|
1 |
| -import json |
2 |
| -import numpy as np |
3 | 1 | import argparse
|
4 | 2 | from os import path as osp
|
5 | 3 | from tabulate import tabulate
|
6 |
| - |
7 |
| - |
8 |
| -def load_values(dir_logs, metrics, nb_epochs=-1, best=None): |
9 |
| - json_files = {} |
10 |
| - values = {} |
11 |
| - |
12 |
| - # load argsup of best |
13 |
| - if best: |
14 |
| - if best['json'] not in json_files: |
15 |
| - with open(osp.join(dir_logs, f'{best["json"]}.json')) as f: |
16 |
| - json_files[best['json']] = json.load(f) |
17 |
| - |
18 |
| - jfile = json_files[best['json']] |
19 |
| - vals = jfile[best['name']] |
20 |
| - end = len(vals) if nb_epochs == -1 else nb_epochs |
21 |
| - argsup = np.__dict__[f'arg{best["order"]}'](vals[:end]) |
22 |
| - |
23 |
| - # load logs |
24 |
| - for _key, metric in metrics.items(): |
25 |
| - # open json_files |
26 |
| - if metric['json'] not in json_files: |
27 |
| - with open(osp.join(dir_logs, f'{metric["json"]}.json')) as f: |
28 |
| - json_files[metric['json']] = json.load(f) |
29 |
| - |
30 |
| - jfile = json_files[metric['json']] |
31 |
| - |
32 |
| - if 'train' in metric['name']: |
33 |
| - epoch_key = 'train_epoch.epoch' |
34 |
| - else: |
35 |
| - epoch_key = 'eval_epoch.epoch' |
36 |
| - |
37 |
| - if epoch_key in jfile: |
38 |
| - epochs = jfile[epoch_key] |
39 |
| - else: |
40 |
| - epochs = jfile['epoch'] |
41 |
| - |
42 |
| - vals = jfile[metric['name']] |
43 |
| - if not best: |
44 |
| - end = len(vals) if nb_epochs == -1 else nb_epochs |
45 |
| - argsup = np.__dict__[f'arg{metric["order"]}'](vals[:end]) |
46 |
| - |
47 |
| - try: |
48 |
| - values[metric['name']] = epochs[argsup], vals[argsup] |
49 |
| - except IndexError: |
50 |
| - values[metric['name']] = epochs[argsup - 1], vals[argsup - 1] |
51 |
| - return values |
52 |
| - |
53 |
| - |
54 |
| -def main(args): |
55 |
| - dir_logs = {} |
56 |
| - for raw in args.dir_logs: |
57 |
| - tmp = raw.split(':') |
58 |
| - if len(tmp) == 2: |
59 |
| - key, path = tmp |
60 |
| - elif len(tmp) == 1: |
61 |
| - path = tmp[0] |
62 |
| - key = osp.basename(osp.normpath(path)) |
63 |
| - else: |
64 |
| - raise ValueError(raw) |
65 |
| - dir_logs[key] = path |
66 |
| - |
67 |
| - metrics = {} |
68 |
| - for json_obj, name, order in args.metrics: |
69 |
| - metrics[f'{json_obj}_{name}'] = { |
70 |
| - 'json': json_obj, |
71 |
| - 'name': name, |
72 |
| - 'order': order |
73 |
| - } |
74 |
| - |
75 |
| - if args.best: |
76 |
| - json_obj, name, order = args.best |
77 |
| - best = { |
78 |
| - 'json': json_obj, |
79 |
| - 'name': name, |
80 |
| - 'order': order |
81 |
| - } |
82 |
| - else: |
83 |
| - best = None |
84 |
| - |
85 |
| - logs = {} |
86 |
| - for name, dir_log in dir_logs.items(): |
87 |
| - logs[name] = load_values(dir_log, metrics, |
88 |
| - nb_epochs=args.nb_epochs, |
89 |
| - best=best) |
90 |
| - |
91 |
| - for _key, metric in metrics.items(): |
92 |
| - names = [] |
93 |
| - values = [] |
94 |
| - epochs = [] |
95 |
| - for name, vals in logs.items(): |
96 |
| - if metric['name'] in vals: |
97 |
| - names.append(name) |
98 |
| - epoch, value = vals[metric['name']] |
99 |
| - epochs.append(epoch) |
100 |
| - values.append(value) |
101 |
| - if values: |
102 |
| - values_names = sorted(zip(values, names, epochs), reverse=metric['order'] == 'max') |
