def test(cfg, model, distributed):
if distributed:
model = model.module
torch.cuda.empty_cache() # TODO check if it helps
iou_types = ("bbox",)
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm",)
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg, is_train=False, is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
synchronize()
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.deprecated.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
)
return model
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"--ckpt",
help=
"The path to the checkpoint for test, default is the latest checkpoint.",
default=None,
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
ckpt = cfg.MODEL.WEIGHT if args.ckpt is None else args.ckpt
_ = checkpointer.load(ckpt, use_latest=args.ckpt is None)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
synchronize()
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = torch.optim.lr_scheduler.StepLR(
optimizer, step_size=1000,
gamma=0.9) #make_lr_scheduler(cfg, optimizer)
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model,
optimizer,
opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg,
model,
None,
None,
output_dir,
save_to_disk # , optimizer, scheduler
)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
)
return model
def train(cfg, local_rank, distributed, use_tensorboard=False, logger=None):
arguments = {"iteration": 0}
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
if cfg.SOLVER.UNFREEZE_CONV_BODY:
for p in model.backbone.parameters():
p.requires_grad = True
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg,
model,
optimizer,
scheduler,
output_dir,
save_to_disk,
logger=logger)
print(cfg.TRAIN.IGNORE_LIST)
extra_checkpoint_data = checkpointer.load(
cfg.MODEL.WEIGHT, ignore_list=cfg.TRAIN.IGNORE_LIST)
arguments.update(extra_checkpoint_data)
if cfg.SOLVER.KEEP_LR:
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
tensorboard_logdir = cfg.OUTPUT_DIR
tensorboard_exp_name = cfg.TENSORBOARD_EXP_NAME
snapshot = cfg.SOLVER.SNAPSHOT_ITERS
do_train(model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
snapshot,
tensorboard_logdir,
tensorboard_exp_name,
use_tensorboard=use_tensorboard)
return model
def do_train(
cfg,
model,
data_loader,
data_loader_val,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
test_period,
arguments,
):
logger = logging.getLogger("maskrcnn_benchmark.trainer")
logger.info("Start training")
meters = MetricLogger(delimiter=" ")
max_iter = len(data_loader)
start_iter = arguments["iteration"]
model.train()
start_training_time = time.time()
end = time.time()
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
dataset_names = cfg.DATASETS.TEST
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
if any(len(target) < 1 for target in targets):
logger.error(
f"Iteration={iteration + 1} || Image Ids used for training {_} || targets Length={[len(target) for target in targets]}"
)
continue
data_time = time.time() - end
iteration = iteration + 1
arguments["iteration"] = iteration
images = images.to(device)
targets = [target.to(device) for target in targets]
# 计算损失，model会计算images与targets的损失，并返回
# 实际调用的是GeneralizedRCNN类的forward方法
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
# 多GPU的时候，会有无效损失，去除这些损失
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
# Note: If mixed precision is not used, this ends up doing nothing
# Otherwise apply loss scaling for mixed-precision recipe
# amp的用法
with amp.scale_loss(losses, optimizer) as scaled_losses:
scaled_losses.backward()
optimizer.step()
# scheduler也需要step
scheduler.step()
# 每个batch要多久
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time, data=data_time)
# 还需要多久
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
# 每20次迭代或者最好一次迭代了，输出信息
if iteration % 20 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join([
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string, # 还需要多久
iter=
iteration, # 第几次迭代，是从maskrcnn_benchmark.utils.checkpoint记录的迭代次数开始算起
meters=str(
meters
), # 包含loss、loss_classifier、loss_box_reg、loss_objectness、loss_rpn_box_reg
lr=optimizer.param_groups[0]["lr"], # 学习率
memory=torch.cuda.max_memory_allocated() / 1024.0 /
1024.0, # 占用最大的GPU内存
))
# 每隔checkpoint_period保存一次checkpoint
if iteration % checkpoint_period == 0:
checkpointer.save("model_{:07d}".format(iteration), **arguments)
if data_loader_val is not None and test_period > 0 and iteration % test_period == 0:
meters_val = MetricLogger(delimiter=" ")
synchronize()
_ = inference( # The result can be used for additional logging, e. g. for TensorBoard
model,
# The method changes the segmentation mask format in a data loader,
# so every time a new data loader is created:
make_data_loader(cfg,
is_train=False,
is_distributed=(get_world_size() > 1),
is_for_period=True),
dataset_name="[Validation]",
iou_types=iou_types,
box_only=False
if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=None,
)
synchronize()
model.train()
with torch.no_grad():
# Should be one image for each GPU:
for iteration_val, (images_val, targets_val,
_) in enumerate(tqdm(data_loader_val)):
images_val = images_val.to(device)
targets_val = [target.to(device) for target in targets_val]
loss_dict = model(images_val, targets_val)
losses = sum(loss for loss in loss_dict.values())
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(
loss for loss in loss_dict_reduced.values())
meters_val.update(loss=losses_reduced, **loss_dict_reduced)
synchronize()
logger.info(
meters_val.delimiter.join([
"[Validation]: ",
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters_val),
lr=optimizer.param_groups[0]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
))
# 最后一次迭代的时候保存checkpoint
if iteration == max_iter:
checkpointer.save("model_final", **arguments)
# train总共花了多少时间
total_training_time = time.time() - start_training_time
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / (max_iter)))
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument(
"--json-file",
default="",
metavar="FILE",
help="path to prediction bbox json file",
)
# parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
# num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
# distributed = num_gpus > 1
# if distributed:
# torch.cuda.set_device(args.local_rank)
# torch.distributed.init_process_group(
# backend="nccl", init_method="env://"
# )
# synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
# logger.info("Using {} GPUs".format(num_gpus))
# logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
# model = build_detection_model(cfg)
# model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
# checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
# _ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox", )
# if cfg.MODEL.MASK_ON:
# iou_types = iou_types + ("segm",)
# if cfg.MODEL.KEYPOINT_ON:
# iou_types = iou_types + ("keypoints",)
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=False)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
# inference(
# model,
# data_loader_val,
# dataset_name=dataset_name,
# iou_types=iou_types,
# box_only=False if cfg.MODEL.FCOS_ON or cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
# device=cfg.MODEL.DEVICE,
# expected_results=cfg.TEST.EXPECTED_RESULTS,
# expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
# output_folder=output_folder,
# )
# extra_args = dict(
# box_only=False,
# iou_types=iou_types,
# expected_results=cfg.TEST.EXPECTED_RESULTS,
# expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
# )
dataset = data_loader_val.dataset
# evaluate(dataset=dataset,
# predictions=predictions,
# output_folder=output_folder,
# only_human=True,
# **extra_args)
do_coco_json_evaluation(
dataset=dataset,
json_file=args.json_file,
box_only=False,
output_folder=output_folder,
iou_types=iou_types,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL)
synchronize()
def train(cfg, local_rank, distributed, logger):
if is_main_process():
wandb.init(project='scene-graph',
entity='sgg-speaker-listener',
config=cfg.LISTENER)
debug_print(logger, 'prepare training')
model = build_detection_model(cfg)
listener = build_listener(cfg)
speaker_listener = SpeakerListener(model,
listener,
cfg,
is_joint=cfg.LISTENER.JOINT)
if is_main_process():
wandb.watch(listener)
debug_print(logger, 'end model construction')
# modules that should be always set in eval mode
# their eval() method should be called after model.train() is called
eval_modules = (
model.rpn,
model.backbone,
model.roi_heads.box,
)
fix_eval_modules(eval_modules)
# NOTE, we slow down the LR of the layers start with the names in slow_heads
if cfg.MODEL.ROI_RELATION_HEAD.PREDICTOR == "IMPPredictor":
slow_heads = [
"roi_heads.relation.box_feature_extractor",
