From: François Fleuret Date: Wed, 5 Jul 2023 06:45:04 +0000 (+0200) Subject: Update. X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=commitdiff_plain;h=02c4828834319a5b7818bafb8821fce66b3a1bb1;p=picoclvr.git Update. --- diff --git a/expr.py b/expr.py index 723022c..ca33daf 100755 --- a/expr.py +++ b/expr.py @@ -1,6 +1,6 @@ #!/usr/bin/env python -import math +import math, re import torch, torchvision @@ -53,6 +53,15 @@ def generate_program(nb_variables, length): return s, variables +def extract_results(seq): + f = lambda a: (a[0], -1 if a[1] == "" else int(a[1])) + results = [ + dict([f(tuple(x.split(":"))) for x in re.findall("[A-Z]:[0-9]*", s)]) + for s in seq + ] + return results + + def generate_sequences(nb, nb_variables=5, length=20, randomize_length=False): sequences = [] for n in range(nb): @@ -78,8 +87,10 @@ if __name__ == "__main__": import time start_time = time.perf_counter() - sequences = generate_sequences(1000, randomize_length=True) + sequences = generate_sequences(1000) end_time = time.perf_counter() for s in sequences[:10]: print(s) print(f"{len(sequences) / (end_time - start_time):.02f} samples per second") + + print(extract_results(sequences[:10])) diff --git a/main.py b/main.py index 35bf02c..e1f619c 100755 --- a/main.py +++ b/main.py @@ -1030,8 +1030,9 @@ class TaskExpr(Task): train_sequences = expr.generate_sequences( nb_train_samples, nb_variables=nb_variables, - length=2 * sequence_length, - randomize_length=True, + length=sequence_length, + # length=2 * sequence_length, + # randomize_length=True, ) test_sequences = expr.generate_sequences( nb_test_samples, @@ -1115,6 +1116,24 @@ class TaskExpr(Task): nb_total = input.size(0) nb_correct = (input == result).long().min(1).values.sum() + values_input = expr.extract_results([self.seq2str(s) for s in input]) + max_input = max([max(x.values()) for x in values_input]) + values_result = expr.extract_results([self.seq2str(s) for s in result]) + max_result = max( + [-1 if len(x) == 0 else max(x.values()) for x in values_result] + ) + + nb_missing, nb_predicted = torch.zeros(max_input + 1), torch.zeros( + max_input + 1, max_result + 1 + ) + for i, r in zip(values_input, values_result): + for n, vi in i.items(): + vr = r.get(n) + if vr is None or vr < 0: + nb_missing[vi] += 1 + else: + nb_predicted[vi, vr] += 1 + return nb_total, nb_correct test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])