumentParser(description="A maze shortest path solving with a GPT.")
 
 parser.add_argument("--log_filename", type=str, default="train.log")
 
@@ -66,6 +64,10 @@ parser.add_argument("--dropout", type=float, default=0.1)
 
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 
+parser.add_argument("--random_regression_order", action="store_true", default=False)
+
+parser.add_argument("--noncausal_prompt", 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)
@@ -75,11 +77,20 @@ parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 ##############################
 # maze options
 
-parser.add_argument("--world_height", type=int, default=13)
+parser.add_argument("--maze_height", type=int, default=13)
 
-parser.add_argument("--world_width", type=int, default=21)
+parser.add_argument("--maze_width", type=int, default=21)
+
+parser.add_argument("--maze_nb_walls", type=int, default=15)
+
+##############################
+# one-shot prediction
 
-parser.add_argument("--world_nb_walls", type=int, default=15)
+parser.add_argument("--oneshot", action="store_true", default=False)
+
+parser.add_argument("--oneshot_input", type=str, default="head")
+
+parser.add_argument("--oneshot_output", type=str, default="trace")
 
 ######################################################################
 
@@ -116,26 +127,59 @@ def log_string(s):
     sys.stdout.flush()
 
 
+log_string(f"cmd {' '.join(sys.argv)}")
+
 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
 ######################################################################
 
 
+def reorder(x, order, reverse=False):  # x is NxTxD1x...xDk, order is NxT'
+    u = x.reshape(x.size()[:2] + (-1,))
+    order = order.unsqueeze(-1).expand(-1, -1, u.size(-1))
+    if reverse:
+        v = u.new(u.size()).scatter_(1, order, u)
+    else:
+        v = u.gather(1, order)
+    v = v.reshape(v.size()[:2] + x.size()[2:])
+    return v
+
+
+def shuffle(x, prompt_len):
+    if args.random_regression_order:
+        order = torch.rand(x.size(), device=x.device)
+        order[:, :prompt_len] = torch.arange(-prompt_len, 0, device=x.device)
+        order = order.sort(1).indices
+    else:
+        order = (
+            torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
+        )
+    return reorder(x, order), order
+
+
+def eval_mygpt(model, input, mode="standard", prompt_len=0):
+    x, order = shuffle(input, prompt_len)
+    x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
+    return reorder(x, order, reverse=True)
+
+
+######################################################################
+
 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 # tokens that should be generated
 
 
-def masked_inplace_autoregression(model, batch_size, input, ar_mask):
-
-    for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
+    for input, ar_mask, order in zip(
+        input.split(batch_size), ar_mask.split(batch_size), order.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
+            # Needed to initialize the model's cache
+            model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
         for s in range(i.min(), i.max() + 1):
-            output = model(mygpt.BracketedSequence(input, s, 1)).x
+            output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
             logits = output[:, s]
             if args.deterministic_synthesis:
                 t_next = logits.argmax(1)
@@ -148,8 +192,262 @@ def masked_inplace_autoregression(model, batch_size, input, ar_mask):
 ######################################################################
 