"roi_heads.relation.union_feature_extractor.feature_extractor",
]
else:
slow_heads = []
# load pretrain layers to new layers
load_mapping = {
"roi_heads.relation.box_feature_extractor":
"roi_heads.box.feature_extractor",
"roi_heads.relation.union_feature_extractor.feature_extractor":
"roi_heads.box.feature_extractor"
}
if cfg.MODEL.ATTRIBUTE_ON:
load_mapping[
"roi_heads.relation.att_feature_extractor"] = "roi_heads.attribute.feature_extractor"
load_mapping[
"roi_heads.relation.union_feature_extractor.att_feature_extractor"] = "roi_heads.attribute.feature_extractor"
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
listener.to(device)
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
num_batch = cfg.SOLVER.IMS_PER_BATCH
optimizer = make_optimizer(cfg,
model,
logger,
slow_heads=slow_heads,
slow_ratio=10.0,
rl_factor=float(num_batch))
listener_optimizer = make_listener_optimizer(cfg, listener)
scheduler = make_lr_scheduler(cfg, optimizer, logger)
listener_scheduler = None
debug_print(logger, 'end optimizer and schedule')
if cfg.LISTENER.JOINT:
speaker_listener_optimizer = make_speaker_listener_optimizer(
cfg, speaker_listener.speaker, speaker_listener.listener)
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
if cfg.LISTENER.JOINT:
speaker_listener, speaker_listener_optimizer = amp.initialize(
speaker_listener, speaker_listener_optimizer, opt_level='O0')
else:
speaker_listener, listener_optimizer = amp.initialize(
speaker_listener, listener_optimizer, opt_level='O0')
#listener, listener_optimizer = amp.initialize(listener, listener_optimizer, opt_level='O0')
#[model, listener], [optimizer, listener_optimizer] = amp.initialize([model, listener], [optimizer, listener_optimizer], opt_level='O1', loss_scale=1)
#model = amp.initialize(model, opt_level='O1')
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
find_unused_parameters=True,
)
listener = torch.nn.parallel.DistributedDataParallel(
listener,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
find_unused_parameters=True,
)
debug_print(logger, 'end distributed')
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
listener_dir = cfg.LISTENER_DIR
save_to_disk = get_rank() == 0
speaker_checkpointer = DetectronCheckpointer(cfg,
model,
optimizer,
scheduler,
output_dir,
save_to_disk,
custom_scheduler=True)
listener_checkpointer = Checkpointer(listener,
optimizer=listener_optimizer,
save_dir=listener_dir,
save_to_disk=save_to_disk,
custom_scheduler=False)
speaker_listener.add_listener_checkpointer(listener_checkpointer)
speaker_listener.add_speaker_checkpointer(speaker_checkpointer)
speaker_listener.load_listener()
speaker_listener.load_speaker(load_mapping=load_mapping)
debug_print(logger, 'end load checkpointer')
train_data_loader = make_data_loader(cfg,
mode='train',
is_distributed=distributed,
start_iter=arguments["iteration"],
ret_images=True)
val_data_loaders = make_data_loader(cfg,
mode='val',
is_distributed=distributed,
ret_images=True)
debug_print(logger, 'end dataloader')
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
if cfg.SOLVER.PRE_VAL:
logger.info("Validate before training")
#output = run_val(cfg, model, listener, val_data_loaders, distributed, logger)
#print('OUTPUT: ', output)
#(sg_loss, img_loss, sg_acc, img_acc) = output
logger.info("Start training")
meters = MetricLogger(delimiter=" ")
max_iter = len(train_data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
print_first_grad = True
listener_loss_func = torch.nn.MarginRankingLoss(margin=1, reduction='none')
mistake_saver = None
if is_main_process():
ds_catalog = DatasetCatalog()
dict_file_path = os.path.join(
ds_catalog.DATA_DIR,
ds_catalog.DATASETS['VG_stanford_filtered_with_attribute']
['dict_file'])
ind_to_classes, ind_to_predicates = load_vg_info(dict_file_path)
ind_to_classes = {k: v for k, v in enumerate(ind_to_classes)}
ind_to_predicates = {k: v for k, v in enumerate(ind_to_predicates)}
print('ind to classes:', ind_to_classes, '/n ind to predicates:',
ind_to_predicates)
mistake_saver = MistakeSaver(
'/Scene-Graph-Benchmark.pytorch/filenames_masked', ind_to_classes,
ind_to_predicates)
#is_printed = False
while True:
try:
listener_iteration = 0
for iteration, (images, targets,
image_ids) in enumerate(train_data_loader,
start_iter):
if cfg.LISTENER.JOINT:
speaker_listener_optimizer.zero_grad()
else:
listener_optimizer.zero_grad()
#print(f'ITERATION NUMBER: {iteration}')
if any(len(target) < 1 for target in targets):
logger.error(
f"Iteration={iteration + 1} || Image Ids used for training {_} || targets Length={[len(target) for target in targets]}"
)
if len(images) <= 1:
continue
data_time = time.time() - end
iteration = iteration + 1
listener_iteration += 1
arguments["iteration"] = iteration
model.train()
fix_eval_modules(eval_modules)
images_list = deepcopy(images)
images_list = to_image_list(
images_list, cfg.DATALOADER.SIZE_DIVISIBILITY).to(device)
for i in range(len(images)):
images[i] = images[i].unsqueeze(0)
images[i] = F.interpolate(images[i],
size=(224, 224),
mode='bilinear',
align_corners=False)
images[i] = images[i].squeeze()
images = torch.stack(images).to(device)
#images.requires_grad_()
targets = [target.to(device) for target in targets]
speaker_loss_dict = {}
if not cfg.LISTENER.JOINT:
score_matrix = speaker_listener(images_list, targets,
images)
else:
score_matrix, _, speaker_loss_dict = speaker_listener(
images_list, targets, images)
speaker_summed_losses = sum(
loss for loss in speaker_loss_dict.values())
# reduce losses over all GPUs for logging purposes
if not not cfg.LISTENER.JOINT:
speaker_loss_dict_reduced = reduce_loss_dict(
speaker_loss_dict)
speaker_losses_reduced = sum(
loss for loss in speaker_loss_dict_reduced.values())
speaker_losses_reduced /= num_gpus
if is_main_process():
wandb.log(
{"Train Speaker Loss": speaker_losses_reduced},
listener_iteration)
listener_loss = 0
gap_reward = 0
avg_acc = 0
num_correct = 0
score_matrix = score_matrix.to(device)
# fill loss matrix
loss_matrix = torch.zeros((2, images.size(0), images.size(0)),
device=device)
# sg centered scores
for true_index in range(loss_matrix.size(1)):
row_score = score_matrix[true_index]
(true_scores, predicted_scores,
binary) = format_scores(row_score, true_index, device)
loss_vec = listener_loss_func(true_scores,
predicted_scores, binary)
loss_matrix[0][true_index] = loss_vec
# image centered scores
transposted_score_matrix = score_matrix.t()
for true_index in range(loss_matrix.size(1)):
row_score = transposted_score_matrix[true_index]
(true_scores, predicted_scores,
binary) = format_scores(row_score, true_index, device)
loss_vec = listener_loss_func(true_scores,
predicted_scores, binary)
loss_matrix[1][true_index] = loss_vec
print('iteration:', listener_iteration)
sg_acc = 0
img_acc = 0
# calculate accuracy
for i in range(loss_matrix.size(1)):
temp_sg_acc = 0
temp_img_acc = 0
for j in range(loss_matrix.size(2)):
if loss_matrix[0][i][i] > loss_matrix[0][i][j]:
temp_sg_acc += 1
else:
if cfg.LISTENER.HTML:
if is_main_process(
) and listener_iteration >= 600 and listener_iteration % 25 == 0 and i != j:
detached_sg_i = (sgs[i][0].detach(),
sgs[i][1],
sgs[i][2].detach())
detached_sg_j = (sgs[j][0].detach(),
sgs[j][1],
sgs[j][2].detach())
mistake_saver.add_mistake(
(image_ids[i], image_ids[j]),
(detached_sg_i, detached_sg_j),
listener_iteration, 'SG')
if loss_matrix[1][i][i] > loss_matrix[1][j][i]:
temp_img_acc += 1
else:
if cfg.LISTENER.HTML:
if is_main_process(
) and listener_iteration >= 600 and listener_iteration % 25 == 0 and i != j:
detached_sg_i = (sgs[i][0].detach(),
sgs[i][1],
sgs[i][2].detach())
detached_sg_j = (sgs[j][0].detach(),
sgs[j][1],
sgs[j][2].detach())
mistake_saver.add_mistake(
(image_ids[i], image_ids[j]),
(detached_sg_i, detached_sg_j),
listener_iteration, 'IMG')
temp_sg_acc = temp_sg_acc * 100 / (loss_matrix.size(1) - 1)
temp_img_acc = temp_img_acc * 100 / (loss_matrix.size(1) -
1)
sg_acc += temp_sg_acc
img_acc += temp_img_acc
if cfg.LISTENER.HTML:
if is_main_process(
) and listener_iteration % 100 == 0 and listener_iteration >= 600:
mistake_saver.toHtml('/www')
sg_acc /= loss_matrix.size(1)
img_acc /= loss_matrix.size(1)
avg_sg_acc = torch.tensor([sg_acc]).to(device)
avg_img_acc = torch.tensor([img_acc]).to(device)
# reduce acc over all gpus
avg_acc = {'sg_acc': avg_sg_acc, 'img_acc': avg_img_acc}
avg_acc_reduced = reduce_loss_dict(avg_acc)
sg_acc = sum(acc for acc in avg_acc_reduced['sg_acc'])
img_acc = sum(acc for acc in avg_acc_reduced['img_acc'])
# log acc to wadb
if is_main_process():
wandb.log({
"Train SG Accuracy": sg_acc.item(),
"Train IMG Accuracy": img_acc.item()
})
sg_loss = 0
img_loss = 0
for i in range(loss_matrix.size(0)):
for j in range(loss_matrix.size(1)):
loss_matrix[i][j][j] = 0.
