def make_config(experiment: sacred.Experiment) -> None:
"""Adds configs and named configs to `experiment`.
The standard config parameters it defines are:
- env_name (str): The environment name in the Gym registry of the rewards to compare.
- x_reward_cfgs (Iterable[common_config.RewardCfg]): tuples of reward_type and reward_path
for x-axis.
- y_reward_cfgs (Iterable[common_config.RewardCfg]): tuples of reward_type and reward_path
for y-axis.
- log_root (str): the root directory to log; subdirectory path automatically constructed.
- n_bootstrap (int): the number of bootstrap samples to take.
- alpha (float): percentile confidence interval
- aggregate_kinds (Iterable[str]): the type of aggregations to perform across seeds.
Not used in `plot_return_heatmap` which only supports its own kind of bootstrapping.
- heatmap_kwargs (dict): passed through to `analysis.compact_heatmaps`.
- styles (Iterable[str]): styles to apply from `evaluating_rewards.analysis.stylesheets`.
- save_kwargs (dict): passed through to `analysis.save_figs`.
"""
# pylint: disable=unused-variable,too-many-statements
@experiment.config
def default_config():
"""Default configuration values."""
data_root = serialize.get_output_dir() # where models are read from
log_root = serialize.get_output_dir() # where results are written to
n_bootstrap = 1000 # number of bootstrap samples
alpha = 95 # percentile confidence interval
aggregate_kinds = ("bootstrap", "studentt", "sample")
_ = locals()
del _
@experiment.config
def aggregate_fns(aggregate_kinds, n_bootstrap, alpha):
"""Make a mapping of aggregate functions of kinds `subset` with specified parameters.
Used in scripts.distances.{epic,npec}; currently ignored by erc since
it does not use multiple seeds and instead bootstraps over samples.
"""
aggregate_fns = {}
if "bootstrap" in aggregate_kinds:
aggregate_fns["bootstrap"] = functools.partial(
bootstrap_ci, n_bootstrap=n_bootstrap, alpha=alpha)
if "studentt" in aggregate_kinds:
aggregate_fns["studentt"] = functools.partial(studentt_ci,
alpha=alpha)
if "sample" in aggregate_kinds:
aggregate_fns["sample"] = sample_mean_sd
@experiment.config
def point_mass_as_default():
"""Default to PointMass as environment so scripts work out-of-the-box."""
locals().update(**common_config.COMMON_CONFIGS["point_mass"])
for name, cfg in common_config.COMMON_CONFIGS.items():
experiment.add_named_config(name, cfg)
from tracktor.motion.model_reid import MotionModelReID
from tracktor.motion.model_simple_reid import MotionModelSimpleReID
from tracktor.motion.model_simple_reid_v2 import MotionModelSimpleReIDV2
from tracktor.motion.model_v3 import MotionModelV3
from tracktor.motion.vis_simple_reid import VisSimpleReID
from tracktor.motion.refine_model import RefineModel
from tracktor.reid.resnet import resnet50
from tracktor.utils import interpolate, plot_sequence, get_mot_accum, evaluate_mot_accums
ex = Experiment()
ex.add_config('experiments/cfgs/tracktor.yaml')
# hacky workaround to load the corresponding configs and not having to hardcode paths here
ex.add_config(ex.configurations[0]._conf['tracktor']['reid_config'])
ex.add_named_config('oracle', 'experiments/cfgs/oracle_tracktor.yaml')
@ex.automain
def main(tracktor, reid, _config, _log, _run):
sacred.commands.print_config(_run)
# set all seeds
torch.manual_seed(tracktor['seed'])
torch.cuda.manual_seed(tracktor['seed'])
np.random.seed(tracktor['seed'])
torch.backends.cudnn.deterministic = True
output_dir = osp.join(get_output_dir(tracktor['module_name']), tracktor['name'])
sacred_config = osp.join(output_dir, 'sacred_config.yaml')
import torchvision
import yaml
from tqdm import tqdm
import sacred
from sacred import Experiment
from tracktor.frcnn_fpn import FRCNN_FPN
from tracktor.config import get_output_dir