 
+def compute_perplexity(model, task, prompt_len, split="train"):
+    with torch.autograd.no_grad():
+        t = model.training
+        model.eval()
+
+        nb_samples, acc_loss = 0, 0.0
+
+        for input in task.batches(split=split):
+            input = input.to(device)
+            output = eval_mygpt(model, input, prompt_len=prompt_len)
+            if args.noncausal_prompt:
+                d = input.size(1) // 2
+                loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
+            else:
+                loss = F.cross_entropy(output.transpose(1, 2), input)
+            acc_loss += loss.item() * input.size(0)
+            nb_samples += input.size(0)
+
+        model.train(t)
+
+        return math.exp(min(100, acc_loss / nb_samples))
+
+
+######################################################################
+
+
+def oneshot_policy_loss(mazes, output, policies, height, width):
+    masks = (mazes == maze.v_empty).unsqueeze(-1)
+    targets = policies.permute(0, 2, 1) * masks
+    output = output * masks
+    return -(output.log_softmax(-1) * targets).sum() / masks.sum()
+
+
+def oneshot_trace_loss(mazes, output, policies, height, width):
+    masks = mazes == maze.v_empty
+    targets = maze.stationary_densities(
+        mazes.view(-1, height, width), policies.view(-1, 4, height, width)
+    ).flatten(-2)
+    targets = targets * masks
+    output = output.squeeze(-1) * masks
+    return (output - targets).abs().sum() / masks.sum()
+
+
+def oneshot(model, learning_rate_scheduler, task):
+    t = model.training
+    model.eval()
+    mazes = task.test_input[:48].clone()
+    mazes[:, task.height * task.width :] = 0
+    policies = task.test_policies[:48]
+    targets = maze.stationary_densities(
+        mazes[:, : task.height * task.width].view(-1, task.height, task.width),
+        policies.view(-1, 4, task.height, task.width),
+    ).flatten(-2)
+    output = eval_mygpt(model, mazes, prompt_len=task.height * task.width)
+    output = F.softmax(output, dim=2)
+    print(f"{output.size()=}")
+    proba_path = output[:, task.height * task.width :, 4].reshape(
+        -1, task.height, task.width
+    )
+    mazes = mazes[:, : task.height * task.width].reshape(-1, task.height, task.width)
+    targets = targets.reshape(-1, task.height, task.width)
+    paths = task.test_input[:48, task.height * task.width :].reshape(
+        -1, task.height, task.width
+    )
+    filename = f"oneshot.png"
+    maze.save_image(
+        os.path.join(args.result_dir, filename),
+        mazes=mazes,
+        # target_paths=paths,
+        score_paths=proba_path,
+        score_truth=targets,
+    )
+    log_string(f"wrote {filename}")
+
+
+def oneshot_old(gpt, learning_rate_scheduler, task):
+    t = gpt.training
+    gpt.eval()
+
+    if args.oneshot_input == "head":
+        dim_in = args.dim_model
+    elif args.oneshot_input == "deep":
+        dim_in = args.dim_model * args.nb_blocks * 2
+    else:
+        raise ValueError(f"{args.oneshot_input=}")
+
+    if args.oneshot_output == "policy":
+        dim_out = 4
+        compute_loss = oneshot_policy_loss
+    elif args.oneshot_output == "trace":
+        dim_out = 1
+        compute_loss = oneshot_trace_loss
+    else:
+        raise ValueError(f"{args.oneshot_output=}")
+
+    model = nn.Sequential(
+        nn.Linear(dim_in, args.dim_model),
+        nn.ReLU(),
+        nn.Linear(args.dim_model, args.dim_model),
+        nn.ReLU(),
+        nn.Linear(args.dim_model, dim_out),
+    ).to(device)
+
+    learning_rate_scheduler.reset()
+
+    for n_epoch in range(args.nb_epochs):
+        learning_rate = learning_rate_scheduler.get_learning_rate()
+        log_string(f"learning_rate {n_epoch} {learning_rate}")
+
+        optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
+
+        acc_train_loss, nb_train_samples = 0, 0
+        for mazes, policies in task.policy_batches(split="train"):
+            output_gpt = eval_mygpt(
+                gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
+            )
+            output = model(output_gpt)
+
+            loss = compute_loss(mazes, output, policies, task.height, task.width)
+            acc_train_loss += loss.item() * mazes.size(0)
