from lib.fpn.box_utils import bbox_loss
import torch.backends.cudnn as cudnn
from pycocotools.cocoeval import COCOeval
from lib.pytorch_misc import optimistic_restore, clip_grad_norm
from torch.optim.lr_scheduler import ReduceLROnPlateau
cudnn.benchmark = True
conf = ModelConfig()
if conf.coco:
train, val = CocoDetection.splits()
val.ids = val.ids[:conf.val_size]
train.ids = train.ids
train_loader, val_loader = CocoDataLoader.splits(
train,
val,
batch_size=conf.batch_size,
num_workers=conf.num_workers,
num_gpus=conf.num_gpus)
else:
train, val, _ = VG.splits(num_val_im=conf.val_size,
filter_non_overlap=False,
filter_empty_rels=False,
use_proposals=conf.use_proposals)
train_loader, val_loader = VGDataLoader.splits(
train,
val,
batch_size=conf.batch_size,
num_workers=conf.num_workers,
num_gpus=conf.num_gpus)
detector = ObjectDetector(
def main(args):
print(args)
check_args(args)
if not exists(args.output_dir):
os.makedirs(args.output_dir)
summary_writer = SummaryWriter(args.output_dir)
if args.coco:
train, val = CocoDetection.splits()
val.ids = val.ids[:args.val_size]
train.ids = train.ids
train_loader, val_loader = CocoDataLoader.splits(train, val, batch_size=args.batch_size,
num_workers=args.num_workers,
num_gpus=args.num_gpus)
else:
train, val, _ = VG.splits(num_val_im=args.val_size, filter_non_overlap=False,
filter_empty_rels=False, use_proposals=args.use_proposals)
train_loader, val_loader = VGDataLoader.splits(train, val, batch_size=args.batch_size,
num_workers=args.num_workers,
num_gpus=args.num_gpus)
print(train.ind_to_classes)
os._exit(0)
all_in_one_model = neural_motifs_sg2im_model(args, train.ind_to_classes)
# Freeze the detector
for n, param in all_in_one_model.detector.named_parameters():
param.requires_grad = False
all_in_one_model.cuda()
gan_g_loss, gan_d_loss = get_gan_losses(args.gan_loss_type)
t, epoch, checkpoint = all_in_one_model.t, all_in_one_model.epoch, all_in_one_model.checkpoint
while True:
if t >= args.num_iterations:
break
epoch += 1
print('Starting epoch %d' % epoch)
for step, batch in enumerate(tqdm(train_loader, desc='Training Epoch %d' % epoch, total=len(train_loader))):
if t == args.eval_mode_after:
print('switching to eval mode')
all_in_one_model.model.eval()
all_in_one_model.optimizer = optim.Adam(all_in_one_model.parameters(), lr=args.learning_rate)
t += 1
with timeit('forward', args.timing):
result = all_in_one_model[batch]
imgs, imgs_pred, objs, g_scores_fake_crop, g_obj_scores_fake_crop, g_scores_fake_img, \
d_scores_fake_crop, d_obj_scores_fake_crop, d_scores_real_crop, d_obj_scores_real_crop, \
d_scores_fake_img, d_scores_real_img = result.imgs, result.imgs_pred, result.objs, \
result.g_scores_fake_crop, result.g_obj_scores_fake_crop, result.g_scores_fake_img, \
result.d_scores_fake_crop, result.d_obj_scores_fake_crop, result.d_scores_real_crop, \
result.d_obj_scores_real_crop, result.d_scores_fake_img, result.d_scores_real_img
with timeit('loss', args.timing):
total_loss, losses = calculate_model_losses(
args, imgs, imgs_pred)
if all_in_one_model.obj_discriminator is not None:
total_loss = add_loss(total_loss, F.cross_entropy(g_obj_scores_fake_crop, objs), losses, 'ac_loss',
args.ac_loss_weight)
weight = args.discriminator_loss_weight * args.d_obj_weight
total_loss = add_loss(total_loss, gan_g_loss(g_scores_fake_crop), losses,
'g_gan_obj_loss', weight)
if all_in_one_model.img_discriminator is not None:
weight = args.discriminator_loss_weight * args.d_img_weight
total_loss = add_loss(total_loss, gan_g_loss(g_scores_fake_img), losses,
'g_gan_img_loss', weight)
losses['total_loss'] = total_loss.item()
if not math.isfinite(losses['total_loss']):
print('WARNING: Got loss = NaN, not backpropping')
continue
with timeit('backward', args.timing):
all_in_one_model.optimizer.zero_grad()
total_loss.backward()
all_in_one_model.optimizer.step()
if all_in_one_model.obj_discriminator is not None:
