def main(args):
cfg = setup(args)
if args.eval_only:
model = Trainer.build_model(cfg)
DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(
cfg.MODEL.WEIGHTS, resume=args.resume)
res = Trainer.test(cfg, model)
if cfg.TEST.AUG.ENABLED:
res.update(Trainer.test_with_TTA(cfg, model))
if comm.is_main_process():
verify_results(cfg, res)
return res
"""
If you'd like to do anything fancier than the standard training logic,
consider writing your own training loop or subclassing the trainer.
"""
trainer = Trainer(cfg)
trainer.resume_or_load(resume=args.resume)
if cfg.TEST.AUG.ENABLED:
trainer.register_hooks([
hooks.EvalHook(0,
lambda: trainer.test_with_TTA(cfg, trainer.model))
])
return trainer.train()
def evaluate(self):
if self._distributed:
comm.synchronize()
predictions = comm.gather(self._predictions, dst=0)
predictions = list(itertools.chain(*predictions))
if not comm.is_main_process():
return {}
else:
predictions = self._predictions
if len(predictions) == 0:
self._logger.warning(
"[COCOEvaluator] Did not receive valid predictions.")
return {}
if self._output_dir:
PathManager.mkdirs(self._output_dir)
file_path = os.path.join(self._output_dir,
"instances_predictions.pth")
with PathManager.open(file_path, "wb") as f:
torch.save(predictions, f)
self._results = OrderedDict()
if "proposals" in predictions[0]:
self._eval_box_proposals(predictions)
if "instances" in predictions[0]:
self._eval_predictions(set(self._tasks), predictions)
# Copy so the caller can do whatever with results
return copy.deepcopy(self._results)
def _write_metrics(self, metrics_dict: dict):
"""
Args:
metrics_dict (dict): dict of scalar metrics
"""
metrics_dict = {
k: v.detach().cpu().item()
if isinstance(v, torch.Tensor) else float(v)
for k, v in metrics_dict.items()
}
# gather metrics among all workers for logging
# This assumes we do DDP-style training, which is currently the only
# supported method in mydl.
all_metrics_dict = comm.gather(metrics_dict)
if comm.is_main_process():
if "data_time" in all_metrics_dict[0]:
# data_time among workers can have high variance. The actual latency
# caused by data_time is the maximum among workers.
data_time = np.max(
[x.pop("data_time") for x in all_metrics_dict])
self.storage.put_scalar("data_time", data_time)
# average the rest metrics
metrics_dict = {
k: np.mean([x[k] for x in all_metrics_dict])
for k in all_metrics_dict[0].keys()
}
total_losses_reduced = sum(loss for loss in metrics_dict.values())
self.storage.put_scalar("total_loss", total_losses_reduced)
if len(metrics_dict) > 1:
self.storage.put_scalars(**metrics_dict)
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
return ret
def default_setup(cfg, args):
"""
Perform some basic common setups at the beginning of a job, including:
1. Set up the mydl logger
2. Log basic information about environment, cmdline arguments, and config
3. Backup the config to the output directory
Args:
cfg (CfgNode): the full config to be used
args (argparse.NameSpace): the command line arguments to be logged
"""
output_dir = cfg.OUTPUT_DIR
if comm.is_main_process() and output_dir:
PathManager.mkdirs(output_dir)
rank = comm.get_rank()
setup_logger(output_dir, distributed_rank=rank, name="fvcore")
logger = setup_logger(output_dir, distributed_rank=rank)
logger.info("Rank of current process: {}. World size: {}".format(
rank, comm.get_world_size()))
logger.info("Environment info:\n" + collect_env_info())
logger.info("Command line arguments: " + str(args))
if hasattr(args, "config_file") and args.config_file != "":
logger.info("Contents of args.config_file={}:\n{}".format(
args.config_file,
PathManager.open(args.config_file, "r").read()))
logger.info("Running with full config:\n{}".format(cfg))
if comm.is_main_process() and output_dir:
# Note: some of our scripts may expect the existence of
# config.yaml in output directory
path = os.path.join(output_dir, "config.yaml")
with PathManager.open(path, "w") as f:
f.write(cfg.dump())
logger.info("Full config saved to {}".format(os.path.abspath(path)))
if cfg.VERSION == 2:
# make sure each worker has a different, yet deterministic seed if specified
seed_all_rng(None if cfg.SEED < 0 else cfg.SEED + rank)
