X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=5019aed3b0953a7037bc213296ca7371d1b3c279;hb=6045e9a7dd61f0dab60bd1c6ff71f6bd5c32778b;hp=e7c2f75897160bec75e1ec2bb91c762a4be29b04;hpb=d6f73f1d5093fb098e822e14db382dd3a1c63a2a;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index e7c2f75..5019aed 100755
--- a/tasks.py
+++ b/tasks.py
@@ -1,5 +1,10 @@
 #!/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, os, tqdm
 
 import torch, torchvision
@@ -7,6 +12,13 @@ import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 
+from mygpt import BracketedSequence
+
+try:
+    from graph import save_attention_image
+except ImportError:
+    save_attention_image = None
+
 ######################################################################
 
 
@@ -29,7 +41,7 @@ def masked_inplace_autoregression(
             batches,
             dynamic_ncols=True,
             desc=progress_bar_desc,
-            # total=input.size(0) // batch_size,
+            total=(input.size(0) + batch_size - 1) // batch_size,
         )
 
     with torch.autograd.no_grad():
@@ -60,158 +72,9 @@ class Task:
         pass
 
 
-######################################################################
-
-
-class Problem:
-    def generate_sequences(self, nb):
-        pass
-
-    def seq2str(self, seq):
-        return "[NOT IMPLEMENTED]"
-
-
 ####################
 
-
-class ProblemLevel0(Problem):
-    def __init__(self, nb_sentences=100, len_prompt=5, len_result=5):
-        self.seq = torch.randint(10, (nb_sentences, len_prompt + 1 + len_result))
-        self.seq[:, len_prompt] = 10
-
-    def generate_sequences(self, nb):
-        sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
-        ar_mask = (sequences == 10).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_mask
-
-
-class ProblemLevel1(Problem):
-    def __init__(self, nb_operators=100, len_source=5, len_result=8):
-        self.len_source = len_source
-        self.len_result = len_result
-        self.len_nb_operator = int(math.log(nb_operators) / math.log(10)) + 1
-        self.operators = F.one_hot(
-            torch.rand(nb_operators, len_result, len_source).argmax(-1),
-            num_classes=len_source,
-        )
-
-    def generate_sequences(self, nb):
-        nb_operators = torch.randint(self.operators.size(0), (nb,))
-        operators = self.operators[nb_operators]
-        nb_operators = (
-            nb_operators[:, None]
-            // 10 ** torch.arange(self.len_nb_operator - 1, -1, -1)
-        ) % 10
-        marker1 = torch.full((nb, 1), 10)
-        source = torch.randint(10, (nb, self.len_source))
-        marker2 = torch.full((nb, 1), 11)
-        result = operators.bmm(source[:, :, None]).squeeze(-1)
-        print(f"{nb_operators.dtype=} {marker1.dtype=}")
-        sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
-        print(f"{sequences.size()=}")
-        ar_mask = (sequences == 11).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_mask
-
-    def seq2str(self, seq):
-        return "".join("0123456789|>"[x.item()] for x in seq)
-
-
-class ProblemLevel2(Problem):
-    def __init__(self, len_source=5, len_result=8):
-        self.len_source = len_source