from tracktor.datasets.factory import Datasets
from tracktor.oracle_tracker import OracleTracker
from tracktor.tracker import Tracker
from tracktor.reid.resnet import resnet50
from tracktor.utils import interpolate, plot_sequence, get_mot_accum, evaluate_mot_accums
ex = Experiment()
ex.add_config('experiments/cfgs/tracktor.yaml')
ex.add_named_config('cfg1', 'experiments/cfgs/hp_search/cfg1.yaml')
ex.add_named_config('cfg_caro_ws1',
'experiments/cfgs/hp_search/cfg_caro_ws1.yaml')
ex.add_named_config('cfg_caro_ws2',
'experiments/cfgs/hp_search/cfg_caro_ws2.yaml')
ex.add_named_config('cfg_caro_ws3',
'experiments/cfgs/hp_search/cfg_caro_ws3.yaml')
ex.add_named_config('cfg_caro_local',
'experiments/cfgs/hp_search/cfg_caro_local.yaml')
########### DEFAULT################
# hacky workaround to load the corresponding configs and not having to har
# dcode paths here
ex.add_config(ex.configurations[0]._conf['tracktor']['reid_config'])
ex.add_named_config('oracle', 'experiments/cfgs/oracle_tracktor.yaml')
from tracktor.frcnn_fpn import FRCNN_FPN
from tracktor.config import get_output_dir
from tracktor.datasets.factory import Datasets
from tracktor.oracle_tracker import OracleTracker
from tracktor.motion import Seq2Seq, CorrelationSeq2Seq, RelativeCorrelationModel
from tracktor.tracker import Tracker
from tracktor.reid.resnet import resnet50
from tracktor.utils import interpolate, plot_sequence, get_mot_accum, evaluate_mot_accums
mm.lap.default_solver = 'lap'
ex = Experiment()
ex.add_config('experiments/cfgs/tracktor.yaml')
# hacky workaround to load the corresponding configs and not having to hardcode paths here
ex.add_config(ex.configurations[0]._conf['tracktor']['reid_config'])
ex.add_named_config('oracle', 'experiments/cfgs/oracle_tracktor.yaml')
ex.add_named_config('correlation_model',
'experiments/cfgs/correlation_model.yaml')
ex.add_named_config('pos_model', 'experiments/cfgs/pos_model.yaml')
ex.add_named_config('cva_model', 'experiments/cfgs/cva_model.yaml')
@ex.automain
def main(tracktor, reid, _config, _log, _run):
sacred.commands.print_config(_run)
# set all seeds
torch.manual_seed(tracktor['seed'])
torch.cuda.manual_seed(tracktor['seed'])
np.random.seed(tracktor['seed'])
torch.backends.cudnn.deterministic = True
num_env_steps = 100e6
eval_dir = "./results/video/{id}"
loss_dir = "./results/loss/{id}"
save_dir = "./results/trained_models/{id}"
log_interval = 2000
save_interval = int(1e6)
eval_interval = int(1e6)
episodes_per_eval = 8
for conf in glob.glob("configs/*.yaml"):
name = f"{Path(conf).stem}"
ex.add_named_config(name, conf)
def _squash_info(info):
info = [i for i in info if i]
new_info = {}
keys = set([k for i in info for k in i.keys()])
keys.discard("TimeLimit.truncated")
for key in keys:
mean = np.mean([np.array(d[key]).sum() for d in info if key in d])
new_info[key] = mean
return new_info
@ex.capture
def evaluate(
agents,
help='bbox score threshold')
args = parser.parse_args()
return args
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.cuda.set_device(0)
print('Using device:', device)
ex = Experiment()
ex.add_config('tracking_wo_bnw/experiments/cfgs/tracktor.yaml')
# hacky workaround to load the corresponding configs and not having to hardcode paths here
ex.add_config(ex.configurations[0]._conf['tracktor']['reid_config'])
ex.add_named_config('oracle',
'tracking_wo_bnw/experiments/cfgs/oracle_tracktor.yaml')
tracktor = ex.configurations[0].__call__()['tracktor']
reid = ex.configurations[1].__call__()['reid']
# set all seeds
torch.manual_seed(tracktor['seed'])
torch.cuda.manual_seed(tracktor['seed'])
np.random.seed(tracktor['seed'])
torch.backends.cudnn.deterministic = True
output_dir = tracktor['output_dir']
sacred_config = osp.join(output_dir, 'sacred_config.yaml')
