def test_complexity(self, cfg_file, cx_expected):
"""Test complexity of a single model with the specified config."""
cfg_init = cfg.clone()
cfg.merge_from_file(cfg_file)
cx = net.complexity(builders.get_model())
cfg.merge_from_other_cfg(cfg_init)
self.assertEqual(cx_expected, cx)
def dump_complexity():
"""Measure the complexity of every model in the configs/ directory."""
complexity = {"date-created": str(datetime.datetime.now())}
cfg_files = [os.path.join(r, f) for r, _, fs in os.walk("configs/") for f in fs]
cfg_files = sorted(f for f in cfg_files if ".yaml" in f)
for cfg_file in cfg_files:
cfg_init = cfg.clone()
cfg.merge_from_file(cfg_file)
complexity[cfg_file] = net.complexity(builders.get_model())
cfg.merge_from_other_cfg(cfg_init)
with open(_COMPLEXITY_FILE, "w") as file:
json.dump(complexity, file, sort_keys=True, indent=4)
def sample_cfgs(seed):
"""Samples chunk configs and return those that are unique and valid."""
# Fix RNG seed (every call to this function should use a unique seed)
np.random.seed(seed)
setup_cfg = sweep_cfg.SETUP
cfgs = {}
for _ in range(setup_cfg.CHUNK_SIZE):
# Sample parameters [key, val, ...] list based on the samplers
params = samplers.sample_parameters(setup_cfg.SAMPLERS)
# Check if config is unique, if not continue
key = zip(params[0::2], params[1::2])
key = " ".join(["{} {}".format(k, v) for k, v in key])
if key in cfgs:
continue
# Generate config from parameters
reset_cfg()
cfg.merge_from_other_cfg(setup_cfg.BASE_CFG)
cfg.merge_from_list(params)
# Check if config is valid, if not continue
is_valid = samplers.check_regnet_constraints(setup_cfg.CONSTRAINTS)
if not is_valid:
continue
# Special logic for dealing w model scaling (side effect is to standardize cfg)
if cfg.MODEL.TYPE in ["anynet", "effnet", "regnet"]:
scaler.scale_model()
# Check if config is valid, if not continue
is_valid = samplers.check_complexity_constraints(setup_cfg.CONSTRAINTS)
if not is_valid:
continue
# Set config description to key
cfg.DESC = key
# Store copy of config if unique and valid
cfgs[key] = cfg.clone()
# Stop sampling if already reached quota
if len(cfgs) == setup_cfg.NUM_CONFIGS:
break
return cfgs
_C.SETUP = CfgNode() # Max number of unique job configs to generate _C.SETUP.NUM_CONFIGS = 0 # Max number of attempts for generating NUM_CONFIGS valid configs _C.SETUP.NUM_SAMPLES = 1000000 # Specifies the chunk size to use per process while sampling configs _C.SETUP.CHUNK_SIZE = 5000 # Random seed for generating job configs _C.SETUP.RNG_SEED = 0 # Base config for all jobs, any valid config option in core.config is valid here _C.SETUP.BASE_CFG = cfg.clone() # Samplers to use for generating job configs (see SAMPLERS defined toward end of file) # SETUP.SAMPLERS should consists of a dictionary of SAMPLERS # Each dict key should be a valid parameter in the BASE_CFG (e.g. "MODEL.DEPTH") # Each dict val should be a valid SAMPLER that defines how to sample (e.g. INT_SAMPLER) # See the example sweep configs for more usage information _C.SETUP.SAMPLERS = CfgNode(new_allowed=True) # Constraints on generated configs _C.SETUP.CONSTRAINTS = CfgNode() # Complexity constraints CX on models specified as a [LOW, HIGH] range, e.g. [0, 1.0e+6] # If LOW == HIGH == 0 for a given complexity constraint that constraint is not applied # For RegNets, if flops<F (B), setting params<3+5.5F and acts<6.5*sqrt(F) (M) works well _C.SETUP.CONSTRAINTS.CX = CfgNode()