for i in range(loss_matrix.size(1)):
sg_loss += torch.max(loss_matrix[0][i])
img_loss += torch.max(loss_matrix[1][:][i])
sg_loss = sg_loss / loss_matrix.size(1)
img_loss = img_loss / loss_matrix.size(1)
sg_loss = sg_loss.to(device)
img_loss = img_loss.to(device)
loss_dict = {'sg_loss': sg_loss, 'img_loss': img_loss}
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
sg_loss_reduced = loss_dict_reduced['sg_loss']
img_loss_reduced = loss_dict_reduced['img_loss']
if is_main_process():
wandb.log({"Train SG Loss": sg_loss_reduced})
wandb.log({"Train IMG Loss": img_loss_reduced})
losses_reduced = sum(loss
for loss in loss_dict_reduced.values())
meters.update(loss=losses_reduced, **loss_dict_reduced)
losses = losses + speaker_summed_losses * cfg.LISTENER.LOSS_COEF
# Note: If mixed precision is not used, this ends up doing nothing
# Otherwise apply loss scaling for mixed-precision recipe
#losses.backward()
if not cfg.LISTENER.JOINT:
with amp.scale_loss(losses,
listener_optimizer) as scaled_losses:
scaled_losses.backward()
else:
with amp.scale_loss(
losses,
speaker_listener_optimizer) as scaled_losses:
scaled_losses.backward()
verbose = (iteration % cfg.SOLVER.PRINT_GRAD_FREQ
) == 0 or print_first_grad # print grad or not
print_first_grad = False
#clip_grad_value([(n, p) for n, p in listener.named_parameters() if p.requires_grad], cfg.LISTENER.CLIP_VALUE, logger=logger, verbose=True, clip=True)
if not cfg.LISTENER.JOINT:
listener_optimizer.step()
else:
speaker_listener_optimizer.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time, data=data_time)
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if cfg.LISTENER.JOINT:
if iteration % 200 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join([
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters),
lr=speaker_listener_optimizer.param_groups[-1]
["lr"],
memory=torch.cuda.max_memory_allocated() /
1024.0 / 1024.0,
))
else:
if iteration % 200 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join([
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters),
lr=listener_optimizer.param_groups[-1]["lr"],
memory=torch.cuda.max_memory_allocated() /
1024.0 / 1024.0,
))
if iteration % checkpoint_period == 0:
"""
print('Model before save')
print('****************************')
print(listener.gnn.conv1.node_model.node_mlp_1[0].weight)
print('****************************')
"""
if not cfg.LISTENER.JOINT:
listener_checkpointer.save(
"model_{:07d}".format(listener_iteration),
amp=amp.state_dict())
else:
speaker_checkpointer.save(
"model_speaker{:07d}".format(iteration))
listener_checkpointer.save(
"model_listenr{:07d}".format(listener_iteration),
amp=amp.state_dict())
if iteration == max_iter:
if not cfg.LISTENER.JOINT:
listener_checkpointer.save(
"model_{:07d}".format(listener_iteration),
amp=amp.state_dict())
else:
speaker_checkpointer.save(
"model_{:07d}".format(iteration))
listener_checkpointer.save(
"model_{:07d}".format(listener_iteration),
amp=amp.state_dict())
val_result = None # used for scheduler updating
if cfg.SOLVER.TO_VAL and iteration % cfg.SOLVER.VAL_PERIOD == 0:
logger.info("Start validating")
val_result = run_val(cfg, model, listener,
val_data_loaders, distributed, logger)
(sg_loss, img_loss, sg_acc, img_acc,
speaker_val) = val_result
if is_main_process():
wandb.log({
"Validation SG Accuracy": sg_acc,
"Validation IMG Accuracy": img_acc,
"Validation SG Loss": sg_loss,
"Validation IMG Loss": img_loss,
"Validation Speaker": speaker_val,
})
#logger.info("Validation Result: %.4f" % val_result)
except Exception as err:
raise (err)
print('Dataset finished, creating new')
train_data_loader = make_data_loader(
cfg,
mode='train',
is_distributed=distributed,
start_iter=arguments["iteration"],
ret_images=True)
total_training_time = time.time() - start_training_time
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / (max_iter)))
return listener
def do_train(
cfg,
model,
data_loader,
data_loader_val,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
test_period,
arguments,
writer,
):
logger = logging.getLogger("maskrcnn_benchmark.trainer")
logger.info("Start training")
meters = MetricLogger(delimiter=" ")
max_iter = len(data_loader)
start_iter = arguments["iteration"]
model.train()
start_training_time = time.time()
end = time.time()
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
dataset_names = cfg.DATASETS.TEST
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
if any(len(target) < 1 for target in targets):
logger.error(
f"Iteration={iteration + 1} || Image Ids used for training {_} || targets Length={[len(target) for target in targets]}"
)
continue
data_time = time.time() - end
iteration = iteration + 1
arguments["iteration"] = iteration
images = images.to(device)
targets = [target.to(device) for target in targets]
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
# # Add images every 100 iterations
if iteration % 100 == 0:
# # Display images
# image = images.tensors[0].cpu().numpy()
# means = np.zeros((image.shape[0], image.shape[1], image.shape[2]))
# means[0] = 102.9801
# means[1] = 115.9465
# means[2] = 122.7717
# image = image + means
# image = image[[2, 1, 0]].astype(np.uint8)
# writer.add_image('input image', image, iteration)
# for b in range(len(targets[0].bbox)):
# box = targets[0].bbox[b]
# x1 = np.around(box[0].cpu().numpy())
# y1 = np.around(box[1].cpu().numpy())
# x2 = np.around(box[2].cpu().numpy())
# y2 = np.around(box[3].cpu().numpy())
# rr, cc = rectangle_perimeter(y1, x1, y2-y1, x2-x1)
# image[:, rr, cc] = 255
# writer.add_image('target boxes', image, iteration)
# # Display masks
# masks = targets[0].get_field('masks')[0]
# masks = masks.get_mask_tensor()
# combined_mask = masks[0, :, :]
# for i in range(1,8):
# combined_mask = combined_mask | masks[i, :, :]
# writer.add_image('mask', combined_mask.unsqueeze(0)*255, iteration)
# writer.add_image('single part 2', masks[1, :, :].unsqueeze(0)*255, iteration)
# writer.add_image('single part 3', masks[2, :, :].unsqueeze(0)*255, iteration)
# writer.add_image('single part 4', masks[3, :, :].unsqueeze(0)*255, iteration)
# writer.add_image('single part 5', masks[4, :, :].unsqueeze(0)*255, iteration)
# writer.add_image('single part 6', masks[5, :, :].unsqueeze(0)*255, iteration)
# writer.add_image('single part 7', masks[6, :, :].unsqueeze(0)*255, iteration)
# writer.add_image('single part 8', masks[7, :, :].unsqueeze(0)*255, iteration)
# Display Losses
writer.add_scalar('loss', meters.loss.median, iteration)
writer.add_scalar('loss_classifier',
loss_dict_reduced['loss_classifier'].item(),
iteration)
writer.add_scalar('loss_box_reg',
loss_dict_reduced['loss_box_reg'].item(),
iteration)
writer.add_scalar('loss_objectness',
loss_dict_reduced['loss_objectness'].item(),
iteration)
writer.add_scalar('loss_rpn_box_reg',
loss_dict_reduced['loss_rpn_box_reg'].item(),
iteration)
writer.add_scalar('loss_mask',
loss_dict_reduced['loss_mask'].item(), iteration)
writer.add_scalar('loss_kpt', loss_dict_reduced['loss_kp'].item(),
iteration)
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], iteration)
# Note: If mixed precision is not used, this ends up doing nothing
# Otherwise apply loss scaling for mixed-precision recipe
with amp.scale_loss(losses, optimizer) as scaled_losses:
scaled_losses.backward()
optimizer.step()
scheduler.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time, data=data_time)
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if iteration % 20 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join([
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters),
lr=optimizer.param_groups[0]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
))
if iteration % checkpoint_period == 0:
checkpointer.save("model_{:07d}".format(iteration), **arguments)