+            nb_train_samples += mazes.size(0)
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+        learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
+
+        acc_test_loss, nb_test_samples = 0, 0
+        for mazes, policies in task.policy_batches(split="test"):
+            output_gpt = eval_mygpt(
+                gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
+            )
+            output = model(output_gpt)
+            loss = compute_loss(mazes, output, policies, task.height, task.width)
+            acc_test_loss += loss.item() * mazes.size(0)
+            nb_test_samples += mazes.size(0)
+
+        log_string(
+            f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
+        )
+
+        # -------------------
+        mazes = task.test_input[:48, : task.height * task.width]
+        policies = task.test_policies[:48]
+        output_gpt = eval_mygpt(
+            gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
+        )
+        output = model(output_gpt)
+        if args.oneshot_output == "policy":
+            targets = policies.permute(0, 2, 1)
+            scores = (
+                (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
+            ).float()
+        elif args.oneshot_output == "trace":
+            targets = maze.stationary_densities(
+                mazes.view(-1, task.height, task.width),
+                policies.view(-1, 4, task.height, task.width),
+            ).flatten(-2)
+            scores = output
+        else:
+            raise ValueError(f"{args.oneshot_output=}")
+
+        scores = scores.reshape(-1, task.height, task.width)
+        mazes = mazes.reshape(-1, task.height, task.width)
+        targets = targets.reshape(-1, task.height, task.width)
+        filename = (
+            f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
+        )
+        maze.save_image(
+            os.path.join(args.result_dir, filename),
+            mazes=mazes,
+            score_paths=scores,
+            score_truth=targets,
+        )
+        log_string(f"wrote {filename}")
+
+        # -------------------
+
+    gpt.train(t)
+
+
+######################################################################
+
+
+class LearningRateScheduler:
+    def get_learning_rate(self):
+        pass
+
+    def update(self, nb_finished_epochs, loss):
+        pass
+
+    def reset(self):
+        pass
+
+    def get_state(self):
+        return vars(self)
+
+    def set_state(self, state):
+        print(f"{state=}")
+        for k, v in state.items():
+            setattr(self, k, v)
+
+
+class StepWiseScheduler(LearningRateScheduler):
+    def __init__(self, schedule):
+        self.nb_finished_epochs = 0
+        self.schedule = schedule
+
+    def get_learning_rate(self):
+        return self.schedule[self.nb_finished_epochs]
+
+    def update(self, nb_finished_epochs, loss):
+        self.nb_finished_epochs = nb_finished_epochs
+
+    def reset(self):
+        self.nb_finished_epochs = 0
+
+    def get_state(self):
+        return {"nb_finished_epochs": self.nb_finished_epochs}
+
+
+class AutoScheduler(LearningRateScheduler):
+    def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
+        self.learning_rate_init = learning_rate_init
+        self.learning_rate = learning_rate_init
+        self.growth = growth
+        self.degrowth = degrowth
+        self.pred_loss = None
+
+    def get_learning_rate(self):
+        return self.learning_rate
+
+    def update(self, nb_finished_epochs, loss):
+        if self.pred_loss is not None:
+            if loss >= self.pred_loss:
+                self.learning_rate *= self.degrowth
+            else:
+                self.learning_rate *= self.growth
+        self.pred_loss = loss
+
+    def reset(self):
+        self.learning_rate = self.learning_rate_init
+
+    def get_state(self):
+        return {
+            "learning_rate_init": self.learning_rate_init,
+            "pred_loss": self.pred_loss,
+        }
+
+
+######################################################################
+
+
 class Task:
-    def batches(self, split="train"):
+    def batches(self, split="train", nb_to_use=-1, desc=None):
         pass
 
