Update.
authorFrançois Fleuret <francois@fleuret.org>
Sun, 10 Dec 2023 15:00:57 +0000 (16:00 +0100)
committerFrançois Fleuret <francois@fleuret.org>
Sun, 10 Dec 2023 15:00:57 +0000 (16:00 +0100)
main.py
problems.py
tasks.py

diff --git a/main.py b/main.py
index 17936c3..1d52b6d 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -33,7 +33,7 @@ parser.add_argument(
     "--task",
     type=str,
     default="twotargets",
-    help="byheart, learnop, guessop, mixing, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
+    help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
 )
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
@@ -256,6 +256,12 @@ default_task_args = {
         "nb_train_samples": 50000,
         "nb_test_samples": 10000,
     },
+    "memory": {
+        "model": "4M",
+        "batch_size": 100,
+        "nb_train_samples": 5000,
+        "nb_test_samples": 1000,
+    },
     "mixing": {
         "model": "37M",
         "batch_size": 25,
@@ -285,6 +291,13 @@ default_model_args = {
         "nb_heads": 2,
         "nb_blocks": 2,
     },
+    "4M": {
+        "dim_model": 256,
+        "dim_keys": 32,
+        "dim_hidden": 1024,
+        "nb_heads": 4,
+        "nb_blocks": 6,
+    },
     "37M": {
         "dim_model": 512,
         "dim_keys": 64,
@@ -418,6 +431,16 @@ elif args.task == "twotargets":
         device=device,
     )
 
+elif args.task == "memory":
+    task = tasks.SandBox(
+        problem=problems.ProblemMemory(),
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        logger=log_string,
+        device=device,
+    )
+
 elif args.task == "mixing":
     task = tasks.SandBox(
         problem=problems.ProblemMixing(
index ac16df4..d7dbc54 100755 (executable)
@@ -110,6 +110,48 @@ class ProblemDegradation(Problem):
 ####################
 
 
+class ProblemMemory(Problem):
+    def __init__(self, len_total=25):
+        self.len_total = len_total
+        self.max_len_pattern = 5
+        self.nb_noise_tokens = 10
+        self.start_pattern_token = 0
+        self.end_pattern_token = 1
+        self.start_result_token = 2
+        self.end_result_token = 3
+        self.token_string = "[]<>" + "".join(
+            [chr(ord("a") + k) for k in range(self.nb_noise_tokens)]
+        )
+
+    def generate_sequences(self, nb):
+        sequences = (
+            torch.randint(self.nb_noise_tokens, (nb, self.len_total))
+            + self.end_result_token
+            + 1
+        )
+        len_patterns = torch.randint(self.max_len_pattern, (nb,)) + 1
+        pattern_positions = torch.randint(
+            self.len_total - (5 + 2 * self.max_len_pattern), (nb,)
+        )
+        k = self.len_total - (3 + self.max_len_pattern)
+        for i in range(nb):
+            l = len_patterns[i]
+            j = pattern_positions[i]
+            sequences[i, j] = self.start_pattern_token
+            sequences[i, j + l + 2] = self.end_pattern_token
+            sequences[i, k] = self.start_result_token
+            sequences[i, k + l + 2] = self.end_result_token
+            sequences[i, k + 1 : k + 2 + l] = sequences[i, j + 1 : j + 2 + l]
+
+        j = torch.arange(self.len_total)[None, :]
+        ar_mask = (j > k).long() * (j <= k + 1 + len_patterns[:, None]).long()
+
+        return sequences, ar_mask
+
+    def seq2str(self, seq):
+        return "".join(self.token_string[x.item()] for x in seq)
+
+
 class ProblemTwoTargets(Problem):
     def __init__(self, len_total=10, len_targets=3):
         assert len_targets >= 3
@@ -325,22 +367,38 @@ class ProblemMixing(Problem):
         return y
 