if data_loader_val is not None and test_period > 0 and iteration % test_period == 0:
meters_val = MetricLogger(delimiter=" ")
synchronize()
_ = inference( # The result can be used for additional logging, e. g. for TensorBoard
model,
# The method changes the segmentation mask format in a data loader,
# so every time a new data loader is created:
make_data_loader(cfg,
is_train=False,
is_distributed=(get_world_size() > 1),
is_for_period=True),
dataset_name="[Validation]",
iou_types=iou_types,
box_only=False
if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=None,
)
synchronize()
model.train()
with torch.no_grad():
# Should be one image for each GPU:
for iteration_val, (images_val, targets_val,
_) in enumerate(tqdm(data_loader_val)):
images_val = images_val.to(device)
targets_val = [target.to(device) for target in targets_val]
loss_dict = model(images_val, targets_val)
losses = sum(loss for loss in loss_dict.values())
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(
loss for loss in loss_dict_reduced.values())
meters_val.update(loss=losses_reduced, **loss_dict_reduced)
synchronize()
logger.info(
meters_val.delimiter.join([
"[Validation]: ",
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters_val),
lr=optimizer.param_groups[0]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
))
if iteration == max_iter:
checkpointer.save("model_final", **arguments)
total_training_time = time.time() - start_training_time
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / (max_iter)))
def train(cfg, local_rank, distributed):
# Model logging
print_mlperf(key=mlperf_log.INPUT_BATCH_SIZE, value=cfg.SOLVER.IMS_PER_BATCH)
print_mlperf(key=mlperf_log.BATCH_SIZE_TEST, value=cfg.TEST.IMS_PER_BATCH)
print_mlperf(key=mlperf_log.INPUT_MEAN_SUBTRACTION, value = cfg.INPUT.PIXEL_MEAN)
print_mlperf(key=mlperf_log.INPUT_NORMALIZATION_STD, value=cfg.INPUT.PIXEL_STD)
print_mlperf(key=mlperf_log.INPUT_RESIZE)
print_mlperf(key=mlperf_log.INPUT_RESIZE_ASPECT_PRESERVING)
print_mlperf(key=mlperf_log.MIN_IMAGE_SIZE, value=cfg.INPUT.MIN_SIZE_TRAIN)
print_mlperf(key=mlperf_log.MAX_IMAGE_SIZE, value=cfg.INPUT.MAX_SIZE_TRAIN)
print_mlperf(key=mlperf_log.INPUT_RANDOM_FLIP)
print_mlperf(key=mlperf_log.RANDOM_FLIP_PROBABILITY, value=0.5)
print_mlperf(key=mlperf_log.FG_IOU_THRESHOLD, value=cfg.MODEL.RPN.FG_IOU_THRESHOLD)
print_mlperf(key=mlperf_log.BG_IOU_THRESHOLD, value=cfg.MODEL.RPN.BG_IOU_THRESHOLD)
print_mlperf(key=mlperf_log.RPN_PRE_NMS_TOP_N_TRAIN, value=cfg.MODEL.RPN.PRE_NMS_TOP_N_TRAIN)
print_mlperf(key=mlperf_log.RPN_PRE_NMS_TOP_N_TEST, value=cfg.MODEL.RPN.PRE_NMS_TOP_N_TEST)
print_mlperf(key=mlperf_log.RPN_POST_NMS_TOP_N_TRAIN, value=cfg.MODEL.RPN.FPN_POST_NMS_TOP_N_TRAIN)
print_mlperf(key=mlperf_log.RPN_POST_NMS_TOP_N_TEST, value=cfg.MODEL.RPN.FPN_POST_NMS_TOP_N_TEST)
print_mlperf(key=mlperf_log.ASPECT_RATIOS, value=cfg.MODEL.RPN.ASPECT_RATIOS)
print_mlperf(key=mlperf_log.BACKBONE, value=cfg.MODEL.BACKBONE.CONV_BODY)
print_mlperf(key=mlperf_log.NMS_THRESHOLD, value=cfg.MODEL.RPN.NMS_THRESH)
# /root/ssy/ssynew/maskrcnn-benchmark/maskrcnn_benchmark/modeling/detector/detectors.py
# building bare mode without doing anthing
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
# Optimizer logging
print_mlperf(key=mlperf_log.OPT_NAME, value=mlperf_log.SGD_WITH_MOMENTUM)
print_mlperf(key=mlperf_log.OPT_LR, value=cfg.SOLVER.BASE_LR)
print_mlperf(key=mlperf_log.OPT_MOMENTUM, value=cfg.SOLVER.MOMENTUM)
print_mlperf(key=mlperf_log.OPT_WEIGHT_DECAY, value=cfg.SOLVER.WEIGHT_DECAY)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
print("output_dir "+str(output_dir))
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
# no such SAVE_CHECKPOINTS
#arguments["save_checkpoints"] = cfg.SAVE_CHECKPOINTS
arguments["save_checkpoints"] = False
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader, iters_per_epoch = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"]
)
print("SSY iters_per_epoch "+str(iters_per_epoch))
#print("SSY iters_per_epoch change to 100 ")
#iters_per_epoch = 100
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# set the callback function to evaluate and potentially
# early exit each epoch
# SSY
# I already add PER_EPOCH_EVAL and MIN_BBOX_MAP MIN_SEGM_MAP to ./configs/e2e_mask_rcnn_R_50_FPN_1x.yaml
# but it still can not find it
# so I manually set them here
#if cfg.PER_EPOCH_EVAL:
# per_iter_callback_fn = functools.partial(
# mlperf_test_early_exit,
# iters_per_epoch=iters_per_epoch,
# tester=functools.partial(test, cfg=cfg),
# model=model,
# distributed=distributed,
# min_bbox_map=cfg.MLPERF.MIN_BBOX_MAP,
# min_segm_map=cfg.MLPERF.MIN_SEGM_MAP)
#else:
# per_iter_callback_fn = None
per_iter_callback_fn = functools.partial(
mlperf_test_early_exit,
iters_per_epoch=iters_per_epoch,
# /root/ssy/ssynew/maskrcnn-benchmark/maskrcnn_benchmark/engine/tester.py
tester=functools.partial(test, cfg=cfg),
model=model,
distributed=distributed,
min_bbox_map=0.377,
min_segm_map=0.339)
start_train_time = time.time()
# /root/ssy/ssynew/maskrcnn-benchmark/maskrcnn_benchmark/engine/trainer.py
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
per_iter_start_callback_fn=functools.partial(mlperf_log_epoch_start, iters_per_epoch=iters_per_epoch),
per_iter_end_callback_fn=per_iter_callback_fn,
)
end_train_time = time.time()
total_training_time = end_train_time - start_train_time
print(
"&&&& MLPERF METRIC THROUGHPUT per GPU={:.4f} iterations / s".format((arguments["iteration"] * 1.0) / total_training_time)
)
return model
def train(cfg, local_rank, distributed, logger):
debug_print(logger, 'prepare training')
model = build_detection_model(cfg)
debug_print(logger, 'end model construction')
# modules that should be always set in eval mode
# their eval() method should be called after model.train() is called
eval_modules = (model.rpn, model.backbone, model.roi_heads.box,)
fix_eval_modules(eval_modules)
# NOTE, we slow down the LR of the layers start with the names in slow_heads
if cfg.MODEL.ROI_RELATION_HEAD.PREDICTOR == "IMPPredictor":
slow_heads = ["roi_heads.relation.box_feature_extractor",
"roi_heads.relation.union_feature_extractor.feature_extractor",]
else:
slow_heads = []
# load pretrain layers to new layers
load_mapping = {"roi_heads.relation.box_feature_extractor" : "roi_heads.box.feature_extractor",
"roi_heads.relation.union_feature_extractor.feature_extractor" : "roi_heads.box.feature_extractor"}
if cfg.MODEL.ATTRIBUTE_ON:
load_mapping["roi_heads.relation.att_feature_extractor"] = "roi_heads.attribute.feature_extractor"
load_mapping["roi_heads.relation.union_feature_extractor.att_feature_extractor"] = "roi_heads.attribute.feature_extractor"
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
num_batch = cfg.SOLVER.IMS_PER_BATCH
optimizer = make_optimizer(cfg, model, logger, slow_heads=slow_heads, slow_ratio=10.0, rl_factor=float(num_batch))
scheduler = make_lr_scheduler(cfg, optimizer, logger)
debug_print(logger, 'end optimizer and shcedule')
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model, optimizer, opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
find_unused_parameters=True,
)
debug_print(logger, 'end distributed')
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk, custom_scheduler=True
)
# if there is certain checkpoint in output_dir, load it, else load pretrained detector
if checkpointer.has_checkpoint():
extra_checkpoint_data = checkpointer.load(cfg.MODEL.PRETRAINED_DETECTOR_CKPT,
update_schedule=cfg.SOLVER.UPDATE_SCHEDULE_DURING_LOAD)
arguments.update(extra_checkpoint_data)
if cfg.SOLVER.UPDATE_SCHEDULE_DURING_LOAD:
checkpointer.scheduler.last_epoch = extra_checkpoint_data["iteration"]
logger.info("update last epoch of scheduler to iter: {}".format(str(extra_checkpoint_data["iteration"])))
else:
