def get(config_path, trained: bool = False):
"""
Get a model specified by relative path under mydl's official ``configs/`` directory.
Args:
config_path (str): config file name relative to mydl's "configs/"
directory, e.g., "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml"
trained (bool): If True, will initialize the model with the trained model zoo weights.
If False, the checkpoint specified in the config file's ``MODEL.WEIGHTS`` is used
instead; this will typically (though not always) initialize a subset of weights using
an ImageNet pre-trained model, while randomly initializing the other weights.
Example:
.. code-block:: python
from mydl import model_zoo
model = model_zoo.get("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml", trained=True)
"""
cfg_file = get_config_file(config_path)
cfg = get_cfg()
cfg.merge_from_file(cfg_file)
if trained:
cfg.MODEL.WEIGHTS = get_checkpoint_url(config_path)
if not torch.cuda.is_available():
cfg.MODEL.DEVICE = "cpu"
model = build_model(cfg)
DetectionCheckpointer(model).load(cfg.MODEL.WEIGHTS)
return model
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 benchmark_train(args):
cfg = setup(args)
model = build_model(cfg)
logger.info("Model:\n{}".format(model))
if comm.get_world_size() > 1:
model = DistributedDataParallel(model,
device_ids=[comm.get_local_rank()],
broadcast_buffers=False)
optimizer = build_optimizer(cfg, model)
checkpointer = DetectionCheckpointer(model, optimizer=optimizer)
checkpointer.load(cfg.MODEL.WEIGHTS)
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0
data_loader = build_detection_train_loader(cfg)
dummy_data = list(itertools.islice(data_loader, 100))
def f():
data = DatasetFromList(dummy_data, copy=False)
while True:
yield from data
max_iter = 400
trainer = SimpleTrainer(model, f(), optimizer)
trainer.register_hooks([
hooks.IterationTimer(),
hooks.PeriodicWriter([CommonMetricPrinter(max_iter)])
])
trainer.train(1, max_iter)
def benchmark_eval(args):
cfg = setup(args)
model = build_model(cfg)
model.eval()
logger.info("Model:\n{}".format(model))
DetectionCheckpointer(model).load(cfg.MODEL.WEIGHTS)
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0
data_loader = build_detection_test_loader(cfg, cfg.DATASETS.TEST[0])
dummy_data = list(itertools.islice(data_loader, 100))
def f():
while True:
yield from DatasetFromList(dummy_data, copy=False)
for _ in range(5): # warmup
model(dummy_data[0])
max_iter = 400
timer = Timer()
with tqdm.tqdm(total=max_iter) as pbar:
for idx, d in enumerate(f()):
if idx == max_iter:
break
model(d)
pbar.update()
logger.info("{} iters in {} seconds.".format(max_iter, timer.seconds()))
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)
return res
trainer = Trainer(cfg)
trainer.resume_or_load(resume=args.resume)
return trainer.train()
def __init__(self, cfg):
self.cfg = cfg.clone() # cfg can be modified by model
self.model = build_model(self.cfg)
self.model.eval()
self.metadata = MetadataCatalog.get(cfg.DATASETS.TEST[0])
checkpointer = DetectionCheckpointer(self.model)
checkpointer.load(cfg.MODEL.WEIGHTS)
self.transform_gen = T.ResizeShortestEdge(
[cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MIN_SIZE_TEST],
cfg.INPUT.MAX_SIZE_TEST)
self.input_format = cfg.INPUT.FORMAT
assert self.input_format in ["RGB", "BGR"], self.input_format
def main(args):
cfg = setup(args)
model = build_model(cfg)
logger.info("Model:\n{}".format(model))
if args.eval_only:
DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(
cfg.MODEL.WEIGHTS, resume=args.resume
)
return do_test(cfg, model)
distributed = comm.get_world_size() > 1
if distributed:
model = DistributedDataParallel(
model, device_ids=[comm.get_local_rank()], broadcast_buffers=False
)
do_train(cfg, model)
return do_test(cfg, model)
def __init__(self, cfg):
"""
Args:
cfg (CfgNode):
"""
logger = logging.getLogger("mydl")
if not logger.isEnabledFor(
logging.INFO): # setup_logger is not called for d2
setup_logger()
# Assume these objects must be constructed in this order.
model = self.build_model(cfg)
optimizer = self.build_optimizer(cfg, model)
data_loader = self.build_train_loader(cfg)
# For training, wrap with DDP. But don't need this for inference.
if comm.get_world_size() > 1:
model = DistributedDataParallel(model,
device_ids=[comm.get_local_rank()],
broadcast_buffers=False)
super().__init__(model, data_loader, optimizer)
self.scheduler = self.build_lr_scheduler(cfg, optimizer)
# Assume no other objects need to be checkpointed.
# We can later make it checkpoint the stateful hooks
self.checkpointer = DetectionCheckpointer(
# Assume you want to save checkpoints together with logs/statistics
model,
cfg.OUTPUT_DIR,
optimizer=optimizer,
scheduler=self.scheduler,
)
self.start_iter = 0
self.max_iter = cfg.SOLVER.MAX_ITER
self.cfg = cfg
self.register_hooks(self.build_hooks())
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)
parser.add_argument("--output", help="output directory for the converted caffe2 model")
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
logger = setup_logger()
logger.info("Command line arguments: " + str(args))
cfg = setup_cfg(args)
# create a torch model
torch_model = build_model(cfg)
DetectionCheckpointer(torch_model).resume_or_load(cfg.MODEL.WEIGHTS)
# get a sample data
data_loader = build_detection_test_loader(cfg, cfg.DATASETS.TEST[0])
first_batch = next(iter(data_loader))
# convert and save caffe2 model
caffe2_model = export_caffe2_model(cfg, torch_model, first_batch)
caffe2_model.save_protobuf(args.output)
# draw the caffe2 graph
caffe2_model.save_graph(os.path.join(args.output, "model.svg"), inputs=first_batch)
# run evaluation with the converted model
if args.run_eval:
dataset = cfg.DATASETS.TEST[0]
data_loader = build_detection_test_loader(cfg, dataset)