# cudnn benchmark has large overhead. It shouldn't be used considering the small size of
# typical validation set.
if not (hasattr(args, "eval_only") and args.eval_only):
torch.backends.cudnn.benchmark = cfg.CUDNN_BENCHMARK
def evaluate(self):
"""
Evaluates standard semantic segmentation metrics (http://cocodataset.org/#stuff-eval):
* Mean intersection-over-union averaged across classes (mIoU)
* Frequency Weighted IoU (fwIoU)
* Mean pixel accuracy averaged across classes (mACC)
* Pixel Accuracy (pACC)
"""
if self._distributed:
synchronize()
conf_matrix_list = all_gather(self._conf_matrix)
self._predictions = all_gather(self._predictions)
self._predictions = list(itertools.chain(*self._predictions))
if not is_main_process():
return
self._conf_matrix = np.zeros_like(self._conf_matrix)
for conf_matrix in conf_matrix_list:
self._conf_matrix += conf_matrix
if self._output_dir:
PathManager.mkdirs(self._output_dir)
file_path = os.path.join(self._output_dir,
"sem_seg_predictions.json")
with PathManager.open(file_path, "w") as f:
f.write(json.dumps(self._predictions))
acc = np.zeros(self._num_classes, dtype=np.float)
iou = np.zeros(self._num_classes, dtype=np.float)
tp = self._conf_matrix.diagonal()[:-1].astype(np.float)
pos_gt = np.sum(self._conf_matrix[:-1, :-1], axis=0).astype(np.float)
class_weights = pos_gt / np.sum(pos_gt)
pos_pred = np.sum(self._conf_matrix[:-1, :-1], axis=1).astype(np.float)
acc_valid = pos_gt > 0
acc[acc_valid] = tp[acc_valid] / pos_gt[acc_valid]
iou_valid = (pos_gt + pos_pred) > 0
union = pos_gt + pos_pred - tp
iou[acc_valid] = tp[acc_valid] / union[acc_valid]
macc = np.sum(acc) / np.sum(acc_valid)
miou = np.sum(iou) / np.sum(iou_valid)
fiou = np.sum(iou * class_weights)
pacc = np.sum(tp) / np.sum(pos_gt)
res = {}
res["mIoU"] = 100 * miou
res["fwIoU"] = 100 * fiou
res["mACC"] = 100 * macc
res["pACC"] = 100 * pacc
if self._output_dir:
file_path = os.path.join(self._output_dir,
"sem_seg_evaluation.pth")
with PathManager.open(file_path, "wb") as f:
torch.save(res, f)
results = OrderedDict({"sem_seg": res})
self._logger.info(results)
return results
def evaluate(self):
if self._distributed:
synchronize()
self._predictions = all_gather(self._predictions)
self._predictions = list(itertools.chain(*self._predictions))
if not is_main_process():
return
return copy.deepcopy(self._eval_predictions())
def evaluate(self):
results = OrderedDict()
for evaluator in self._evaluators:
result = evaluator.evaluate()
if is_main_process() and result is not None:
for k, v in result.items():
assert (
k not in results
), "Different evaluators produce results with the same key {}".format(
k)
results[k] = v
return results
def train(self):
"""
Run training.
Returns:
OrderedDict of results, if evaluation is enabled. Otherwise None.
"""
super().train(self.start_iter, self.max_iter)
if len(self.cfg.TEST.EXPECTED_RESULTS) and comm.is_main_process():
assert hasattr(self, "_last_eval_results"
), "No evaluation results obtained during training!"
verify_results(self.cfg, self._last_eval_results)
return self._last_eval_results
def test(cls, cfg, model, evaluators=None):
"""
Args:
cfg (CfgNode):
model (nn.Module):
evaluators (list[DatasetEvaluator] or None): if None, will call
:meth:`build_evaluator`. Otherwise, must have the same length as
`cfg.DATASETS.TEST`.
Returns:
dict: a dict of result metrics
"""
logger = logging.getLogger(__name__)
if isinstance(evaluators, DatasetEvaluator):
evaluators = [evaluators]
if evaluators is not None:
assert len(
cfg.DATASETS.TEST) == len(evaluators), "{} != {}".format(
len(cfg.DATASETS.TEST), len(evaluators))
results = OrderedDict()
for idx, dataset_name in enumerate(cfg.DATASETS.TEST):
data_loader = cls.build_test_loader(cfg, dataset_name)
# When evaluators are passed in as arguments,
# implicitly assume that evaluators can be created before data_loader.
if evaluators is not None:
evaluator = evaluators[idx]
else:
try:
evaluator = cls.build_evaluator(cfg, dataset_name)
except NotImplementedError:
logger.warn(
"No evaluator found. Use `DefaultTrainer.test(evaluators=)`, "
"or implement its `build_evaluator` method.")