# load_mapping is only used when we init current model from detection model.
checkpointer.load(cfg.MODEL.PRETRAINED_DETECTOR_CKPT, with_optim=False, load_mapping=load_mapping)
debug_print(logger, 'end load checkpointer')
train_data_loader = make_data_loader(
cfg,
mode='train',
is_distributed=distributed,
start_iter=arguments["iteration"],
)
val_data_loaders = make_data_loader(
cfg,
mode='val',
is_distributed=distributed,
)
debug_print(logger, 'end dataloader')
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
if cfg.SOLVER.PRE_VAL:
logger.info("Validate before training")
run_val(cfg, model, val_data_loaders, distributed, logger)
logger.info("Start training")
meters = MetricLogger(delimiter=" ")
max_iter = len(train_data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
print_first_grad = True
for iteration, (images, targets, _) in enumerate(train_data_loader, start_iter):
if any(len(target) < 1 for target in targets):
logger.error(f"Iteration={iteration + 1} || Image Ids used for training {_} || targets Length={[len(target) for target in targets]}" )
data_time = time.time() - end
iteration = iteration + 1
arguments["iteration"] = iteration
model.train()
fix_eval_modules(eval_modules)
images = images.to(device)
targets = [target.to(device) for target in targets]
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
# Note: If mixed precision is not used, this ends up doing nothing
# Otherwise apply loss scaling for mixed-precision recipe
with amp.scale_loss(losses, optimizer) as scaled_losses:
scaled_losses.backward()
# add clip_grad_norm from MOTIFS, tracking gradient, used for debug
verbose = (iteration % cfg.SOLVER.PRINT_GRAD_FREQ) == 0 or print_first_grad # print grad or not
print_first_grad = False
clip_grad_norm([(n, p) for n, p in model.named_parameters() if p.requires_grad], max_norm=cfg.SOLVER.GRAD_NORM_CLIP, logger=logger, verbose=verbose, clip=True)
optimizer.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time, data=data_time)
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if iteration % 200 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join(
[
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]
).format(
eta=eta_string,
iter=iteration,
meters=str(meters),
lr=optimizer.param_groups[-1]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
)
)
if iteration % checkpoint_period == 0:
checkpointer.save("model_{:07d}".format(iteration), **arguments)
if iteration == max_iter:
checkpointer.save("model_final", **arguments)
val_result = None # used for scheduler updating
if cfg.SOLVER.TO_VAL and iteration % cfg.SOLVER.VAL_PERIOD == 0:
logger.info("Start validating")
val_result = run_val(cfg, model, val_data_loaders, distributed, logger)
logger.info("Validation Result: %.4f" % val_result)
# scheduler should be called after optimizer.step() in pytorch>=1.1.0
# https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
if cfg.SOLVER.SCHEDULE.TYPE == "WarmupReduceLROnPlateau":
scheduler.step(val_result, epoch=iteration)
if scheduler.stage_count >= cfg.SOLVER.SCHEDULE.MAX_DECAY_STEP:
logger.info("Trigger MAX_DECAY_STEP at iteration {}.".format(iteration))
break
else:
scheduler.step()
total_training_time = time.time() - start_training_time
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info(
"Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / (max_iter)
)
)
return model
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model,
optimizer,
opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
arguments['phase'] = 1
arguments['plot_median'], arguments['plot_global_avg'] = defaultdict(
list), defaultdict(list)
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg, model, optimizer, scheduler,
output_dir, save_to_disk)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
test_period = cfg.SOLVER.TEST_PERIOD
if test_period > 0:
data_loader_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed,
is_for_period=True)
else:
data_loader_val = None
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
if arguments['phase'] == 1:
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
phase=1,
)
do_train(
cfg,
model,
data_loader,
data_loader_val,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
test_period,
arguments,
training_phase=1,
)
arguments["iteration"] = 0
arguments["phase"] = 2
data_loader_phase2 = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
phase=2,
)
do_train(
cfg,
model,
data_loader_phase2,
data_loader_val,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
test_period,
arguments,
training_phase=2,
)
return model
def pred_with_weight(args):
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.deprecated.init_process_group(backend="nccl",
init_method="env://")
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
for weight in args.weights:
# skipping evaluations already performed
out_json = "{}/detections/{}.json".format(
cfg.OUTPUT_DIR,
weight.split('/')[-1].split('_')[-1].split('.')[0])
if os.path.exists(out_json):
print('skipping', out_json)
continue
checkpointer = DetectronCheckpointer(cfg, model)
_ = checkpointer.load(weight)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
output_folders = [None] * len(cfg.DATASETS.TEST)
if cfg.OUTPUT_DIR:
dataset_names = cfg.DATASETS.TEST
for idx, dataset_name in enumerate(dataset_names):
output_folder = cfg.OUTPUT_DIR
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, data_loader_val in zip(output_folders,
data_loaders_val):
inference(
model,
data_loader_val,
iou_types=iou_types,
box_only=cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
name=weight.split('_')[-1].split('.')[0])
synchronize()
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
summary_writer = SummaryWriter(log_dir=output_dir)
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
if cfg.MODEL.WEIGHT.upper() == 'CONTINUE':
model_weight = last_checkpoint(output_dir)
else:
model_weight = cfg.MODEL.WEIGHT
extra_checkpoint_data = checkpointer.load(model_weight)
arguments.update(extra_checkpoint_data)
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
data_loader_val = make_data_loader(
cfg,
is_train=False,
is_distributed=distributed)[0]
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
do_train(
model=model,
data_loader=data_loader,
data_loader_val=data_loader_val,
optimizer=optimizer,
scheduler=scheduler,
checkpointer=checkpointer,
device=device,
checkpoint_period=checkpoint_period,
arguments=arguments,
summary_writer=summary_writer
)
return model
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox",)
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm",)
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg, is_train=False, is_distributed=distributed)
results = []
for output_folder, dataset_name, data_loader_val in zip(output_folders, dataset_names, data_loaders_val):
result = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
skip_eval=args.skip_eval,
dllogger=dllogger,
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"--data-dir",
default=".",
metavar="DIR",
help="data dir for training",
type=str,
)
parser.add_argument(
"--out-dir",
default=".",
metavar="DIR",
help="output dir for model",
type=str,
)
parser.add_argument(
"--gpu_ids",
default="-1",
help="gpu id",
type=str,
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
if args.gpu_ids != '-1':
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_ids
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.MODEL.WEIGHT = args.data_dir + cfg.MODEL.WEIGHT[1:]
cfg.DATA_DIR = args.data_dir + cfg.DATA_DIR[1:]
cfg.OUTPUT_DIR = args.out_dir + cfg.OUTPUT_DIR[1:]
print("cfg.OUTPUT_DIR: ", cfg.OUTPUT_DIR)
print("cfg.MODEL.WEIGHT: ", cfg.MODEL.WEIGHT)
print("cfg.DATA_DIR: ", cfg.DATA_DIR)
print("cfg.MODEL.ATTRIBUTE_ON: ", cfg.MODEL.ATTRIBUTE_ON)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
# evaluate object detection
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
result_obj = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
eval_attributes=False,
)
synchronize()
# evaluate attribute detection
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
result_attr = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
eval_attributes=True,
)
synchronize()
# evaluate RPN
cfg.defrost()
cfg.MODEL.RPN_ONLY = True
cfg.freeze()
logger.info(cfg)
# pdb.set_trace()
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
result_rpn = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