-        self.len_result = len_result
-
-    def generate_sequences(self, nb):
-        operators = F.one_hot(
-            torch.rand(nb, self.len_result, self.len_source).argmax(-1),
-            num_classes=self.len_source,
-        )
-        source1 = torch.randint(10, (nb, self.len_source))
-        marker1 = torch.full((nb, 1), 10)
-        result1 = operators.bmm(source1[:, :, None]).squeeze(-1)
-        marker2 = torch.full((nb, 1), 11)
-        source2 = torch.randint(10, (nb, self.len_source))
-        marker3 = torch.full((nb, 1), 12)
-        result2 = operators.bmm(source2[:, :, None]).squeeze(-1)
-
-        sequences = torch.cat(
-            (source1, marker1, result1, marker2, source2, marker3, result2), 1
-        )
-        ar_mask = (sequences == 12).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_mask
-
-    def seq2str(self, seq):
-        return "".join("0123456789>|~"[x.item()] for x in seq)
-
-
-####################
-
-
-class ProblemAddition(Problem):
-    def __init__(self, nb_digits=10, zero_padded=False, inverted_result=False):
-        self.nb_digits = nb_digits
-        self.zero_padded = zero_padded
-        self.inverted_result = inverted_result
-        self.char2id = dict([(c, n) for n, c in enumerate("0123456789+=$")])
-        self.id2char = dict([(n, c) for c, n in self.char2id.items()])
-
-    def tensorize(self, strings):
-        len_max = max([len(x) for x in strings])
-        return torch.cat(
-            [
-                torch.tensor(
-                    [
-                        [self.char2id[c] for c in s + "$" * (len_max - len(s))]
-                        for s in strings
-                    ]
-                )
-            ],
-            0,
-        )
-
-    def generate_sequences(self, nb):
-        sequences = []
-        for k in range(nb):
-            a, b = torch.randint(10**self.nb_digits, (2,))
-            c = a + b
-            a, b, c = str(a.item()), str(b.item()), str(c.item())
-            if self.zero_padded:
-                a = "0" * (self.nb_digits - len(a)) + a
-                b = "0" * (self.nb_digits - len(b)) + b
-                c = "0" * (self.nb_digits + 1 - len(c)) + c
-            if self.inverted_result:
-                c = c[::-1]
-            sequences.append(f"{a}+{b}={c}$")
-
-        sequences = self.tensorize(sequences)
-        ar_mask = (sequences == self.char2id["="]).long()
-        ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
-        return sequences, ar_mask
-
-    def seq2str(self, seq):
-        return "".join(self.id2char[x.item()] for x in seq)
-
-
-# class ProblemUnion(Problem):
-# problems = [ProblemByheart()]
-# nb_common_codes = 100
-
-# def generate_sequences(nb_samples):
-# problem_indexes = torch.randint(len(problems), (nb_samples,))
-# nb_samples_per_problem = torch.one_hot(problem_indexes).sum(0)
-# print(f"{nb_samples_per_problem}")
-# all_seq = []
-# for nb, p in zip(nb_samples_per_problem, problems):
-# all_seq.append(p.generate_sequences(nb_samples_per_problem[nb]))
-# return all_seq
-
-# for strain, stest in zip(train_seq, test_seq):
-# s = torch.cat((strain, stest), 0)
-
-####################
+import problems
 