     def vocabulary_size(self):
@@ -187,34 +485,57 @@ class TaskMaze(Task):
         self.width = width
         self.device = device
 
-        mazes_train, paths_train = maze.create_maze_data(
+        train_mazes, train_paths, train_policies = maze.create_maze_data(
             nb_train_samples,
             height=height,
             width=width,
             nb_walls=nb_walls,
             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
         )
-        mazes_train, paths_train = mazes_train.to(device), paths_train.to(device)
-        self.train_input = self.map2seq(mazes_train, paths_train)
-        self.nb_codes = self.train_input.max() + 1
+        self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
+        self.train_policies = train_policies.flatten(-2).to(device)
 
-        mazes_test, paths_test = maze.create_maze_data(
+        test_mazes, test_paths, test_policies = maze.create_maze_data(
             nb_test_samples,
             height=height,
             width=width,
             nb_walls=nb_walls,
             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
         )
-        mazes_test, paths_test = mazes_test.to(device), paths_test.to(device)
-        self.test_input = self.map2seq(mazes_test, paths_test)
+        self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
+        self.test_policies = test_policies.flatten(-2).to(device)
 
-    def batches(self, split="train", nb_to_use=-1):
+        self.nb_codes = self.train_input.max() + 1
+
+    def batches(self, split="train", nb_to_use=-1, desc=None):
         assert split in {"train", "test"}
         input = self.train_input if split == "train" else self.test_input
         if nb_to_use > 0:
             input = input[:nb_to_use]
+        if desc is None:
+            desc = f"epoch-{split}"
         for batch in tqdm.tqdm(
-            input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
+            input.split(self.batch_size), dynamic_ncols=True, desc=desc
+        ):
+            yield batch
+
+    def policy_batches(self, split="train", nb_to_use=-1, desc=None):
+        assert split in {"train", "test"}
+        input = self.train_input if split == "train" else self.test_input
+        policies = self.train_policies if split == "train" else self.test_policies
+        input = input[:, : self.height * self.width]
+        policies = policies * (input != maze.v_wall)[:, None]
+
+        if nb_to_use > 0:
+            input = input[:nb_to_use]
+            policies = policies[:nb_to_use]
+
+        if desc is None:
+            desc = f"epoch-{split}"
+        for batch in tqdm.tqdm(
+            zip(input.split(self.batch_size), policies.split(self.batch_size)),
+            dynamic_ncols=True,
+            desc=desc,
         ):
             yield batch
 
@@ -227,7 +548,12 @@ class TaskMaze(Task):
             result = input.clone()
             ar_mask = result.new_zeros(result.size())
             ar_mask[:, self.height * self.width :] = 1
-            masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
+            result *= 1 - ar_mask
+            x, order = shuffle(result, self.height * self.width)
+            masked_inplace_autoregression(
+                model, self.batch_size, x, ar_mask, order=order
+            )
+            result = reorder(x, order, reverse=True)
             mazes, paths = self.seq2map(result)
             nb_correct += maze.path_correctness(mazes, paths).long().sum()
             nb_total += mazes.size(0)
@@ -253,16 +579,28 @@ class TaskMaze(Task):
                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
             )
 
-            input = self.test_input[:32]
+            input = self.test_input[:48]
             result = input.clone()
             ar_mask = result.new_zeros(result.size())
-
             ar_mask[:, self.height * self.width :] = 1
-            masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
+            result *= 1 - ar_mask
+            x, order = shuffle(result, self.height * self.width)
+            masked_inplace_autoregression(
+                model, self.batch_size, x, ar_mask, order=order
+            )
+            result = reorder(x, order, reverse=True)
 
             mazes, paths = self.seq2map(input)
             _, predicted_paths = self.seq2map(result)
-            maze.save_image(f"result_{n_epoch:04d}.png", mazes, paths, predicted_paths)
+            filename = f"result_{n_epoch:04d}.png"
+            maze.save_image(
+                os.path.join(args.result_dir, filename),
+                mazes=mazes,
+                target_paths=paths,
+                predicted_paths=predicted_paths,
+                path_correct=maze.path_correctness(mazes, predicted_paths),
+            )
+            log_string(f"wrote {filename}")
 
             model.train(t)
 
@@ -276,9 +614,9 @@ task = TaskMaze(
     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,
     batch_size=args.batch_size,
-    height=args.world_height,
-    width=args.world_width,
-    nb_walls=args.world_nb_walls,
+    height=args.maze_height,
+    width=args.maze_width,
+    nb_walls=args.maze_nb_walls,
     device=device,
 )
 
@@ -289,6 +627,30 @@ log_string(f"vocabulary_size {vocabulary_size}")
 
 ##############################
 
+
+def noncausal_prompt_amm_generator(d):
+    q = torch.arange(d)[:, None]
+    k = torch.arange(d)[None, :]
+    s = args.maze_height * args.maze_width
+    return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
+    # return q < k
+
+
+def noncausal_prompt_oneshot_amm_generator(d):
+    q = torch.arange(d)[:, None]
+    k = torch.arange(d)[None, :]
+    s = args.maze_height * args.maze_width
+    return k >= s
+    # return q < k
+
+
+if args.oneshot:
+    amm_generator = noncausal_prompt_oneshot_amm_generator
+elif args.noncausal_prompt:
+    amm_generator = noncausal_prompt_amm_generator
+else:
+    amm_generator = None
+
 model = mygpt.MyGPT(
     vocabulary_size=vocabulary_size,
     dim_model=args.dim_model,
@@ -298,6 +660,7 @@ model = mygpt.MyGPT(
     nb_blocks=args.nb_blocks,
     causal=True,
     dropout=args.dropout,
+    amm_generator=amm_generator,
 )
 