eval_attributes=False,
)
synchronize()
if is_main_process():
results = {**result_rpn, **result_obj, **result_attr}
print(results)
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
params = get_model_parameters_number(model)
print('{:<30} {:<8}'.format('Number of parameters: ', params))
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
if cfg.MODEL.DOMAIN_ADAPTATION_ON:
source_data_loader = make_data_loader(
cfg,
is_train=True,
is_source=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
target_data_loader = make_data_loader(
cfg,
is_train=True,
is_source=False,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
do_da_train(
model,
source_data_loader,
target_data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
cfg,
)
else:
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
)
return model
def main():
parser = argparse.ArgumentParser(description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default="/home/guli/Desktop/VOS_ICCV2019/maskrcnn-benchmark/configs/davis/e2e_mask_rcnn_R_50_FPN_1x_davis.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.deprecated.init_process_group(
backend="nccl", init_method="env://"
)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("DAVIS_MaskRCNN_baseline_test", save_dir, args.local_rank)
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
checkpointer = Checkpointer(model)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox",)
# if cfg.MODEL.MASK_ON:
# iou_types = iou_types + ("segm",)
output_folders = [None] * len(cfg.DATASETS.TEST)
if cfg.OUTPUT_DIR:
dataset_names = cfg.DATASETS.TEST
exp_name = cfg.EXP.NAME
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name + "_" + exp_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg, is_train=False, is_distributed=distributed)
for output_folder, data_loader_val in zip(output_folders, data_loaders_val):
inference_davis(
model,
data_loader_val,
iou_types=iou_types,
box_only=False,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
debug=cfg.TEST.DEBUG,
generate_annotation=cfg.TEST.GENERATE_ANNOTATION,
overlay_box=cfg.TEST.OVERLAY_BOX,
matching=cfg.TEST.MATCHING,
skip_computation_network=cfg.TEST.SKIP_NETWORK,
select_top_predictions_flag=cfg.TEST.SELECT_TOP_PREDICTIONS,
cfg=cfg
)
synchronize()
def train(cfg, local_rank, distributed, logger):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg,
model,
logger,
rl_factor=float(cfg.SOLVER.IMS_PER_BATCH))
scheduler = make_lr_scheduler(cfg, optimizer)
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model,
optimizer,
opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg, model, optimizer, scheduler,
output_dir, save_to_disk)
extra_checkpoint_data = checkpointer.load(
cfg.MODEL.WEIGHT,
update_schedule=cfg.SOLVER.UPDATE_SCHEDULE_DURING_LOAD)
arguments.update(extra_checkpoint_data)
train_data_loader = make_data_loader(
cfg,
mode='train',
is_distributed=distributed,
start_iter=arguments["iteration"],
)
val_data_loaders = make_data_loader(
cfg,
mode='val',
is_distributed=distributed,
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
if cfg.SOLVER.PRE_VAL:
logger.info("Validate before training")
run_val(cfg, model, val_data_loaders, distributed)
logger.info("Start training")
meters = MetricLogger(delimiter=" ")
max_iter = len(train_data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
for iteration, (images, targets, _) in enumerate(train_data_loader,
start_iter):
model.train()
if any(len(target) < 1 for target in targets):
logger.error(
f"Iteration={iteration + 1} || Image Ids used for training {_} || targets Length={[len(target) for target in targets]}"
)
data_time = time.time() - end
iteration = iteration + 1
arguments["iteration"] = iteration
scheduler.step()
images = images.to(device)
targets = [target.to(device) for target in targets]
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
# Note: If mixed precision is not used, this ends up doing nothing
# Otherwise apply loss scaling for mixed-precision recipe
with amp.scale_loss(losses, optimizer) as scaled_losses:
scaled_losses.backward()
optimizer.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time, data=data_time)
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if iteration % 200 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join([
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters),
lr=optimizer.param_groups[0]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
))
if cfg.SOLVER.TO_VAL and iteration % cfg.SOLVER.VAL_PERIOD == 0:
logger.info("Start validating")
run_val(cfg, model, val_data_loaders, distributed)
if iteration % checkpoint_period == 0:
checkpointer.save("model_{:07d}".format(iteration), **arguments)
if iteration == max_iter:
checkpointer.save("model_final", **arguments)
total_training_time = time.time() - start_training_time
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / (max_iter)))
return model
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
# if use_amp:
# # Initialize mixed-precision training
# use_mixed_precision = cfg.DTYPE == "float16"
# amp_handle = amp.init(enabled=use_mixed_precision, verbose=cfg.AMP_VERBOSE)
# # wrap the optimizer for mixed precision
# if cfg.SOLVER.ACCUMULATE_GRAD:
# # also specify number of steps to accumulate over
# optimizer = amp_handle.wrap_optimizer(optimizer, num_loss=cfg.SOLVER.ACCUMULATE_STEPS)
# else:
# optimizer = amp_handle.wrap_optimizer(optimizer)
model, optimizer = amp.initialize(model, optimizer,opt_level='O1')
if distributed:
if use_apex_ddp:
model = DDP(model, delay_allreduce=True)
else:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader, iters_per_epoch = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# set the callback function to evaluate and potentially
# early exit each epoch
if 1==1:
per_iter_callback_fn = functools.partial(
mlperf_test_early_exit,
iters_per_epoch=iters_per_epoch,
tester=functools.partial(test, cfg=cfg),
model=model,
distributed=distributed,
min_bbox_map=cfg.MIN_BBOX_MAP,
min_segm_map=cfg.MIN_MASK_MAP)
else:
per_iter_callback_fn = None
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
use_amp,
cfg,
per_iter_end_callback_fn=per_iter_callback_fn,
)
return model
def test_maskscoring_rcnn(config_file):
import sys
sys.path.append('./detection_model/maskscoring_rcnn')
import argparse
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '2'
import torch
#from maskrcnn_benchmark.config import cfg
from maskrcnn_benchmark.data import make_data_loader
from maskrcnn_benchmark.engine.inference import inference
from maskrcnn_benchmark.modeling.detector import build_detection_model
from maskrcnn_benchmark.utils.checkpoint import DetectronCheckpointer
from maskrcnn_benchmark.utils.collect_env import collect_env_info
from maskrcnn_benchmark.utils.comm import synchronize, get_rank
from maskrcnn_benchmark.utils.logger import setup_logger
from maskrcnn_benchmark.utils.miscellaneous import mkdir
from yacs.config import CfgNode as CN
def read_config_file(config_file):
"""
read config information form yaml file
"""
f = open(config_file)
opt = CN.load_cfg(f)
return opt
opt = read_config_file(config_file)
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(opt.local_rank)
torch.distributed.deprecated.init_process_group(
backend="nccl", init_method="env://"
)
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(opt)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(opt)
model.to(opt.MODEL.DEVICE)
output_dir = opt.OUTPUT_DIR
checkpointer = DetectronCheckpointer(opt, model, save_dir=output_dir)
_ = checkpointer.load(opt.MODEL.WEIGHT)
iou_types = ("bbox",)
if opt.MODEL.MASK_ON:
iou_types = iou_types + ("segm",)
output_folders = [None] * len(opt.DATASETS.TEST)
if opt.OUTPUT_DIR:
dataset_names = opt.DATASETS.TEST
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(opt.OUTPUT_DIR, "inference", dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(opt, is_train=False, is_distributed=distributed)
for output_folder, data_loader_val in zip(output_folders, data_loaders_val):
inference(
model,
data_loader_val,
iou_types=iou_types,
box_only=opt.MODEL.RPN_ONLY,
device=opt.MODEL.DEVICE,
expected_results=opt.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=opt.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
maskiou_on=opt.MODEL.MASKIOU_ON
)
synchronize()
def train(cfg, cfg_origial, local_rank, distributed):
## The one with modified number of classes
model = build_detection_model(cfg)
