def after_step(self):
next_iter = self.trainer.iter + 1
is_final = next_iter == self.trainer.max_iter
if is_final or (self._period > 0 and next_iter % self._period == 0):
results = self._func()
if results:
assert isinstance(
results, dict
), "Eval function must return a dict. Got {} instead.".format(
results)
flattened_results = flatten_results_dict(results)
for k, v in flattened_results.items():
try:
v = float(v)
except Exception:
raise ValueError(
"[EvalHook] eval_function should return a nested dict of float. "
"Got '{}: {}' instead.".format(k, v))
self.trainer.storage.put_scalars(**flattened_results,
smoothing_hint=False)
# Evaluation may take different time among workers.
# A barrier make them start the next iteration together.
comm.synchronize()
def _distributed_worker(
local_rank, main_func, world_size, num_gpus_per_machine, machine_rank, dist_url, args
):
assert torch.cuda.is_available(), "cuda is not available. Please check your installation."
global_rank = machine_rank * num_gpus_per_machine + local_rank
try:
dist.init_process_group(
backend="NCCL", init_method=dist_url, world_size=world_size, rank=global_rank
)
except Exception as e:
logger = logging.getLogger(__name__)
logger.error("Process group URL: {}".format(dist_url))
raise e
# synchronize is needed here to prevent a possible timeout after calling init_process_group
# See: https://github.com/facebookresearch/maskrcnn-benchmark/issues/172
comm.synchronize()
assert num_gpus_per_machine <= torch.cuda.device_count()
torch.cuda.set_device(local_rank)
# Setup the local process group (which contains ranks within the same machine)
assert comm._LOCAL_PROCESS_GROUP is None
num_machines = world_size // num_gpus_per_machine
for i in range(num_machines):
ranks_on_i = list(range(i * num_gpus_per_machine, (i + 1) * num_gpus_per_machine))
pg = dist.new_group(ranks_on_i)
if i == machine_rank:
comm._LOCAL_PROCESS_GROUP = pg
main_func(*args)
def evaluate(self):
if self._distributed:
comm.synchronize()
self._predictions = comm.gather(self._predictions, dst=0)
self._predictions = list(itertools.chain(*self._predictions))
if not comm.is_main_process():
return {}
if len(self._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(self._predictions, f)
self._results = OrderedDict()
if "proposals" in self._predictions[0]:
self._eval_box_proposals()
if "instances" in self._predictions[0]:
self._eval_predictions(set(self._tasks))
# Copy so the caller can do whatever with results
return copy.deepcopy(self._results)
def evaluate(self):
"""
Returns:
dict: has a key "segm", whose value is a dict of "AP" and "AP50".
"""
comm.synchronize()
if comm.get_rank() > 0:
return
os.environ["CITYSCAPES_DATASET"] = os.path.abspath(
os.path.join(self._metadata.gt_dir, "..", ".."))
# Load the Cityscapes eval script *after* setting the required env var,
# since the script reads CITYSCAPES_DATASET into global variables at load time.
import cityscapesscripts.evaluation.evalInstanceLevelSemanticLabeling as cityscapes_eval
self._logger.info("Evaluating results under {} ...".format(
self._temp_dir))
# set some global states in cityscapes evaluation API, before evaluating
cityscapes_eval.args.predictionPath = os.path.abspath(self._temp_dir)
cityscapes_eval.args.predictionWalk = None
cityscapes_eval.args.JSONOutput = False
cityscapes_eval.args.colorized = False
cityscapes_eval.args.gtInstancesFile = os.path.join(
self._temp_dir, "gtInstances.json")
# These lines are adopted from
# https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalInstanceLevelSemanticLabeling.py # noqa
groundTruthImgList = glob.glob(cityscapes_eval.args.groundTruthSearch)
assert len(
groundTruthImgList
), "Cannot find any ground truth images to use for evaluation. Searched for: {}".format(
cityscapes_eval.args.groundTruthSearch)
predictionImgList = []
for gt in groundTruthImgList:
predictionImgList.append(
cityscapes_eval.getPrediction(gt, cityscapes_eval.args))
results = cityscapes_eval.evaluateImgLists(
predictionImgList, groundTruthImgList,
cityscapes_eval.args)["averages"]
ret = OrderedDict()
ret["segm"] = {
"AP": results["allAp"] * 100,
"AP50": results["allAp50%"] * 100
}
self._working_dir.cleanup()
return ret
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))
self._real_flops = all_gather(self._real_flops)
self._real_flops = list(itertools.chain(*self._real_flops))
self._expt_flops = all_gather(self._expt_flops)
self._expt_flops = list(itertools.chain(*self._expt_flops))
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
# add flops calculation
if len(self._real_flops) > 0 and len(self._expt_flops) > 0:
res["mean_real_flops"] = (sum(self._real_flops) /
len(self._real_flops)) / 1e3
res["max_real_flops"] = max(self._real_flops) / 1e3
res["min_real_flops"] = min(self._real_flops) / 1e3
res["mean_expt_flops"] = (sum(self._expt_flops) /
len(self._expt_flops)) / 1e3
res["max_expt_flops"] = max(self._expt_flops) / 1e3
res["min_expt_flops"] = min(self._expt_flops) / 1e3
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 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 for loss in 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)