with timeit('d_obj loss', args.timing):
d_obj_losses = LossManager()
d_obj_gan_loss = gan_d_loss(d_scores_real_crop, d_scores_fake_crop)
d_obj_losses.add_loss(d_obj_gan_loss, 'd_obj_gan_loss')
d_obj_losses.add_loss(F.cross_entropy(d_obj_scores_real_crop, objs), 'd_ac_loss_real')
d_obj_losses.add_loss(F.cross_entropy(d_obj_scores_fake_crop, objs), 'd_ac_loss_fake')
with timeit('d_obj backward', args.timing):
all_in_one_model.optimizer_d_obj.zero_grad()
d_obj_losses.total_loss.backward()
all_in_one_model.optimizer_d_obj.step()
if all_in_one_model.img_discriminator is not None:
with timeit('d_img loss', args.timing):
d_img_losses = LossManager()
d_img_gan_loss = gan_d_loss(d_scores_real_img, d_scores_fake_img)
d_img_losses.add_loss(d_img_gan_loss, 'd_img_gan_loss')
with timeit('d_img backward', args.timing):
all_in_one_model.optimizer_d_img.zero_grad()
d_img_losses.total_loss.backward()
all_in_one_model.optimizer_d_img.step()
if t % args.print_every == 0:
print('t = %d / %d' % (t, args.num_iterations))
G_loss_list = []
for name, val in losses.items():
G_loss_list.append('[%s]: %.4f' % (name, val))
checkpoint['losses'][name].append(val)
summary_writer.add_scalar("G_%s" % name, val, t)
print("G: %s" % ", ".join(G_loss_list))
checkpoint['losses_ts'].append(t)
if all_in_one_model.obj_discriminator is not None:
D_obj_loss_list = []
for name, val in d_obj_losses.items():
D_obj_loss_list.append('[%s]: %.4f' % (name, val))
checkpoint['d_losses'][name].append(val)
summary_writer.add_scalar("D_obj_%s" % name, val, t)
print("D_obj: %s" % ", ".join(D_obj_loss_list))
if all_in_one_model.img_discriminator is not None:
D_img_loss_list = []
for name, val in d_img_losses.items():
D_img_loss_list.append('[%s]: %.4f' % (name, val))
checkpoint['d_losses'][name].append(val)
summary_writer.add_scalar("D_img_%s" % name, val, t)
print("D_img: %s" % ", ".join(D_img_loss_list))
if t % args.checkpoint_every == 0:
print('checking on train')
train_results = check_model(args, train_loader, all_in_one_model)
t_losses, t_samples = train_results
checkpoint['train_samples'].append(t_samples)
checkpoint['checkpoint_ts'].append(t)
for name, images in t_samples.items():
summary_writer.add_image("train_%s" % name, images, t)
print('checking on val')
val_results = check_model(args, val_loader, all_in_one_model)
val_losses, val_samples = val_results
checkpoint['val_samples'].append(val_samples)
for name, images in val_samples.items():
summary_writer.add_image("val_%s" % name, images, t)
for k, v in val_losses.items():
checkpoint['val_losses'][k].append(v)
summary_writer.add_scalar("val_%s" % k, v, t)
checkpoint['model_state'] = all_in_one_model.model.state_dict()
if all_in_one_model.obj_discriminator is not None:
checkpoint['d_obj_state'] = all_in_one_model.obj_discriminator.state_dict()
checkpoint['d_obj_optim_state'] = all_in_one_model.optimizer_d_obj.state_dict()
if all_in_one_model.img_discriminator is not None:
checkpoint['d_img_state'] = all_in_one_model.img_discriminator.state_dict()
checkpoint['d_img_optim_state'] = all_in_one_model.optimizer_d_img.state_dict()
checkpoint['optim_state'] = all_in_one_model.optimizer.state_dict()
checkpoint['counters']['t'] = t
checkpoint['counters']['epoch'] = epoch
checkpoint_path = os.path.join(args.output_dir,
'%s_with_model.pt' % args.checkpoint_name)
print('Saving checkpoint to ', checkpoint_path)
torch.save(checkpoint, checkpoint_path)
# Save another checkpoint without any model or optim state
checkpoint_path = os.path.join(args.output_dir,
'%s_no_model.pt' % args.checkpoint_name)
key_blacklist = ['model_state', 'optim_state', 'model_best_state',
'd_obj_state', 'd_obj_optim_state', 'd_obj_best_state',
'd_img_state', 'd_img_optim_state', 'd_img_best_state']
small_checkpoint = {}
for k, v in checkpoint.items():
if k not in key_blacklist:
small_checkpoint[k] = v
torch.save(small_checkpoint, checkpoint_path)