     def start_error(self, x):
-        i = torch.arange(self.height, device=x.device).reshape(1, -1, 1).expand_as(x)
-        j = torch.arange(self.width, device=x.device).reshape(1, 1, -1).expand_as(x)
-
-        ri = (
-            (x == self.height * self.width).long().sum(dim=-1).argmax(-1).view(-1, 1, 1)
-        )
-        rj = (
-            (x == self.height * self.width).long().sum(dim=-2).argmax(-1).view(-1, 1, 1)
-        )
+        if self.random_start:
+            i = (
+                torch.arange(self.height, device=x.device)
+                .reshape(1, -1, 1)
+                .expand_as(x)
+            )
+            j = torch.arange(self.width, device=x.device).reshape(1, 1, -1).expand_as(x)
+
+            ri = (
+                (x == self.height * self.width)
+                .long()
+                .sum(dim=-1)
+                .argmax(-1)
+                .view(-1, 1, 1)
+            )
+            rj = (
+                (x == self.height * self.width)
+                .long()
+                .sum(dim=-2)
+                .argmax(-1)
+                .view(-1, 1, 1)
+            )
 
-        m = 1 - torch.logical_or(i == ri, j == rj).long().flatten(1)
+            m = 1 - torch.logical_or(i == ri, j == rj).long().flatten(1)
+        else:
+            m = 1
 
         x = x.flatten(1)
         u = torch.arange(self.height * self.width, device=x.device).reshape(1, -1)
 
         d = (x - (m * u + (1 - m) * self.height * self.width)).abs().sum(-1)
+
         return d
 
     def moves(self, x):
@@ -424,7 +482,8 @@ class ProblemMixing(Problem):
 ####################
 
 if __name__ == "__main__":
-    p = ProblemMixing()
+    p = ProblemMixing(height=3, width=3, random_start=False)
+
     s, m = p.generate_sequences(10000)
     for x in s[:5]:
         print(p.seq2str(x))
index 7a4abbe..f4be293 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -125,6 +125,12 @@ class SandBox(Task):
             (0, 1),
         }
 
+        if logger is not None:
+            for s, a in zip(self.train_input[:100], self.train_ar_mask[:100]):
+                logger(f"train_sequences {self.problem.seq2str(s)}")
+                a = "".join(["01"[x.item()] for x in a])
+                logger(f"                {a}")
+
     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
@@ -206,30 +212,30 @@ class SandBox(Task):
                 with torch.autograd.no_grad():
                     t = model.training
                     model.eval()
-                    model.record_attention(True)
+                    model.record_attention(True)
                     model(BracketedSequence(input))
                     model.train(t)
-                    ram = model.retrieve_attention()
-                    model.record_attention(False)
-
-                tokens_output = [c for c in self.problem.seq2str(input[0])]
-                tokens_input = ["n/a"] + tokens_output[:-1]
-                for n_head in range(ram[0].size(1)):
-                    filename = os.path.join(
-                        result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
-                    )
-                    attention_matrices = [m[0, n_head] for m in ram]
-                    save_attention_image(
-                        filename,
-                        tokens_input,
-                        tokens_output,
-                        attention_matrices,
-                        k_top=10,
-                        # min_total_attention=0.9,
-                        token_gap=12,
-                        layer_gap=50,
-                    )
-                    logger(f"wrote {filename}")
+                    ram = model.retrieve_attention()
+                    model.record_attention(False)
+
+                tokens_output = [c for c in self.problem.seq2str(input[0])]
+                tokens_input = ["n/a"] + tokens_output[:-1]
+                for n_head in range(ram[0].size(1)):
+                # filename = os.path.join(
+                # result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
+                # )
+                # attention_matrices = [m[0, n_head] for m in ram]
+                # save_attention_image(
+                # filename,
+                # tokens_input,
+                # tokens_output,
+                # attention_matrices,
+                # k_top=10,
+                ##min_total_attention=0.9,
+                # token_gap=12,
+                # layer_gap=50,
+                # )
+                # logger(f"wrote {filename}")
 
 
 ######################################################################