if not osp.exists(output_dir):
os.makedirs(output_dir)
dt = 10 # how many steps to include in the last loss
log_dir = 'debug_out' # directory to dump logs and debug plots
net_path = None # path of to network file to initialize weights with
# config to control run_from_file
run_config = {
'usage': 'test', # what dataset to use {training, validation, test}
'batch_size': 100,
'rollout_steps': 10,
'debug_samples':
[0, 1, 2], # sample ids to generate plots for (None, int, list)
}
ex.add_named_config('no_score', {'record_grouping_score': False})
ex.add_named_config('no_collisions', {'record_relational_loss': None})
@ex.capture
def add_noise(data, noise):
noise_type = noise['noise_type']
if noise_type in ['None', 'none', None]:
return data
n = torch.bernoulli(noise['prob'] * torch.ones(data.size()))
p = 2 * torch.mul(data, n)
corrupted = data + n - p # hacky way of implementing (data XOR n)
return corrupted
finalvideo=12
if finalvideo==1:
ex = Experiment()
ex.add_config('C:/Users/EloiMartins/PycharmProjects/tracking_wo_bnw/experiments/cfgs/tracktor.yaml')
# hacky workaround to load the corresponding configs and not having to hardcode paths here
ex.add_config(ex.configurations[0]._conf['tracktor']['reid_config'])
ex.add_named_config('oracle', 'C:/Users/EloiMartins/PycharmProjects/tracking_wo_bnw/experiments/cfgs/tracktor.yaml')
@ex.automain
def main(tracktor, reid, _config, _log, _run):
sacred.commands.print_config(_run)
# set all seeds
torch.manual_seed(tracktor['seed'])
torch.cuda.manual_seed(tracktor['seed'])
np.random.seed(tracktor['seed'])
torch.backends.cudnn.deterministic = True
output_dir = osp.join(get_output_dir(tracktor['module_name']), tracktor['name'])
sacred_config = osp.join(output_dir, 'sacred_config.yaml')
if context.epoch % context.cfg.tracktor_val_every == 0 and context.validate:
with torch.no_grad():
metrics = {
**metrics,
**self.validate_tracktor(context.model, context.epoch)
}
return metrics
ex = Experiment()
ex.add_config('experiments/cfgs/tracktor.yaml')
ex.add_config('experiments/cfgs/train_motion_im.yaml')
ex.add_config(ex.configurations[0]._conf['tracktor']['reid_config'])
ex.add_named_config('correlation_model',
'experiments/cfgs/correlation_model.yaml')
ex.add_named_config('pos_model', 'experiments/cfgs/pos_model.yaml')
@ex.automain
def main(tracktor, reid, train, _config, _log, _run):
# instantiate tracktor
sacred.commands.print_config(_run)
# set all seeds
seed_everything(tracktor['seed'])
output_dir = osp.abspath(
osp.join(osp.dirname(__file__), '..', '..', 'output', 'motion',
train['name']))
sacred_config = osp.join(output_dir, 'sacred_config.yaml')
from tracktor.frcnn_fpn import FRCNN_FPN
from tracktor.frcnn_fpn_detectron2 import FRCNN_FPN_Detectron2
from tracktor.config import get_output_dir
from tracktor.datasets.factory import Datasets
from tracktor.oracle_tracker import OracleTracker
from tracktor.tracker import Tracker
from tracktor.reid.resnet import resnet50
from tracktor.utils import interpolate, plot_sequence, get_mot_accum, evaluate_mot_accums
ex = Experiment()
ex.add_config('experiments/cfgs/tracktor.yaml')
# hacky workaround to load the corresponding configs and not having to hardcode paths here
ex.add_config(ex.configurations[0]._conf['tracktor']['reid_config'])
ex.add_named_config('oracle', 'experiments/cfgs/oracle_tracktor.yaml')
ex.add_named_config(
'detectron2_coco_r50fpn1x_no_camera_no_reid',
'experiments/cfgs/tracktor_detectron2_r50fpn1x_coco_no_camera_no_reid.yaml'
)
@ex.automain
def main(tracktor, reid, _config, _log, _run):
sacred.commands.print_config(_run)
# set all seeds
torch.manual_seed(tracktor['seed'])
torch.cuda.manual_seed(tracktor['seed'])
np.random.seed(tracktor['seed'])
torch.backends.cudnn.deterministic = True