 
 class SandBox(Task):
@@ -318,6 +181,43 @@ class SandBox(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
+        if save_attention_image is None:
+            logger("no save_attention_image (is pycairo installed?)")
+        else:
+            for k in range(10):
+                ns = torch.randint(self.test_input.size(0), (1,)).item()
+                input = self.test_input[ns : ns + 1].clone()
+
+                with torch.autograd.no_grad():
+                    t = model.training
+                    model.eval()
+                    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}")
+
 
 ######################################################################
 
@@ -473,6 +373,10 @@ class PicoCLVR(Task):
             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
         )
 
+        logger(
+            f"main_test_accuracy {n_epoch} {1-nb_missing_properties/nb_requested_properties}"
+        )
+
     ######################################################################
 
     def produce_results(
@@ -743,6 +647,8 @@ class Maze(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         if count is not None:
             proportion_optimal = count.diagonal().sum().float() / count.sum()
             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
@@ -882,6 +788,8 @@ class Snake(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
 
 ######################################################################
 
@@ -991,6 +899,8 @@ class Stack(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         ##############################################################
         # Log a few generated sequences
         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
@@ -1019,6 +929,297 @@ class Stack(Task):
         ##############################################################
 
 
+######################################################################
+
+import rpl
+
+
+class RPL(Task):
+    def tensorize(self, sequences):
+        len_max = max([len(x) for x in sequences])
+        return torch.cat(
+            [
+                torch.tensor(
+                    [
+                        [
+                            self.token2id[str(c)]
+                            for c in s + ["<nul>"] * (len_max - len(s))
+                        ]
+                        for s in sequences
+                    ]
+                )
+            ],
+            0,
+        )
+
+    def seq2str(self, seq):
+        return " ".join([self.id2token[i] for i in seq])
+
+    def __init__(
+        self,
+        nb_train_samples,
+        nb_test_samples,
+        batch_size,
+        nb_starting_values=3,
+        max_input=9,
+        prog_len=6,
+        nb_runs=5,
+        no_prog=False,
+        logger=None,
+        device=torch.device("cpu"),
+    ):
+        super().__init__()
+
+        self.batch_size = batch_size
+        self.device = device
+        self.no_prog = no_prog
+
+        train_sequences = [
+            rpl.generate(
+                nb_starting_values=nb_starting_values,
+                nb_result_values_max=4 * nb_starting_values,
+                max_input=max_input,
+                prog_len=prog_len,
+                nb_runs=nb_runs,
+            )
+            for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
+        ]
+
+        test_sequences = [
+            rpl.generate(
+                nb_starting_values=nb_starting_values,
+                nb_result_values_max=4 * nb_starting_values,
+                max_input=max_input,
+                prog_len=prog_len,
+                nb_runs=nb_runs,
+            )
+            for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
+        ]
+
+        symbols = list(
+            set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
+        )
+        val_max = max([x if type(x) is int else 0 for x in symbols])
+        symbols = list(filter(lambda x: type(x) is str, symbols))
+        symbols.sort()
+        symbols += [str(n) for n in range(val_max + 1)]
+        self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
+        self.id2token = dict([(n, c) for c, n in self.token2id.items()])
+
+        self.t_nul = self.token2id["<nul>"]
+        self.t_input = self.token2id["<in>"]
+        self.t_output = self.token2id["<out>"]
+        self.t_prog = self.token2id["<prg>"]
+        self.t_end = self.token2id["<end>"]
+
+        self.train_input = self.tensorize(train_sequences)
+        self.test_input = self.tensorize(test_sequences)
+
+        if no_prog:
+            # Excise the program from every train and test example
+            k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
+                None, :
+            ]
+            p = (
+                ((self.train_input == self.t_prog).long() * k)
+                .max(1, keepdim=True)
+                .values
+            )
+            self.train_input = (
+                self.train_input * (k <= p).long()
+                + self.t_end * (k == p + 1).long()
+                + self.t_nul * (k > p + 1).long()
+            )
+            k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
+                None, :
+            ]
+            p = (
+                ((self.test_input == self.t_prog).long() * k)
+                .max(1, keepdim=True)
+                .values
+            )
+            self.test_input = (
+                self.test_input * (k <= p).long()
+                + self.t_end * (k == p + 1).long()
+                + self.t_nul * (k > p + 1).long()
+            )
+
+        if logger is not None:
+            logger(f"value_max {val_max}")
+            for x in self.train_input[:25]:
+                end = (x != self.t_nul).nonzero().max().item() + 1
+                seq = [self.id2token[i.item()] for i in x[:end]]
+                s = " ".join(seq)
+                logger(f"example_seq {s}")
+
+        self.nb_codes = max(self.train_input.max(), self.test_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=desc
+        ):
