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, 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 train(cfg, local_rank, distributed, logger):
model = SGEncode()
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, 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')
train_data_loader = get_loader(cfg, train_ids, test_ids, sg_data=sg_data, test_on=False, val_on=False, num_test=5000, num_val=1000)
val_data_loader = get_loader(cfg, train_ids, test_ids, sg_data=sg_data, test_on=False, val_on=True, num_test=5000, num_val=1000)
test_data_loader = get_loader(cfg, train_ids, test_ids, sg_data=sg_data, test_on=True, val_on=False, num_test=5000, num_val=1000)
debug_print(logger, 'end dataloader')
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
if os.path.exists(cfg.MODEL.PRETRAINED_DETECTOR_CKPT):
checkpoint = torch.load(cfg.MODEL.PRETRAINED_DETECTOR_CKPT, map_location=torch.device("cpu"))
model.load_state_dict(checkpoint)
if cfg.SOLVER.PRE_VAL:
logger.info("Validate before training")
run_val(cfg, model, val_data_loader, distributed, logger)
logger.info("Start training")
max_iter = len(train_data_loader)
start_training_time = time.time()
end = time.time()
test_result = run_test(cfg, model, test_data_loader, distributed, logger)
evaluator(logger, test_result)
torch.save(test_result, output_path % ('test', 0))
print_first_grad = True
for epoch in range(cfg.SOLVER.MAX_ITER):
epoch_loss = []
for iteration, (fg_imgs, fg_txts, bg_imgs, bg_txts) in enumerate(train_data_loader):
data_time = time.time() - end
iteration = iteration + 1
model.train()
for fg_img, fg_txt, bg_img, bg_txt in zip(fg_imgs, fg_txts, bg_imgs, bg_txts):
fg_img['entities'] = fg_img['entities'].to(device)
fg_img['relations'] = fg_img['relations'].to(device)
fg_img['graph'] = fg_img['graph'].to(device)
fg_txt['entities'] = fg_txt['entities'].to(device)
fg_txt['relations'] = fg_txt['relations'].to(device)
fg_txt['graph'] = fg_txt['graph'].to(device)
bg_img['entities'] = bg_img['entities'].to(device)
bg_img['relations'] = bg_img['relations'].to(device)
bg_img['graph'] = bg_img['graph'].to(device)
bg_txt['entities'] = bg_txt['entities'].to(device)
bg_txt['relations'] = bg_txt['relations'].to(device)
bg_txt['graph'] = bg_txt['graph'].to(device)
loss_list = model(fg_imgs, fg_txts, bg_imgs, bg_txts)
losses = sum(loss_list) / (len(loss_list) + 1e-9)
epoch_loss.append(float(losses))
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, 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()
logger.info("epoch: {epoch} loss: {loss:.6f} lr: {lr:.6f}".format(epoch=epoch, loss=float(sum(epoch_loss) / len(epoch_loss)), lr=optimizer.param_groups[-1]["lr"]))
if epoch % checkpoint_period == 0:
torch.save(model.state_dict(), os.path.join(cfg.OUTPUT_DIR, "model_{}.pytorch".format(str(epoch))))
if epoch == max_iter:
torch.save(model.state_dict(), os.path.join(cfg.OUTPUT_DIR, "model_final.pytorch"))
val_result = None # used for scheduler updating
if cfg.SOLVER.TO_VAL and epoch % cfg.SOLVER.VAL_PERIOD == 0:
logger.info("Start testing")
test_result = run_test(cfg, model, test_data_loader, distributed, logger)
test_similarity = evaluator(logger, test_result)
torch.save({'result' : test_result, 'similarity' : test_similarity}, output_path % ('test', epoch))
logger.info("Start validating")
val_result = run_test(cfg, model, val_data_loader, distributed, logger)
val_similarity = evaluator(logger, val_result)
torch.save({'result' : val_result, 'similarity' : val_similarity}, output_path % ('val', epoch))
# scheduler should be called after optimizer.step() in pytorch>=1.1.0
assert cfg.SOLVER.SCHEDULE.TYPE != "WarmupReduceLROnPlateau"
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, test_result