X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=5abe39b767c13299d8dcffcc3369682bfbfff69f;hb=HEAD;hp=f5bd9245319a37f65f655ed7bb05cfbf81563383;hpb=cd3fae5ff5e9b4c76b533dac8bc64cc331ecf51e;p=beaver.git

diff --git a/beaver.py b/beaver.py
index f5bd924..5abe39b 100755
--- a/beaver.py
+++ b/beaver.py
@@ -64,6 +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)
@@ -86,7 +90,7 @@ 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="policy")
+parser.add_argument("--oneshot_output", type=str, default="trace")
 
 ######################################################################
 
@@ -123,24 +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:
             # Needed to initialize the model's cache
-            model(mygpt.BracketedSequence(input, 0, i.min()))
+            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)
@@ -153,7 +192,7 @@ def masked_inplace_autoregression(model, batch_size, input, ar_mask):
 ######################################################################
 
 
-def compute_perplexity(model, split="train"):
+def compute_perplexity(model, task, prompt_len, split="train"):
     with torch.autograd.no_grad():
         t = model.training
         model.eval()
@@ -162,9 +201,12 @@ def compute_perplexity(model, split="train"):
 
         for input in task.batches(split=split):
             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=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)
 
@@ -193,7 +235,39 @@ def oneshot_trace_loss(mazes, output, policies, height, width):
     return (output - targets).abs().sum() / masks.sum()
 
 
-def oneshot(gpt, task):
+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()
 
@@ -221,13 +295,19 @@ def oneshot(gpt, task):
         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_schedule[n_epoch]
+        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 = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
+            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)
@@ -238,9 +318,13 @@ def oneshot(gpt, task):
             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 = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
+            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)
@@ -251,9 +335,11 @@ def oneshot(gpt, task):
         )
 
         # -------------------
-        mazes = task.test_input[:32, : task.height * task.width]
-        policies = task.test_policies[:32]
-        output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
+        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)
@@ -272,15 +358,17 @@ def oneshot(gpt, task):
         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,
-                f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
-            ),
+            os.path.join(args.result_dir, filename),
             mazes=mazes,
             score_paths=scores,
             score_truth=targets,
         )
+        log_string(f"wrote {filename}")
+
         # -------------------
 
     gpt.train(t)
@@ -289,8 +377,77 @@ def oneshot(gpt, task):
 ######################################################################
 
 
+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):
@@ -350,17 +507,19 @@ class TaskMaze(Task):
 
         self.nb_codes = self.train_input.max() + 1
 
-    def batches(self, split="train", nb_to_use=-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):
+    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
@@ -371,10 +530,12 @@ class TaskMaze(Task):
             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=f"epoch-{split}",
+            desc=desc,
         ):
             yield batch
 
@@ -388,7 +549,11 @@ class TaskMaze(Task):
             ar_mask = result.new_zeros(result.size())
             ar_mask[:, self.height * self.width :] = 1
             result *= 1 - ar_mask
-            masked_inplace_autoregression(model, self.batch_size, result, 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)
@@ -414,22 +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
             result *= 1 - ar_mask
-            masked_inplace_autoregression(model, self.batch_size, result, 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)
+            filename = f"result_{n_epoch:04d}.png"
             maze.save_image(
-                os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
+                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)
 
@@ -456,6 +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,
@@ -465,6 +660,7 @@ model = mygpt.MyGPT(
     nb_blocks=args.nb_blocks,
     causal=True,
     dropout=args.dropout,
+    amm_generator=amm_generator,
 )
 
 model.to(device)
@@ -474,6 +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:
@@ -485,6 +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"])
@@ -494,9 +721,15 @@ 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)
+
+######################################################################
+
+if args.oneshot:
+    oneshot(model, learning_rate_scheduler, task)
+    exit(0)
 
 ######################################################################
 
@@ -509,34 +742,14 @@ 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}")
-
-##############################
-
 if nb_epochs_finished >= args.nb_epochs:
     n_epoch = nb_epochs_finished
-    train_perplexity = compute_perplexity(model, split="train")
-    test_perplexity = compute_perplexity(model, split="test")
+    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"
+    )
 
     log_string(
         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
@@ -544,14 +757,13 @@ if nb_epochs_finished >= args.nb_epochs:
 
     task.produce_results(n_epoch, model)
 
-    exit(0)
-
 ##############################
 
-for n_epoch in range(nb_epochs_finished, args.nb_epochs):
-    learning_rate = learning_rate_schedule[n_epoch]
+learning_rate_scheduler.reset()
 
-    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)
@@ -568,8 +780,12 @@ for n_epoch in range(nb_epochs_finished, args.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)
 
@@ -577,8 +793,12 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
         loss.backward()
         optimizer.step()
 
+    learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
+
     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-    test_perplexity = compute_perplexity(model, split="test")
+    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}"
@@ -589,6 +809,7 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
     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(),
     }
 
@@ -600,8 +821,3 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
     log_string(f"saved checkpoint {checkpoint_name}")
 
 ######################################################################
-
-if args.oneshot:
-    oneshot(model, task)
-
-######################################################################