def fit_model(
_run,
ray_server: str,
init_kwargs: Dict[str, Any],
activation_dir: str,
output_root: str,
num_components: int,
num_observations: int,
perplexity: int,
data_type,
):
try:
ray.init(address=ray_server, **init_kwargs)
# Find activation paths for each environment & victim-path tuple
stem_pattern = re.compile(r"(.*)_opponent_.*\.npz")
opponent_pattern = re.compile(r".*_opponent_([^\s]+)_[^\s]+\.npz")
activation_paths = {}
for fname in os.listdir(activation_dir):
stem_match = stem_pattern.match(fname)
if stem_match is None:
logger.debug(f"Skipping {fname}")
continue
stem = stem_match.groups()[0]
opponent_match = opponent_pattern.match(fname)
opponent_type = opponent_match.groups()[0]
path = osp.join(activation_dir, fname)
activation_paths.setdefault(stem, {})[opponent_type] = path
# Create temporary output directory (if needed)
tmp_dir = None
if output_root is None:
tmp_dir = tempfile.TemporaryDirectory()
output_root = tmp_dir.name
# Fit t-SNE and save model weights
results = []
for stem, paths in activation_paths.items():
output_dir = osp.join(output_root, stem)
os.makedirs(output_dir)
future = fit_tsne_helper.remote(
paths, output_dir, num_components, num_observations, perplexity, data_type
)
results.append(future)
ray.get(results) # block until all jobs have finished
utils.add_artifacts(_run, output_root, ingredient=fit_model_ex)
finally:
# Clean up temporary directory (if needed)
if tmp_dir is not None:
tmp_dir.cleanup()
ray.shutdown()
def visualize(_run, model_glob, output_root):
# Output directory
tmp_dir = None
if output_root is None:
tmp_dir = tempfile.TemporaryDirectory()
output_root = tmp_dir.name
for model_dir in glob.glob(model_glob):
model_name = os.path.basename(model_dir)
output_dir = osp.join(output_root, model_name)
os.makedirs(output_dir)
_visualize_helper(model_dir, output_dir)
utils.add_artifacts(_run, output_root, ingredient=visualize_ex)
if tmp_dir is not None:
tmp_dir.cleanup()
def generate_activations(
_run, out_dir, score_configs, score_update, adversary_path, ray_upload_dir
):
"""Uses multi.score to generate activations, then extracts them into a convenient
directory structure."""
logger.info("Generating activations")
activation_dirs = run_external(
score_configs,
post_named_configs=["save_activations"],
config_updates=score_update,
adversary_path=adversary_path,
)
os.makedirs(out_dir)
extract_data(_activations_path_generator, out_dir, activation_dirs, ray_upload_dir)
logger.info("Activations saved")
utils.add_artifacts(_run, out_dir)
def fit_model(
_run,
ray_server: str,
init_kwargs: Dict[str, Any],
activation_glob: str,
output_root: str,
max_timesteps: int,
data_type,
model_class,
model_kwargs,
train_opponent,
train_percentage,
):
"""Fits density models for each environment and victim type in activation_dir,
saving resulting models to output_root. Works by repeatedly calling `density_fitter`,
running in parallel via Ray."""
try:
ray.init(address=ray_server, **init_kwargs)
# Find activation paths for each environment & victim-path tuple
stem_pattern = re.compile(r"(.*)_opponent_.*\.npz")
opponent_pattern = re.compile(r".*_opponent_([^\s]+)+\.npz")
# activation_paths is indexed by [env_victim][opponent_type] where env_victim is
# e.g. 'SumoHumans-v0_victim_zoo_1' and opponent_type is e.g. 'ppo2_1'.
activation_paths = {}
for activation_path in glob.glob(activation_glob):
activation_dir = os.path.basename(activation_path)
stem_match = stem_pattern.match(activation_dir)
if stem_match is None:
logger.debug(f"Skipping {activation_path}")
continue
stem = stem_match.groups()[0]
opponent_match = opponent_pattern.match(activation_dir)
opponent_type = opponent_match.groups()[0]
activation_paths.setdefault(stem,
{})[opponent_type] = activation_path
# Create temporary output directory (if needed)
tmp_dir = None
if output_root is None:
tmp_dir = tempfile.TemporaryDirectory()
output_root = tmp_dir.name
else:
exp_name = gen_exp_name(model_class, model_kwargs)
output_root = os.path.join(output_root, exp_name)
# Fit density model and save weights
results = []
for stem, paths in activation_paths.items():
output_dir = osp.join(output_root, stem)
os.makedirs(output_dir)
future = density_fitter.remote(
paths,
output_dir,
model_class,
utils.sacred_copy(model_kwargs),
max_timesteps,
data_type,
train_opponent,
train_percentage,
)
results.append(future)
ray.get(results) # block until all jobs have finished
utils.add_artifacts(_run, output_root, ingredient=fit_model_ex)
finally:
# Clean up temporary directory (if needed)
if tmp_dir is not None:
tmp_dir.cleanup()
ray.shutdown()