103 |
| - values_names = [[i + 1, name, value, epoch] for i, (value, name, epoch) in enumerate(values_names)] |
104 |
| - print('\n\n## {}\n'.format(metric['name'])) |
105 |
| - print(tabulate(values_names, headers=['Place', 'Method', 'Score', 'Epoch'])) |
| 4 | +import sqlite3 |
| 5 | +from contextlib import closing |
| 6 | + |
| 7 | +# def get_internal_table_name(table_name): |
| 8 | +# return f'_{table_name}' |
| 9 | + |
| 10 | +# def run_query(conn, query, parameters=None, cursor=None): |
| 11 | +# return execute(conn, query, parameters, commit=False, cursor=cursor) |
| 12 | + |
| 13 | +# def list_columns(conn, table_name): |
| 14 | +# table_name = get_internal_table_name(table_name) |
| 15 | +# query = "SELECT name FROM PRAGMA_TABLE_INFO(?)" |
| 16 | +# with closing(conn.cursor()) as cursor: |
| 17 | +# qry_cur = run_query(conn, query, (table_name,), cursor=cursor) |
| 18 | +# columns = (res[0] for res in qry_cur) |
| 19 | +# # remove __id and __timestamp columns |
| 20 | +# columns = [c for c in columns if not c.startswith('__')] |
| 21 | +# return columns |
| 22 | + |
| 23 | +# def select(conn, group, columns=None, where=None): |
| 24 | +# table_name = get_internal_table_name(group) |
| 25 | +# table_columns = list_columns(conn, group) |
| 26 | +# if columns is None: |
| 27 | +# column_string = '*' |
| 28 | +# else: |
| 29 | +# for c in columns: |
| 30 | +# if c not in table_columns: |
| 31 | +# Logger()(f'Unknown column "{c}"', log_level=Logger.ERROR) |
| 32 | +# column_string = ', '.join([f'"{c}"' for c in columns]) |
| 33 | +# statement = f'SELECT {column_string} FROM {table_name}' |
| 34 | +# with closing(conn.cursor()) as cursor: |
| 35 | +# return execute(conn, statement, cursor=cursor, commit=False).fetchall() |
| 36 | + |
| 37 | + |
| 38 | +def execute(conn, statement, parameters=None, commit=True, cursor=None): |
| 39 | + assert parameters is None or isinstance(parameters, tuple) |
| 40 | + parameters = parameters or () |
| 41 | + return_value = cursor.execute(statement, parameters) |
| 42 | + if commit: |
| 43 | + conn.commit() |
| 44 | + return return_value |
| 45 | + |
| 46 | + |
| 47 | +def load_table(list_dir, metric, nb_epochs=None, best=None): |
| 48 | + table = [] |
| 49 | + for dir_logs in list_dir: |
| 50 | + # if metric['fname'] == best['fname']: |
| 51 | + # path_sql = osp.join(dir_logs, f'{metric["fname"]}.sqlite') |
| 52 | + # conn = sqlite3.connect(path_sql, check_same_thread=False, isolation_level='IMMEDIATE') |
| 53 | + # statement = f'SELECT m.{metric["column"]}, m.epoch FROM _{metric["group"]} AS m, _{best["group"]} AS b' |
| 54 | + # if nb_epochs: |
| 55 | + # statement += f' WHERE m.epoch < {nb_epochs}' |
| 56 | + # if best['order'] == 'max': |
| 57 | + # order = 'DESC' |
| 58 | + # elif best['order'] == 'min': |
| 59 | + # order = 'ASC' |
| 60 | + # statement += f' ORDER BY b.{best["column"]} {order} LIMIT 1' |
| 61 | + # with closing(conn.cursor()) as cursor: |
| 62 | + # score, epoch = execute(conn, statement, cursor=cursor).fetchone() |
| 63 | + # else: |
| 64 | + path_sql = osp.join(dir_logs, f'{best["fname"]}.sqlite') |
| 65 | + conn = sqlite3.connect(path_sql, check_same_thread=False, isolation_level='IMMEDIATE') |
| 66 | + statement = f'SELECT {best["column"]}, epoch FROM _{best["group"]}' |