 model.to(device)
@@ -307,6 +670,36 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
+if args.learning_rate_schedule == "auto":
+    learning_rate_scheduler = AutoScheduler(args.learning_rate)
+
+elif args.learning_rate_schedule == "cos":
+    schedule = {}
+    for n_epoch in range(args.nb_epochs):
+        u = n_epoch / args.nb_epochs * math.pi
+        schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
+    learning_rate_scheduler = StepWiseScheduler(schedule)
+    log_string(f"learning_rate_schedule {schedule}")
+
+else:
+    u = {
+        int(k): float(v)
+        for k, v in [
+            tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
+        ]
+    }
+
+    schedule = {}
+    learning_rate = args.learning_rate
+    for n_epoch in range(args.nb_epochs):
+        if n_epoch in u:
+            learning_rate = u[n_epoch]
+        schedule[n_epoch] = learning_rate
+    learning_rate_scheduler = StepWiseScheduler(schedule)
+    log_string(f"learning_rate_schedule {schedule}")
+
+######################################################################
+
 nb_epochs_finished = 0
 
 if args.no_checkpoint:
@@ -318,6 +711,7 @@ else:
         checkpoint = torch.load(checkpoint_name)
         nb_epochs_finished = checkpoint["nb_epochs_finished"]
         model.load_state_dict(checkpoint["model_state"])
+        learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
         torch.set_rng_state(checkpoint["rng_state"])
         if torch.cuda.is_available():
             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
@@ -327,13 +721,17 @@ else:
     except FileNotFoundError:
         log_string("starting from scratch.")
 
-    except:
-        log_string("error when loading the checkpoint.")
-        exit(1)
+    # except:
+    # log_string("error when loading the checkpoint.")
+    # exit(1)
 
 ######################################################################
 
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
+if args.oneshot:
+    oneshot(model, learning_rate_scheduler, task)
+    exit(0)
+
+######################################################################
 
 token_count = 0
 for input in task.batches(split="train"):
@@ -344,40 +742,28 @@ 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(",")
-        ]
-    }
+if nb_epochs_finished >= args.nb_epochs:
+    n_epoch = nb_epochs_finished
+    train_perplexity = compute_perplexity(
+        model, task, prompt_len=task.height * task.width, split="train"
+    )
+    test_perplexity = compute_perplexity(
+        model, task, prompt_len=task.height * task.width, split="test"
+    )
 
-    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"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
+    )
 
-log_string(f"learning_rate_schedule {learning_rate_schedule}")
+    task.produce_results(n_epoch, model)
 
 ##############################
 
-nb_samples_seen = 0
+learning_rate_scheduler.reset()
 
-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}")
+for n_epoch in range(nb_epochs_finished, args.nb_epochs):
+    learning_rate = learning_rate_scheduler.get_learning_rate()
+    log_string(f"learning_rate {n_epoch} {learning_rate}")
 
     if args.optim == "sgd":
         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
@@ -386,7 +772,7 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
     elif args.optim == "adamw":
         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
     else:
-        raise ValueError(f"Unknown optimizer {args.optim}.")
+        raise ValueError(f"{args.optim=}")
 
     model.train()
 
@@ -394,45 +780,36 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
 
     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)
+        output = eval_mygpt(model, input, prompt_len=task.height * task.width)
+        if args.noncausal_prompt:
+            d = input.size(1) // 2
+            loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
+        else:
+            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)
+    learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 
-        train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+    test_perplexity = compute_perplexity(
+        model, task, prompt_len=task.height * task.width, split="test"
+    )
 
-        log_string(
-            f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
-        )
+    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)
+    task.produce_results(n_epoch, model)
 
     checkpoint = {
         "nb_epochs_finished": n_epoch + 1,
         "model_state": model.state_dict(),
+        "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
         "rng_state": torch.get_rng_state(),
     }