# cfg_origial = cfg.clone()
# cfg_origial.MODEL.ROI_BOX_HEAD.NUM_CLASSES = 81
# original_model = build_detection_model(cfg_origial) ## Original model with 81 classes
# ## Let's load weights for old class!
# save_dir = cfg.OUTPUT_DIR
# checkpointer = DetectronCheckpointer(cfg_origial, original_model, save_dir=save_dir)
# checkpointer.load(cfg_origial.MODEL.WEIGHT)
# # pretrained_model_pth = "/network/home/bhattdha/.torch/models/_detectron_35861795_12_2017_baselines_e2e_mask_rcnn_R-101-FPN_1x.yaml.02_31_37.KqyEK4tT_output_train_coco_2014_train%3Acoco_2014_valminusminival_generalized_rcnn_model_final.pkl"
# # These keys are to be removed which forms final layers of the network
# removal_keys = ['roi_heads.box.predictor.cls_score.weight', 'roi_heads.box.predictor.cls_score.bias', 'roi_heads.box.predictor.bbox_pred.weight', 'roi_heads.box.predictor.bbox_pred.bias', 'roi_heads.mask.predictor.mask_fcn_logits.weight', 'roi_heads.mask.predictor.mask_fcn_logits.bias']
# model = _transfer_pretrained_weights(new_model, original_model, removal_keys)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
# # Initialize mixed-precision training
# use_mixed_precision = cfg.DTYPE == "float16"
# amp_opt_level = 'O1' if use_mixed_precision else 'O0'
# model, optimizer = amp.initialize(model, optimizer, opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg, model, optimizer, scheduler,
output_dir, save_to_disk)
# cfg.MODEL.WEIGHT = '/network/home/bhattdha/exp.pth' ## Model stored through surgery
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
)
return model
def main():
parser = argparse.ArgumentParser(description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default="/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument("--seq_test", action='store_true')
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = cfg.OUTPUT_DIR
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank(), filename='test_log.txt')
logger.info("Using {} GPUs".format(num_gpus))
# logger.info(cfg)
logger.info("Collecting env info (might take some time)")
# logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
ori_output_dir = cfg.OUTPUT_DIR
if args.seq_test:
load_dir = cfg.TEST.LOAD_DIR
model_files = glob.glob(load_dir+'/*.pth')
model_files.sort()
min_iter = cfg.TEST.MIN_ITER
max_iter = cfg.TEST.MAX_ITER
# print(model_files)
model_files = [model_file for model_file in model_files if 'final' not in model_file and int(model_file[-11:-4])>=min_iter and int(model_file[-11:-4])<=max_iter]
else:
model_files = [cfg.MODEL.WEIGHT]
for model_file in model_files:
cfg.defrost()
cfg.MODEL.WEIGHT = model_file
logger.info('testing from {} '.format(model_file))
cfg.OUTPUT_DIR = os.path.join(ori_output_dir, model_file[-11:-4])
cfg.freeze()
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox",)
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm",)
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints",)
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg, is_train=False, is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False, #False if cfg.MODEL.FCOS_ON or cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
stop_iter=cfg.FEW_SHOT.STOP_ITER
)
synchronize()
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"--ckpt",
help=
"The path to the checkpoint for test, default is the latest checkpoint.",
default=None,
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
parser.add_argument(
"--build-model",
default="",
metavar="FILE",
help="path to NAS model build file",
type=str,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
assert len(args.build_model) != 0, 'args.build_model should be provided'
model_config = json.load(open(args.build_model, 'r'))
if isinstance(model_config, list):
assert len(model_config) == 1
model_config = model_config[0]
print('Testing single model:', model_config)
model = build_detection_model(cfg, model_config)
model.to(cfg.MODEL.DEVICE)
# Initialize mixed-precision if necessary
use_mixed_precision = cfg.DTYPE == 'float16'
amp_handle = amp.init(enabled=use_mixed_precision, verbose=cfg.AMP_VERBOSE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
ckpt = cfg.MODEL.WEIGHT if args.ckpt is None else args.ckpt
_ = checkpointer.load(ckpt, use_latest=args.ckpt is None)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
# output_folders = [None] * len(cfg.DATASETS.TEST)
# dataset_names = cfg.DATASETS.TEST
dataset_names = cfg.DATASETS.NAS_VAL if not cfg.NAS.TRAIN_SINGLE_MODEL else cfg.DATASETS.TEST
output_folders = [None] * len(dataset_names)
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed,
test_only=cfg.TEST_ONLY)
if cfg.NAS.TRAIN_SINGLE_MODEL:
if get_rank() == 0:
print('==' * 20, 'Evaluating single model...', '==' * 20)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False
if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
bbox_aug=cfg.TEST.BBOX_AUG.ENABLED,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
c2d_json_path=cfg.MODEL.SEG_BRANCH.JSON_PATH,
cfg=cfg,
test_only=cfg.TEST_ONLY)
synchronize()
elif not cfg.NAS.SKIP_NAS_TEST:
if get_rank() == 0:
print('==' * 10, 'Start NAS testing', '==' * 10)
timer = Timer()
timer.tic()
searcher = PathPrioritySearch(cfg, './nas_test')
searcher.generate_fair_test(
) # load cache results and generate new model for test
searcher.search(model, output_folders, dataset_names, distributed)
searcher.save_topk()
total_time = timer.toc()
total_time_str = get_time_str(total_time)
if get_rank() == 0:
print('Finish NAS testing, total time:{}'.format(total_time_str))
return
else:
print('Skipping NAS testing...')