+            last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
+            batch = batch[:, :last].to(self.device)
+            yield batch
+
+    def vocabulary_size(self):
+        return self.nb_codes
+
+    def produce_results(
+        self, n_epoch, model, result_dir, logger, deterministic_synthesis
+    ):
+        # --------------------------------------------------------------------
+        def compute_nb_errors_prog(input, nb_to_log=0):
+            result = input.clone()
+            s = (result == self.t_prog).long()
+            ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
+            result = (1 - ar_mask) * result + ar_mask * self.t_nul
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                device=self.device,
+            )
+
+            sum_nb_total, sum_nb_errors = 0, 0
+            for one_input, one_result in zip(input, result):
+                seq = [self.id2token[i.item()] for i in one_result]
+                nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
+                sum_nb_total += 1
+                sum_nb_errors += 0 if nb_errors == 0 else 1
+                if nb_to_log > 0:
+                    gt_seq = [self.id2token[i.item()] for i in one_input]
+                    _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
+                    gt_prog = " ".join([str(x) for x in gt_prog])
+                    prog = " ".join([str(x) for x in prog])
+                    comment = "*" if nb_errors == 0 else "-"
+                    logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
+                    for start_stack, target_stack, result_stack, correct in stacks:
+                        comment = "*" if correct else "-"
+                        start_stack = " ".join([str(x) for x in start_stack])
+                        target_stack = " ".join([str(x) for x in target_stack])
+                        result_stack = " ".join([str(x) for x in result_stack])
+                        logger(
+                            f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
+                        )
+                    nb_to_log -= 1
+
+            return sum_nb_total, sum_nb_errors
+
+        # --------------------------------------------------------------------
+        def compute_nb_errors_output(input, nb_to_log=0):
+            result = input.clone()
+            k = torch.arange(result.size(1), device=result.device)[None, :]
+            last_output_idx = (
+                ((result == self.t_output) * k).max(dim=1, keepdim=True).values
+            )
+            first_prog_idx = (
+                ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
+            )
+            ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
+            result = (1 - ar_mask) * result + ar_mask * self.t_nul
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                device=self.device,
+            )
+
+            sum_nb_total, sum_nb_errors = 0, 0
+            for one_input, one_result, i, j in zip(
+                input, result, last_output_idx, first_prog_idx
+            ):
+                seq = [self.id2token[i.item()] for i in one_result]
+                sum_nb_total += 1
+                correct = (one_input - one_result).abs().max() == 0
+                sum_nb_errors += 0 if correct else 1
+                if nb_to_log > 0:
+                    result_stack = [
+                        self.id2token[i.item()] for i in one_result[i : j + 1]
+                    ]
+                    target_stack = [
+                        self.id2token[i.item()] for i in one_input[i : j + 1]
+                    ]
+                    comment = "*" if correct else "-"
+                    result_stack = " ".join([str(x) for x in result_stack])
+                    target_stack = " ".join([str(x) for x in target_stack])
+                    logger(
+                        f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
+                    )
+                    nb_to_log -= 1
+
+            return sum_nb_total, sum_nb_errors
+
+        # --------------------------------------------------------------------
+
+        if not self.no_prog:
+            test_nb_total, test_nb_errors = compute_nb_errors_prog(
+                self.test_input[:1000].to(self.device), nb_to_log=10
+            )
+
+            logger(
+                f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
+            )
+
+            logger(f"main_test_accuracy {n_epoch} {1-test_nb_errors/test_nb_total}")
+
+        test_nb_total, test_nb_errors = compute_nb_errors_output(
+            self.test_input[:1000].to(self.device), nb_to_log=10
+        )
+
+        logger(
+            f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
+        )
+
+        if save_attention_image is None:
+            logger("no save_attention_image (is pycairo installed?)")
+        else:
+            ns = torch.randint(self.test_input.size(0), (1,)).item()
+            input = self.test_input[ns : ns + 1].clone()
+            last = (input != self.t_nul).max(0).values.nonzero().max() + 3
+            input = input[:, :last].to(self.device)
+
+            with torch.autograd.no_grad():
+                t = model.training
+                model.eval()
+                model.record_attention(True)
+                model(BracketedSequence(input))
+                model.train(t)
+                ram = model.retrieve_attention()
+                model.record_attention(False)
+
+            tokens_output = [self.id2token[i.item()] for i in 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"rpl_attention_{n_epoch}_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}")
+
+
 ######################################################################
 
 
@@ -1172,6 +1373,8 @@ class Expr(Task):
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
         )
 
+        logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
         nb_total = test_nb_delta.sum() + test_nb_missed
         for d in range(test_nb_delta.size(0)):
             logger(