results[dataset_name] = {}
continue
results_i = inference_on_dataset(model, data_loader, evaluator)
results[dataset_name] = results_i
if comm.is_main_process():
assert isinstance(
results_i, dict
), "Evaluator must return a dict on the main process. Got {} instead.".format(
results_i)
logger.info("Evaluation results for {} in csv format:".format(
dataset_name))
print_csv_format(results_i)
if len(results) == 1:
results = list(results.values())[0]
return results
def evaluate(self):
comm.synchronize()
self._predictions = comm.gather(self._predictions)
self._predictions = list(itertools.chain(*self._predictions))
if not comm.is_main_process():
return
# PanopticApi requires local files
gt_json = PathManager.get_local_path(self._metadata.panoptic_json)
gt_folder = PathManager.get_local_path(self._metadata.panoptic_root)
with tempfile.TemporaryDirectory(prefix="panoptic_eval") as pred_dir:
logger.info("Writing all panoptic predictions to {} ...".format(pred_dir))
for p in self._predictions:
with open(os.path.join(pred_dir, p["file_name"]), "wb") as f:
f.write(p.pop("png_string"))
with open(gt_json, "r") as f:
json_data = json.load(f)
json_data["annotations"] = self._predictions
with PathManager.open(self._predictions_json, "w") as f:
f.write(json.dumps(json_data))
from panopticapi.evaluation import pq_compute
with contextlib.redirect_stdout(io.StringIO()):
pq_res = pq_compute(
gt_json,
PathManager.get_local_path(self._predictions_json),
gt_folder=gt_folder,
pred_folder=pred_dir,
)
res = {}
res["PQ"] = 100 * pq_res["All"]["pq"]
res["SQ"] = 100 * pq_res["All"]["sq"]
res["RQ"] = 100 * pq_res["All"]["rq"]
res["PQ_th"] = 100 * pq_res["Things"]["pq"]
res["SQ_th"] = 100 * pq_res["Things"]["sq"]
res["RQ_th"] = 100 * pq_res["Things"]["rq"]
res["PQ_st"] = 100 * pq_res["Stuff"]["pq"]
res["SQ_st"] = 100 * pq_res["Stuff"]["sq"]
res["RQ_st"] = 100 * pq_res["Stuff"]["rq"]
results = OrderedDict({"panoptic_seg": res})
_print_panoptic_results(pq_res)
return results
def do_test(cfg, model):
results = OrderedDict()
for dataset_name in cfg.DATASETS.TEST:
data_loader = build_detection_test_loader(cfg, dataset_name)
evaluator = get_evaluator(
cfg, dataset_name, os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name)
)
results_i = inference_on_dataset(model, data_loader, evaluator)
results[dataset_name] = results_i
if comm.is_main_process():
logger.info("Evaluation results for {} in csv format:".format(dataset_name))
print_csv_format(results_i)
if len(results) == 1:
results = list(results.values())[0]
return results
def main(args):
cfg = setup(args)
if args.eval_only:
model = Trainer.build_model(cfg)
DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(
cfg.MODEL.WEIGHTS, resume=args.resume)
res = Trainer.test(cfg, model)
if comm.is_main_process():
verify_results(cfg, res)
return res
trainer = Trainer(cfg)
trainer.resume_or_load(resume=args.resume)
return trainer.train()
def evaluate(self):
"""
Returns:
dict: has a key "segm", whose value is a dict of "AP", "AP50", and "AP75".
"""
all_predictions = comm.gather(self._predictions, dst=0)
if not comm.is_main_process():
return
predictions = defaultdict(list)
for predictions_per_rank in all_predictions:
for clsid, lines in predictions_per_rank.items():
predictions[clsid].extend(lines)
del all_predictions
self._logger.info(
"Evaluating {} using {} metric. "
"Note that results do not use the official Matlab API.".format(
self._dataset_name, 2007 if self._is_2007 else 2012))
with tempfile.TemporaryDirectory(prefix="pascal_voc_eval_") as dirname:
res_file_template = os.path.join(dirname, "{}.txt")
aps = defaultdict(list) # iou -> ap per class
for cls_id, cls_name in enumerate(self._class_names):
lines = predictions.get(cls_id, [""])
with open(res_file_template.format(cls_name), "w") as f:
f.write("\n".join(lines))
for thresh in range(50, 100, 5):
rec, prec, ap = voc_eval(
res_file_template,
self._anno_file_template,
self._image_set_path,
cls_name,
ovthresh=thresh / 100.0,
use_07_metric=self._is_2007,
)
aps[thresh].append(ap * 100)
ret = OrderedDict()
mAP = {iou: np.mean(x) for iou, x in aps.items()}
ret["bbox"] = {
"AP": np.mean(list(mAP.values())),
"AP50": mAP[50],
"AP75": mAP[75]
}
return ret
def cache_url(url, model_dir=None, progress=True):
r"""Loads the Torch serialized object at the given URL.
