def modify_config(self, config: FaceGroupingConfig, i):
if i < self.num_runs / 2:
config.dataset = self.bf
else:
config.dataset = self.bbt
return super().modify_config(config, i)
def modify_config(self, config: FaceGroupingConfig, i):
if i < 6:
config.dataset = BigBangTheory(episode_index_test=i)
else:
config.dataset = Buffy(episode_index_test=i - 6)
return super().modify_config(config, i)
def modify_config(self, config: FaceGroupingConfig, i):
from FGG.dataset.split_strategy import SplitEveryXFrames
x = 10
config.graph_builder_params["split_strategy"] = SplitEveryXFrames(x=x)
config.pool_before_clustering = True
config.dataset = Buffy(episode_index_val=None,
episode_index_train=None,
episode_index_test=i)
return (str(x), *super().modify_config(config, i))
def next_experiment(self):
"""
Generator for configs.
Should be sent the result WCP of the previous experiment.
:return:
"""
with open(self.file_name, "w") as f:
f.write(f"{','.join(self.header)}\n")
for i in range(self.num_runs):
config = FaceGroupingConfig()
entries = list(map(str, self.modify_config(config, i)))
config.model_name = self.create_model_name(config, i)
config.finalize()
result = yield config
with open(self.file_name, "a") as f:
f.write(f"{','.join(entries + [str(result), ])}\n")
def modify_config(self, config: FaceGroupingConfig, i):
out = super().modify_config(config, i)
assert config.dataset.episode_index_test is not None
config.dataset.episode_index_train = None
config.dataset.episode_index_val = None
# These are the paths for the pretrained weights.
if isinstance(config.dataset.episode_index_test, int):
if isinstance(config.dataset,
BigBangTheory) and self.checkpoint_file is Ellipsis:
config.model_load_file = f"weights/bbt010{config.dataset.episode_index_test + 1}_checkpoint.tar"
elif isinstance(config.dataset,
Buffy) and self.checkpoint_file is Ellipsis:
config.model_load_file = f"weights/bf050{config.dataset.episode_index_test + 1}_checkpoint.tar"
elif isinstance(config.dataset,
Accio) and self.checkpoint_file is Ellipsis:
config.model_load_file = "weights/accio_checkpoint.tar"
elif config.model_load_file is None:
config.model_load_file = self.checkpoint_file
print("Selected checkpoint", config.model_load_file)
assert config.model_load_file is None or Path(
config.model_load_file).is_file(), config.model_load_file
return out
def modify_config(self, config: FaceGroupingConfig, i):
config.dataset = self.accio
config.model_params["sparse_adjacency"] = True
return (str(self.num_clusters), *super().modify_config(config, i))
def modify_config(self, config: FaceGroupingConfig, i):
config.dataset = self.bf
return super().create_model_name(config, i)
def modify_config(self, config: FaceGroupingConfig, i):
config.dataset = self.dataset
return super().modify_config(config, i)
def modify_config(self, config: FaceGroupingConfig, i):
config.dataset = Buffy(episode_index_test=i)
return super().modify_config(config, i)
def modify_config(self, config: FaceGroupingConfig, i):
config.dataset = Accio()
config.model_params["sparse_adjacency"] = True
return super().modify_config(config, i)