def build_search_space(function):
# abuse configscope to parse search space definitions
scope = ConfigScope(function)
space_dict = dict(scope())
# parse generic dict to a search space
space = SearchSpace(space_dict)
return space
def named_config(self, func):
"""
Decorator to turn a function into a named configuration.
See :ref:`named_configurations`.
"""
config_scope = ConfigScope(func)
self._add_named_config(func.__name__, config_scope)
return config_scope
def dataset_default_config(config):
if isinstance(config, list):
configs = []
for c in config:
configs.append(dataset_default_config(c))
return configs
scope = ConfigScope(dataset_cfg_fn)
config = scope(preset=config)
config['transform'] = transform_default_config(
config.get('transform', {}))
return config
def config(self, function):
"""
Decorator to add a function to the configuration of the Experiment.
The decorated function is turned into a
:class:`~sacred.config_scope.ConfigScope` and added to the
Ingredient/Experiment.
When the experiment is run, this function will also be executed and
all json-serializable local variables inside it will end up as entries
in the configuration of the experiment.
"""
self.configurations.append(ConfigScope(function))
return self.configurations[-1]
def test_searcher_bm25_grid(tmpdir_as_cache, tmpdir, dummy_index):
searcher_config = ConfigScope(BM25Grid.config)()
searcher_config["_name"] = BM25Grid.name
searcher = BM25Grid(searcher_config)
searcher.modules["index"] = dummy_index
bs = np.around(np.arange(0.1, 1 + 0.1, 0.1), 1)
k1s = np.around(np.arange(0.1, 1 + 0.1, 0.1), 1)
topics_fn = DummyBenchmark.topic_file
output_fn = searcher.query_from_file(
topics_fn, os.path.join(searcher.get_cache_path(),
DummyBenchmark.name))
assert output_fn == os.path.join(searcher.get_cache_path(),
DummyBenchmark.name)
for k1 in k1s:
for b in bs:
assert os.path.exists(
os.path.join(output_fn, "searcher_k1={0},b={1}".format(k1, b)))
assert os.path.exists(os.path.join(output_fn, "done"))
def test_searcher_bm25(tmpdir_as_cache, tmpdir, dummy_index):
searcher_config = ConfigScope(BM25.config)()
searcher_config["_name"] = BM25.name
searcher = BM25(searcher_config)
searcher.modules["index"] = dummy_index
topics_fn = DummyBenchmark.topic_file
output_fn = searcher.query_from_file(
topics_fn, os.path.join(searcher.get_cache_path(),
DummyBenchmark.name))
assert output_fn == os.path.join(searcher.get_cache_path(),
DummyBenchmark.name)
with open(os.path.join(output_fn, "searcher"), "r") as fp:
file_contents = fp.readlines()
assert file_contents == [
"301 Q0 LA010189-0001 1 0.139500 Anserini\n",
"301 Q0 LA010189-0002 2 0.097000 Anserini\n"
]
"""doc"""
pass
def _config_scope_with_multiline_doc():
"""Multiline
docstring!
"""
pass
@pytest.mark.parametrize(
'indent, path, named_config, expected',
[(0, 'a', None, 'a'), (1, 'b', None, ' b'),
(4, 'a.b.c', None, ' a.b.c'),
(0, 'c', ConfigScope(_config_scope_with_single_line_doc),
'c' + COLOR_DOC + ' # doc' + ENDC),
(0, 'd', ConfigScope(_config_scope_with_multiline_doc),
'd' + COLOR_DOC + '\n """Multiline\n docstring!\n """' + ENDC)])
def test_format_named_config(indent, path, named_config, expected):
assert _format_named_config(indent, path, named_config) == expected
def test_format_named_configs():
ingred = Ingredient('ingred')
ex = Experiment(name='experiment', ingredients=[ingred])
@ingred.named_config
def named_config1():
pass
"""doc"""
pass
def _config_scope_with_multiline_doc():
"""Multiline
docstring!