def main():
args = parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
output_dir = os.path.dirname(cfg.MODEL.WEIGHT)
cfg.OUTPUT_DIR = output_dir
if not os.path.exists(output_dir):
os.makedirs(output_dir)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
checkpointer = DetectronCheckpointer(cfg, model, save_dir=cfg.MODEL.WEIGHT)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
synchronize()
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
## add for training
parser.add_argument(
"--data-dir",
default="",
metavar="DIR",
help="path to data folder",
type=str,
)
parser.add_argument(
"--pretrained-model",
default="",
help="path to pretrained model",
metavar="FILE",
type=str,
)
parser.add_argument(
"--nonlocal-cls-num-group",
default="1",
help="nonlocal num group cls",
metavar="1",
type=int,
)
parser.add_argument(
"--nonlocal-cls-num-stack",
default="0",
help="nonlocal num stack cls",
metavar="1",
type=int,
)
parser.add_argument(
"--nonlocal-reg-num-group",
default="1",
help="nonlocal num group reg",
metavar="1",
type=int,
)
parser.add_argument(
"--nonlocal-reg-num-stack",
default="0",
help="nonlocal num stack reg",
metavar="1",
type=int,
)
parser.add_argument(
"--nonlocal-shared-num-group",
default="1",
help="nonlocal num group reg",
metavar="1",
type=int,
)
parser.add_argument(
"--nonlocal-shared-num-stack",
default="0",
help="nonlocal num stack reg",
metavar="1",
type=int,
)
parser.add_argument(
"--nonlocal-out-channels",
default="2048",
help="nonlocal out channels for fpn, fpn=2048(like c4)",
metavar="2048",
type=int,
)
parser.add_argument(
"--nonlocal-inter-channels",
default="256",
help="nonlocal inter channels, c4 < 2048, fpn < 256",
metavar="256",
type=int,
)
parser.add_argument(
"--nonlocal-use-shared",
default="True",
help="nonlocal use shared non-locael",
metavar="True",
type=str,
)
parser.add_argument(
"--nonlocal-use-bn",
default="True",
help="nonlocal use bn after attention",
metavar="True",
type=str,
)
parser.add_argument(
"--nonlocal-use-softmax",
default="False",
help="nonlocal use softmax other than div",
metavar="False",
type=str,
)
parser.add_argument(
"--nonlocal-use-attention",
default="True",
help="nonlocal use attention before ffconv",
metavar="True",
type=str,
)
parser.add_argument(
"--nonlocal-use-ffconv",
default="True",
help="nonlocal use ffconv after nonlocal with residual",
metavar="True",
type=str,
)
parser.add_argument(
"--nonlocal-use-relu",
default="True",
help="nonlocal use relu after bn",
metavar="True",
type=str,
)
parser.add_argument(
"--conv-bbox-expand",
default="1.0",
help="box expand conv",
metavar="1.0",
type=float,
)
parser.add_argument(
"--fc-bbox-expand",
default="1.0",
help="box expand fc",
metavar="1.0",
type=float,
)
parser.add_argument(
"--backbone-out-channels",
default="256",
help="fpn out channels for fpn, fpn=2048(like c4)",
metavar="256",
type=int,
)
parser.add_argument(
"--evaluation-flags",
nargs='*',
# default=[0, 3],
default=[],
help="model code for evaluation flags",
metavar="1 1 1 1",
type=int,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.DATA_DIR = args.data_dir
cfg.MODEL.WEIGHT = args.pretrained_model
cfg.MODEL.ROI_BOX_HEAD.NEIGHBOR_CONV_EXPAND = args.conv_bbox_expand
cfg.MODEL.ROI_BOX_HEAD.NEIGHBOR_FC_EXPAND = args.fc_bbox_expand
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_CLS_NUM_GROUP = args.nonlocal_cls_num_group
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_CLS_NUM_STACK = args.nonlocal_cls_num_stack
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_REG_NUM_GROUP = args.nonlocal_reg_num_group
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_REG_NUM_STACK = args.nonlocal_reg_num_stack
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_SHARED_NUM_GROUP = args.nonlocal_shared_num_group
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_SHARED_NUM_STACK = args.nonlocal_shared_num_stack
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_INTER_CHANNELS = args.nonlocal_inter_channels
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_OUT_CHANNELS = args.nonlocal_out_channels
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_USE_SHARED = ast.literal_eval(
args.nonlocal_use_shared)
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_USE_BN = ast.literal_eval(
args.nonlocal_use_bn)
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_USE_SOFTMAX = ast.literal_eval(
args.nonlocal_use_softmax)
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_USE_FFCONV = ast.literal_eval(
args.nonlocal_use_ffconv)
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_USE_RELU = ast.literal_eval(
args.nonlocal_use_relu)
cfg.MODEL.ROI_BOX_HEAD.NONLOCAL_USE_ATTENTION = ast.literal_eval(
args.nonlocal_use_attention)
cfg.MODEL.BACKBONE.OUT_CHANNELS = args.backbone_out_channels
# double heads
cfg.TEST.EVALUATION_FLAGS = args.evaluation_flags
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
synchronize()
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg, model, None, None, output_dir,
save_to_disk)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
# arguments.update(extra_checkpoint_data)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
logger = logging.getLogger("maskrcnn_benchmark.trainer")
if cfg.MODEL.META_ARCHITECTURE == 'AdaptionRCNN':
logger.info('AdaptionRCNN trainer is adapted!')
cross_do_train(
cfg,
model,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
distributed,
)
elif cfg.MODEL.META_ARCHITECTURE == 'GeneralizedRCNN':
logger.info('GeneralizedRCNN trainer is adapted!')
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
do_train(
cfg,
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
distributed,
)
return model
try:
cfg.merge_from_file(config_file)
except KeyError as e:
print(e)
cfg.INPUT.PIXEL_MEAN = [0, 0, 0]
cfg.INPUT.HORIZONTAL_FLIP_PROB_TRAIN = 0.5
cfg.INPUT.VERTICAL_FLIP_PROB_TRAIN = 0.5
cfg.INPUT.ROTATE_PROB_TRAIN = 1.0
cfg.INPUT.ROTATE_DEGREES_TRAIN = (-45, 45)
cfg.DATALOADER.NUM_WORKERS = 1
cfg.DATALOADER.SIZE_DIVISIBILITY = 0
cfg.freeze()
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=False,
start_iter=0,
)
device = 'cpu'
is_rotated = 1 #cfg.MODEL.ROTATED
if is_rotated:
from maskrcnn_benchmark.modeling.rrpn.utils import get_boxlist_rotated_rect_tensor
from maskrcnn_benchmark.modeling.rrpn.anchor_generator import draw_anchors
start_iter = 0
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
img_tensors = images.tensors
for id in range(len(targets)):
def run_test(cfg, model, distributed, iter, valid=False):
if distributed:
model = model.module
torch.cuda.empty_cache() # TODO check if it helps
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
if valid:
output_folders = [None] * len(cfg.DATASETS.VALID)
dataset_names = cfg.DATASETS.VALID
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(
cfg.OUTPUT_DIR, "validation", dataset_name,
'{}_{}'.format(iter, cfg.MODEL.ROI_HEADS.SCORE_THRESH))
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
dataset='valid',
is_distributed=distributed)
print(distributed)
results = []
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
# TODO if multiple valid set, result will be a list
result = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False
if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
results.append(result)
return results
else:
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(
cfg.OUTPUT_DIR, "inference", dataset_name,
'{}_{}'.format(iter, cfg.MODEL.ROI_HEADS.SCORE_THRESH))
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_test = make_data_loader(cfg,
dataset='test',
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_test):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False
if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
synchronize()
def train(cfg, local_rank, distributed):
model = build_detection_model(cfg) # 梦开始的地方
device = torch.device(cfg.MODEL.DEVICE) # ！！！！！
model.to(device)
for name, value in model.backbone.body.network.named_children(
): # 冻结主干网络参数
if int(name) > 60:
for param in value.parameters():
param.requires_grad = False
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16" # 这里可以改成float16来加速
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model,
optimizer,
opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg, model, optimizer, scheduler,
output_dir, save_to_disk)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
test_period = cfg.SOLVER.TEST_PERIOD
if test_period > 0:
data_loader_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed,
is_for_period=True)
else:
data_loader_val = None
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
do_train(
cfg,
model,
data_loader,
data_loader_val,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
test_period,
arguments,
)
return model
def do_train(model, data_loader, optimizer, scheduler, checkpointer, device,
checkpoint_period, arguments, meters):
logger = logging.getLogger("maskrcnn_benchmark.trainer")
logger.info("Start training")
max_iter = len(data_loader)
start_iter = arguments["iteration"]
model.train()
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=False)
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
cfg.DATASETS.TEST[0])
mkdir(output_folder)
start_training_time = time.time()
end = time.time()
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
model.train()
data_time = time.time() - end
iteration = iteration + 1
arguments["iteration"] = iteration
scheduler.step()
images = images.to(device)
targets = [target.to(device) for target in targets]
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
losses.backward()
optimizer.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time, data=data_time)
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if iteration % 20 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join([
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters),
lr=optimizer.param_groups[0]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
))
if iteration % checkpoint_period == 0:
checkpointer.save("model_{:07d}".format(iteration), **arguments)
# EVALUATION
# inference(
# model,
# cfg,
# data_loaders_val[0],
# dataset_name=cfg.DATASETS.TEST,
# device=cfg.MODEL.DEVICE,
# meters=meters,
# expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
# output_folder=output_folder,
# )
if iteration == max_iter:
checkpointer.save("model_final", **arguments)
total_training_time = time.time() - start_training_time
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / (max_iter)))
def main():
parser = argparse.ArgumentParser(description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default="/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.deprecated.init_process_group(
backend="nccl", init_method="env://"
)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox",)
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm",)
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg, is_train=False, is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
synchronize()
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
init()
tag = 17
set_epoch_tag(tag)
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
is_fp16 = (cfg.DTYPE == "float16")
if is_fp16:
# convert model to FP16
model.half()
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
# evaluator = get_evaluator()
start_test_time = time.time()
results = []
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
result = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
results.append(result)
end_test_time = time.time()
total_testing_time = end_test_time - start_test_time
if is_main_process():
map_results, raw_results = results[0]
bbox_map = map_results.results["bbox"]['AP']
segm_map = map_results.results["segm"]['AP']
print("BBOX_mAP: ", bbox_map, " MASK_mAP: ", segm_map)
print("Inference time: ", total_testing_time)