| 67 | + if nb_epochs: |
| 68 | + statement += f' WHERE epoch < {nb_epochs}' |
| 69 | + if best['order'] == 'max': |
| 70 | + order = 'DESC' |
| 71 | + elif best['order'] == 'min': |
| 72 | + order = 'ASC' |
| 73 | + statement += f' ORDER BY {best["column"]} {order} LIMIT 1' |
| 74 | + with closing(conn.cursor()) as cursor: |
| 75 | + best_score, best_epoch = execute(conn, statement, cursor=cursor).fetchone() |
| 76 | + |
| 77 | + path_sql = osp.join(dir_logs, f'{metric["fname"]}.sqlite') |
| 78 | + conn = sqlite3.connect(path_sql, check_same_thread=False, isolation_level='IMMEDIATE') |
| 79 | + statement = f'SELECT {metric["column"]}, epoch FROM _{metric["group"]}' |
| 80 | + statement += f' WHERE epoch == {best_epoch}' |
| 81 | + with closing(conn.cursor()) as cursor: |
| 82 | + score, epoch = execute(conn, statement, cursor=cursor).fetchone() |
| 83 | + |
| 84 | + table.append([dir_logs, score, epoch]) |
| 85 | + |
| 86 | + if best['order'] == 'max': |
| 87 | + reverse = True |
| 88 | + elif best['order'] == 'min': |
| 89 | + reverse = False |
| 90 | + table.sort(key=lambda x: x[1], reverse=reverse) |
| 91 | + |
| 92 | + for i, x in enumerate(table): |
| 93 | + x.insert(0, f'# {i+1}') |
| 94 | + return table |
| 95 | + |
| 96 | + |
| 97 | +def metric_str_to_dict(metric): |
| 98 | + split_ = metric.split('.') |
| 99 | + return { |
| 100 | + 'fname': split_[0], |
| 101 | + 'group': split_[1], |
| 102 | + 'column': split_[2], |
| 103 | + 'order': split_[3] |
| 104 | + } |
| 105 | + |
| 106 | + |
| 107 | +def display_metrics(list_dir, metrics, nb_epochs=None, best=None): |
| 108 | + best = metric_str_to_dict(best) |
| 109 | + for mstr in metrics: |
| 110 | + metric = metric_str_to_dict(mstr) |
| 111 | + table = load_table(list_dir, metric, nb_epochs=nb_epochs, best=best) |
| 112 | + print(f'\n\n## {mstr}\n') |
| 113 | + print(tabulate(table, headers=['Place', 'Method', 'Score', 'Epoch'])) |
106 | 114 |
|
107 | 115 |
|
108 | 116 | if __name__ == '__main__':
|
109 | 117 | parser = argparse.ArgumentParser(description='')
|
110 | 118 | parser.add_argument('-n', '--nb_epochs', default=-1, type=int)
|
111 | 119 | parser.add_argument('-d', '--dir_logs', default='', type=str, nargs='*')
|
112 |
| - parser.add_argument('-m', '--metrics', type=str, action='append', nargs=3, |
113 |
| - metavar=('json', 'name', 'order'), |
114 |
| - default=[['logs', 'eval_epoch.accuracy_top1', 'max'], |
115 |
| - ['logs', 'eval_epoch.accuracy_top5', 'max'], |
116 |
| - ['logs', 'eval_epoch.loss', 'min']]) |
117 |
| - parser.add_argument('-b', '--best', type=str, nargs=3, |
118 |
| - metavar=('json', 'name', 'order'), |
119 |
| - default=['logs', 'eval_epoch.accuracy_top1', 'max']) |
| 120 | + parser.add_argument('-m', '--metrics', type=str, nargs='*', |
| 121 | + default=['logs.eval_epoch.accuracy.max', |
| 122 | + 'logs.train_epoch.loss.min', |
| 123 | + 'logs.train_epoch.accuracy.max']) |
| 124 | + parser.add_argument('-b', '--best', type=str, |
| 125 | + default='logs.eval_epoch.accuracy.max') |
120 | 126 | args = parser.parse_args()
|
121 |
| - main(args) |
| 127 | + nb_epochs = None if args.nb_epochs == -1 else args.nb_epochs |
| 128 | + display_metrics(args.dir_logs, args.metrics, nb_epochs, args.best) |
0 commit comments