X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=a97ec2e4c8298dcd764480e114fee099979a9d12;hb=4540ede418ea744e50e0ff0b3a90785015da962b;hp=75cd35ed6c2e2c280fa4ad2ea07e7adb86439bb8;hpb=0c47d4d8ef8c4938f4765af816349cf30da14cb1;p=picoclvr.git diff --git a/tasks.py b/tasks.py index 75cd35e..a97ec2e 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 + import math, os, tqdm import torch, torchvision @@ -108,9 +113,7 @@ class ProblemLevel1(Problem): source = torch.rand(nb, 10).sort(dim=1).indices[:, : 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 @@ -1042,7 +1045,7 @@ class RPL(Task): ) ], 0, - ).to(self.device) + ) def seq2str(self, seq): return " ".join([self.id2token[i] for i in seq]) @@ -1052,6 +1055,11 @@ class RPL(Task): nb_train_samples, nb_test_samples, batch_size, + nb_starting_values=3, + max_input=9, + prog_len=6, + nb_runs=5, + logger=None, device=torch.device("cpu"), ): super().__init__() @@ -1060,11 +1068,23 @@ class RPL(Task): self.device = device train_sequences = [ - rpl.generate() + rpl.generate( + nb_starting_values=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() for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data") + rpl.generate( + nb_starting_values=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( @@ -1074,15 +1094,25 @@ class RPL(Task): symbols = list(filter(lambda x: type(x) is str, symbols)) symbols.sort() symbols += [str(n) for n in range(val_max + 1)] - print(f"{val_max=}") 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.t_prog = self.token2id[""], self.token2id[""] + self.t_nul = self.token2id[""] + self.t_prog = self.token2id[""] + self.t_input = self.token2id[""] + self.t_output = self.token2id[""] self.train_input = self.tensorize(train_sequences) self.test_input = self.tensorize(test_sequences) + 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): @@ -1096,7 +1126,7 @@ class RPL(Task): input.split(self.batch_size), dynamic_ncols=True, desc=desc ): last = (batch != self.t_nul).max(0).values.nonzero().max() + 3 - batch = batch[:, :last] + batch = batch[:, :last].to(self.device) yield batch def vocabulary_size(self): @@ -1105,7 +1135,8 @@ class RPL(Task): def produce_results( self, n_epoch, model, result_dir, logger, deterministic_synthesis ): - def compute_nb_errors(input, nb_to_log=0): + # -------------------------------------------------------------------- + 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) @@ -1131,21 +1162,66 @@ class RPL(Task): _, _, 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]) - logger(f"GROUND-TRUTH PROG [{gt_prog}] PREDICTED PROG [{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 = " CORRECT" if correct else "" + 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" [{start_stack}] -> [{result_stack}] TARGET [{target_stack}]{comment}" + f" {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]" ) nb_to_log -= 1 return sum_nb_total, sum_nb_errors - test_nb_total, test_nb_errors = compute_nb_errors( - self.test_input[:1000], nb_to_log=10 + # -------------------------------------------------------------------- + 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, keep_dim=True) + first_prog_idx = ((result == self.t_prog) * k).min(dim=1, keep_dim=True) + ar_mask = (k > last_output_idx).long() * (k < first_prog_idx) + 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 x, y in zip(input, result): + seq = [self.id2token[i.item()] for i in y] + sum_nb_total += 1 + sum_nb_errors += 0 if (x - y).abs().max() == 0 else 1 + if nb_to_log > 0: + gt_seq = [self.id2token[i.item()] for i in x] + _, _, 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 + + # -------------------------------------------------------------------- + + test_nb_total, test_nb_errors = compute_nb_errors_prog( + self.test_input[:1000].to(self.device), nb_to_log=10 ) logger(