"""
pass
@pytest.mark.parametrize('indent, path, named_config, expected', [
(0, 'a', None, 'a'),
(1, 'b', None, ' b'),
(4, 'a.b.c', None, ' a.b.c'),
(0, 'c', ConfigScope(_config_scope_with_single_line_doc), 'c' + COLOR_DOC
+ ' # doc' + ENDC),
(0, 'd', ConfigScope(_config_scope_with_multiline_doc),
'd' + COLOR_DOC + '\n """Multiline\n docstring!\n """' + ENDC)
])
def test_format_named_config(indent, path, named_config, expected):
assert _format_named_config(indent, path, named_config) == expected
def test_format_named_configs():
ingred = Ingredient('ingred')
ex = Experiment(name='experiment', ingredients=[ingred])
@ingred.named_config
def named_config1():
pass
def build_default_config(name, config):
fn = build_cfg_fn(name)
scope = ConfigScope(fn)
return scope(preset=config)
def build_config(config):
def build_default_config(name, config):
fn = build_cfg_fn(name)
scope = ConfigScope(fn)
return scope(preset=config)
def model_default_config(config):
name = config['name']
return build_default_config(name, config)
def transform_default_config(config):
return build_default_config('transform', config)
def dataset_default_config(config):
if isinstance(config, list):
configs = []
for c in config:
configs.append(dataset_default_config(c))
return configs
scope = ConfigScope(dataset_cfg_fn)
config = scope(preset=config)
config['transform'] = transform_default_config(
config.get('transform', {}))
return config
def single_sampler_default_config(config):
name = config['type']
config['dataset'] = dataset_default_config(config['dataset'])
return build_default_config(name, config)
def sampler_default_config(config):
if "samplers" in config:
#multi sampler
samplers = config['samplers']
for name, c in samplers.items():
samplers[name] = single_sampler_default_config(c)
return config
else:
return single_sampler_default_config(config)
def loss_default_config(config):
if isinstance(config, list):
configs = []
for c in config:
configs.append(loss_default_config(c))
return configs
name = config['name']
cfg = build_default_config(name, config)
return cfg
def scheduler_default_config(config):
if 'preset' in config:
name = "{}_scheduler".format(config['preset'])
else:
name = "{}_scheduler".format(config['name'])
return build_default_config(name, config)
def optimizer_default_config(config):
name = config['name']
return build_default_config(name, config)
def training_default_config(config):
config = build_default_config('train', config)
for key, values in config.items():
if key == 'model':
config['model'] = model_default_config(config['model'])
elif key == 'dataloader':
config['dataloader']['sampler'] = sampler_default_config(
config['dataloader']['sampler'])
elif key == 'losses':
config['losses'] = loss_default_config(config['losses'])
elif key == 'scheduler':
config['scheduler'] = scheduler_default_config(
config['scheduler'])
elif key == 'optimizer':
config['optimizer'] = optimizer_default_config(
config['optimizer'])
elif key == 'checkpoint_frequency':
pass
elif key == 'restore_checkpoint':
pass
elif key == 'epochs':
pass
elif key == 'num_workers':
pass
else:
raise ValueError(key)
return config
# TODO make central or get rid
reid_datasets = ["market-1501", "duke"]
reid_attribute_datasets = ["market-1501-attribute", "duke-attribute"]
pose_datasets = ["mpii"]
def evaluation_dataset_default_config(config):
if isinstance(config, list):
configs = []
for c in config:
configs.append(evaluation_dataset_default_config(c))
return configs
name = config['name'].lower()
config = build_default_config('evaluation_dataset', config)
if name in reid_datasets:
return build_default_config('reid_evaluation', config)
elif name in reid_attribute_datasets:
return build_default_config('reid_attribute_evaluation', config)
elif name in pose_datasets:
return config
else:
raise ValueError(
"Unknown evaluation dataset in config builder: {}.".format(
name))
def evaluation_default_config(config):
if 'experiment' in config:
# TODO restore the config
return config
#config['sampler']['datasets'] = evaluation_dataset_default_config(config['sampler']['datasets'])
config = build_default_config('evaluation', config)
return config
scope = ConfigScope(general_cfg_fn)
config = scope(preset=config)
for key, value in config.items():
if key == 'training':
config['training'] = training_default_config(config['training'])
elif key == 'evaluation':
config['evaluation'] = evaluation_default_config(
config['evaluation'])
elif key == 'validation':
pass
elif key == 'device_id':
pass
elif key == 'num_workers':
pass
elif key == 'restore_checkpoint':
pass
elif key == 'experiment':
pass
else:
pass
return config