If the object is already present in `model_dir`, it's deserialized and
returned. The filename part of the URL should follow the naming convention
``filename-<sha256>.ext`` where ``<sha256>`` is the first eight or more
digits of the SHA256 hash of the contents of the file. The hash is used to
ensure unique names and to verify the contents of the file.
The default value of `model_dir` is ``$TORCH_HOME/models`` where
``$TORCH_HOME`` defaults to ``~/.torch``. The default directory can be
overridden with the ``$TORCH_MODEL_ZOO`` environment variable.
Args:
url (string): URL of the object to download
model_dir (string, optional): directory in which to save the object
progress (bool, optional): whether or not to display a progress bar to stderr
Example:
>>> cached_file = mydl.utils.model_zoo.cache_url('https://s3.amazonaws.com/pytorch/models/resnet18-5c106cde.pth')
"""
if model_dir is None:
torch_home = os.path.expanduser(os.getenv("TORCH_HOME", "~/.torch"))
model_dir = os.getenv("TORCH_MODEL_ZOO",
os.path.join(torch_home, "models"))
if not os.path.exists(model_dir):
os.makedirs(model_dir)
parts = urlparse(url)
filename = os.path.basename(parts.path)
if filename == "model_final.pkl":
# workaround as pre-trained Caffe2 models from Detectron have all the same filename
# so make the full path the filename by replacing / with _
filename = parts.path.replace("/", "_")
cached_file = os.path.join(model_dir, filename)
if not os.path.exists(cached_file) and is_main_process():
sys.stderr.write('Downloading: "{}" to {}\n'.format(url, cached_file))
hash_prefix = HASH_REGEX.search(filename)
if hash_prefix is not None:
hash_prefix = hash_prefix.group(1)
# workaround: Caffe2 models don't have a hash, but follow the R-50 convention,
# which matches the hash PyTorch uses. So we skip the hash matching
# if the hash_prefix is less than 6 characters
if len(hash_prefix) < 6:
hash_prefix = None
_download_url_to_file(url, cached_file, hash_prefix, progress=progress)
synchronize()
return cached_file
def do_train(cfg, model, resume=False):
model.train()
optimizer = build_optimizer(cfg, model)
scheduler = build_lr_scheduler(cfg, optimizer)
checkpointer = DetectionCheckpointer(
model, cfg.OUTPUT_DIR, optimizer=optimizer, scheduler=scheduler
)
start_iter = (
checkpointer.resume_or_load(cfg.MODEL.WEIGHTS, resume=resume).get("iteration", -1) + 1
)
max_iter = cfg.SOLVER.MAX_ITER
periodic_checkpointer = PeriodicCheckpointer(
checkpointer, cfg.SOLVER.CHECKPOINT_PERIOD, max_iter=max_iter
)
writers = (
[
CommonMetricPrinter(max_iter),
JSONWriter(os.path.join(cfg.OUTPUT_DIR, "metrics.json")),
TensorboardXWriter(cfg.OUTPUT_DIR),
]
if comm.is_main_process()
else []
)
# compared to "train_net.py", we do not support accurate timing and
# precise BN here, because they are not trivial to implement
data_loader = build_detection_train_loader(cfg)
logger.info("Starting training from iteration {}".format(start_iter))
with EventStorage(start_iter) as storage:
for data, iteration in zip(data_loader, range(start_iter, max_iter)):
iteration = iteration + 1
storage.step()
loss_dict = model(data)
losses = sum(loss_dict.values())
assert torch.isfinite(losses).all(), loss_dict
loss_dict_reduced = {k: v.item() for k, v in comm.reduce_dict(loss_dict).items()}
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
if comm.is_main_process():
storage.put_scalars(total_loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
losses.backward()
optimizer.step()
storage.put_scalar("lr", optimizer.param_groups[0]["lr"], smoothing_hint=False)
scheduler.step()
if (
cfg.TEST.EVAL_PERIOD > 0
and iteration % cfg.TEST.EVAL_PERIOD == 0
and iteration != max_iter
):
do_test(cfg, model)
# Compared to "train_net.py", the test results are not dumped to EventStorage
comm.synchronize()
if iteration - start_iter > 5 and (iteration % 20 == 0 or iteration == max_iter):
for writer in writers:
writer.write()
periodic_checkpointer.step(iteration)
