X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=da39a830e3b4d899e2f3521f73444fa9cdd8c36b;hb=0c6d29f73e35adbbaab1263de439f73efa98d99e;hp=0a4dd6fa2f880e93aecbbd494621fae26b7dcdbb;hpb=a291e213a152364b74e833200191c08a36451a90;p=picoclvr.git diff --git a/tasks.py b/tasks.py index 0a4dd6f..da39a83 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 @@ -29,7 +34,7 @@ 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(): @@ -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 @@ -1067,6 +1070,7 @@ class RPL(Task): 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, @@ -1077,6 +1081,7 @@ class RPL(Task): 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, @@ -1091,16 +1096,19 @@ 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]] @@ -1130,7 +1138,7 @@ class RPL(Task): 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) @@ -1171,13 +1179,62 @@ class RPL(Task): 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 x, y, i, j in zip(input, result, last_output_idx, first_prog_idx): + seq = [self.id2token[i.item()] for i in y] + sum_nb_total += 1 + correct = (x - y).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 y[i : j + 1]] + target_stack = [self.id2token[i.item()] for i in x[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 + + # -------------------------------------------------------------------- + + 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}%" + ) - test_nb_total, test_nb_errors = compute_nb_errors( + test_nb_total, test_nb_errors = compute_nb_errors_output( self.test_input[:1000].to(self.device), nb_to_log=10 ) logger( - f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%" + 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}%" )