X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=78005274f3a9c9d73a52d96fbf207a3540039461;hb=2ee976b3249254d1eb796678c5408ea45293489b;hp=bd17365c47e355c97824986c5c0d76f7ae8f3279;hpb=7cc9d20fbf195eeec10d171b280089929b30659a;p=beaver.git

diff --git a/beaver.py b/beaver.py
index bd17365..7800527 100755
--- a/beaver.py
+++ b/beaver.py
@@ -66,6 +66,8 @@ parser.add_argument("--deterministic_synthesis", action="store_true", default=Fa
 
 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)
@@ -131,35 +133,35 @@ for n in vars(args):
 ######################################################################
 
 
-def generation_order(x, fixed_len):
-    if args.random_regression_order:
-        order = torch.rand(x.size(), device=x.device)
-        order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=order.device)
-        order = order.sort(1).indices
-    else:
-        order = (
-            torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
-        )
-    return order
-
-
-def reorder(x, order, back=False):  # x is NxTxD1x...xDk, order is NxT'
+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 back:
-        v = u.new(u.size())
-        v.scatter_(1, order, u)
+    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, fixed_len):
-    order = generation_order(x, fixed_len)
+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
@@ -167,7 +169,9 @@ def shuffle(x, fixed_len):
 
 
 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
-    for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+    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
@@ -186,7 +190,7 @@ def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None)
 ######################################################################
 
 
-def compute_perplexity(model, fixed_len, split="train"):
+def compute_perplexity(model, task, prompt_len, split="train"):
     with torch.autograd.no_grad():
         t = model.training
         model.eval()
@@ -195,10 +199,12 @@ def compute_perplexity(model, fixed_len, split="train"):
 
         for input in task.batches(split=split):
             input = input.to(device)
-            x, order = shuffle(input, fixed_len)
-            x = model(mygpt.BracketedSequence(x), order=order).x
-            output = reorder(x, order, back=True)
-            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)
 
@@ -227,7 +233,7 @@ def oneshot_trace_loss(mazes, output, policies, height, width):
     return (output - targets).abs().sum() / masks.sum()
 
 
-def oneshot(gpt, task):
+def oneshot(gpt, learning_rate_scheduler, task):
     t = gpt.training
     gpt.eval()
 
@@ -255,15 +261,19 @@ 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"):
-            x, order = shuffle(mazes, task.height * task.width)
-            x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
-            output_gpt = reorder(x, order, back=True)
+            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)
@@ -274,11 +284,13 @@ 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"):
-            x, order = shuffle(mazes, task.height * task.width)
-            x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
-            output_gpt = reorder(x, order, back=True)
+            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)
@@ -291,9 +303,9 @@ def oneshot(gpt, task):
         # -------------------
         mazes = task.test_input[:32, : task.height * task.width]
         policies = task.test_policies[:32]
-        x, order = shuffle(mazes, task.height * task.width)
-        x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
-        output_gpt = reorder(x, order, back=True)
+        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)
@@ -331,6 +343,75 @@ 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", nb_to_use=-1, desc=None):
         pass
@@ -438,7 +519,7 @@ class TaskMaze(Task):
             masked_inplace_autoregression(
                 model, self.batch_size, x, ar_mask, order=order
             )
-            result = reorder(x, order, back=True)
+            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)
@@ -469,7 +550,11 @@ 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(input)
             _, predicted_paths = self.seq2map(result)
@@ -508,6 +593,20 @@ 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
+
+
+amm_generator = None
+
+if args.noncausal_prompt:
+    amm_generator = noncausal_prompt_amm_generator
+
 model = mygpt.MyGPT(
     vocabulary_size=vocabulary_size,
     dim_model=args.dim_model,
@@ -517,6 +616,7 @@ model = mygpt.MyGPT(
     nb_blocks=args.nb_blocks,
     causal=True,
     dropout=args.dropout,
+    amm_generator=amm_generator,
 )
 
 model.to(device)
@@ -526,6 +626,36 @@ 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:
@@ -537,6 +667,7 @@ 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"])
@@ -546,9 +677,9 @@ 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)
 
 ######################################################################
 
@@ -561,37 +692,13 @@ 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, fixed_len=task.height * task.width, split="train"
+        model, task, prompt_len=task.height * task.width, split="train"
     )
     test_perplexity = compute_perplexity(
-        model, fixed_len=task.height * task.width, split="test"
+        model, task, prompt_len=task.height * task.width, split="test"
     )
 
     log_string(
@@ -602,10 +709,11 @@ if nb_epochs_finished >= args.nb_epochs:
 
 ##############################
 
-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)
@@ -622,10 +730,12 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 
     for input in task.batches(split="train"):
         input = input.to(device)
-        x, order = shuffle(input, task.height * task.width)
-        x = model(mygpt.BracketedSequence(x), order=order).x
-        output = reorder(x, order, back=True)
-        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)
 
@@ -633,9 +743,11 @@ 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, fixed_len=task.height * task.width, split="test"
+        model, task, prompt_len=task.height * task.width, split="test"
     )
 
     log_string(
@@ -647,6 +759,7 @@ 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(),
     }
 
@@ -660,6 +773,6 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 ######################################################################
 
 if args.oneshot:
-    oneshot(model, task)
+    oneshot(model, learning_rate_scheduler, task)
 
 ######################################################################