def main(argv):
"""
Run training of the Deeppose stg-1
"""
args = cmd_options.get_arguments(argv)
if not os.path.exists(args.o_dir):
os.makedirs(args.o_dir)
suffix = datetime.datetime.now().strftime("%y%m%d_%H%M%S")
with open(os.path.join(args.o_dir, 'params.dump_{}.txt'.format(suffix)), 'w') as f:
f.write('{}\n'.format(pprint.pformat(args)))
net, loss_op, pose_loss_op, train_op = regressionnet.create_regression_net(
n_joints=args.n_joints,
init_snapshot_path=args.snapshot,
is_resume=args.resume,
reset_iter_counter=args.reset_iter_counter,
reset_moving_averages=args.reset_moving_averages,
optimizer_type=args.optimizer,
gpu_memory_fraction=0.32, # Set how much GPU memory to reserve for the network
net_type=args.net_type)
with net.graph.as_default():
saver = tf.train.Saver()
print 'args.resume: {}\nargs.snapshot: {}'.format(args.resume, args.snapshot)
bbox_extension_range = (args.bbox_extension_min, args.bbox_extension_max)
if bbox_extension_range[0] is None or bbox_extension_range[1] is None:
bbox_extension_range = None
test_bbox_extension_range = None
else:
test_bbox_extension_range = (bbox_extension_range[1], bbox_extension_range[1])
train_dataset = dataset.PoseDataset(
args.train_csv_fn, args.img_path_prefix, args.im_size,
fliplr=args.fliplr,
rotate=args.rotate,
rotate_range=args.rotate_range,
shift=args.shift,
bbox_extension_range=bbox_extension_range,
min_dim=args.min_dim,
coord_normalize=args.coord_normalize,
gcn=args.gcn,
fname_index=args.fname_index,
joint_index=args.joint_index,
symmetric_joints=args.symmetric_joints,
ignore_label=args.ignore_label,
should_downscale_images=args.should_downscale_images,
downscale_height=args.downscale_height
)
test_dataset = dataset.PoseDataset(
args.test_csv_fn, args.img_path_prefix, args.im_size,
fliplr=False, rotate=False,
shift=None,
bbox_extension_range=test_bbox_extension_range,
coord_normalize=args.coord_normalize,
gcn=args.gcn,
fname_index=args.fname_index,
joint_index=args.joint_index,
symmetric_joints=args.symmetric_joints,
ignore_label=args.ignore_label,
should_return_bbox=True,
should_downscale_images=args.should_downscale_images,
downscale_height=args.downscale_height
)
np.random.seed(args.seed)
train_iterator = iterators.MultiprocessIterator(train_dataset, args.batch_size,
n_processes=args.workers, n_prefetch=3)
test_iterator = iterators.MultiprocessIterator(
test_dataset, args.batch_size,
repeat=False, shuffle=False,
n_processes=1, n_prefetch=1)
val_iterator = None
if args.val_csv_fn is not None and args.val_csv_fn != '':
small_train_dataset = dataset.PoseDataset(
args.val_csv_fn,
args.img_path_prefix, args.im_size,
fliplr=False, rotate=False,
shift=None,
bbox_extension_range=test_bbox_extension_range,
coord_normalize=args.coord_normalize,
gcn=args.gcn,
fname_index=args.fname_index,
joint_index=args.joint_index,
symmetric_joints=args.symmetric_joints,
ignore_label=args.ignore_label,
should_return_bbox=True,
should_downscale_images=args.should_downscale_images,
downscale_height=args.downscale_height
)
val_iterator = iterators.MultiprocessIterator(
small_train_dataset, args.batch_size,
repeat=False, shuffle=False,
n_processes=1, n_prefetch=1)
train_loop(net, saver, loss_op, pose_loss_op, train_op, args.dataset_name,
train_iterator, test_iterator,
val_iterator=val_iterator,
max_iter=args.max_iter,
test_step=args.test_step,
log_step=args.log_step,
snapshot_step=args.snapshot_step,
batch_size=args.batch_size,
conv_lr=args.conv_lr,
fc_lr=args.fc_lr,
fix_conv_iter=args.fix_conv_iter,
output_dir=args.o_dir
)
return optimizer.lr
trainer.extend(Evaluator(
test_iter, model, device=args.gpu), trigger=val_interval)
trainer.extend(extensions.observe_value(
'lr', lambda _: lr_shift()), trigger=(1, 'epoch'))
trainer.extend(extensions.dump_graph('main/loss'))
trainer.extend(extensions.snapshot_object(
model, 'epoch_{.updater.epoch}.model'), trigger=val_interval)
trainer.extend(extensions.snapshot_object(
optimizer, 'epoch_{.updater.epoch}.state'), trigger=val_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
start_time = time.time()
trainer.extend(extensions.observe_value(
'time', lambda _: time.time() - start_time), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'time', 'epoch', 'iteration', 'main/loss', 'validation/main/loss',
'main/accuracy', 'validation/main/accuracy', 'lr',
]), trigger=log_interval)
trainer.extend(extensions.observe_value(
'graph', lambda _: create_fig(args.dir)),
trigger=(1, 'epoch'), priority=50)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.run()
if __name__ == '__main__':
args = cmd_options.get_arguments()
main(args)
model(x, t)
sum_loss += float(model.loss.data) * input_data.shape[0]
num += input_data.shape[0]
logging.info('loss:{}'.format(sum_loss / num))
# quit training data loading thread
input_q.put(None)
data_loader.join()
return sum_loss
if __name__ == '__main__':
args = get_arguments()
if cuda.available and args.gpu >= 0:
cuda.get_device(args.gpu).use()
np.random.seed(args.seed)
# create result dir
create_result_dir(args)
# create model and optimizer
model, optimizer = get_model_optimizer(args)
train_dl, test_dl = load_dataset(args)
N, N_test = len(train_dl), len(test_dl)
logging.info('# of training data:{}'.format(N))
logging.info('# of test data:{}'.format(N_test))
input_data[0] = x.transpose((2, 0, 1))
label[0] = t
return orig, input_data, label
if __name__ == '__main__':
sys.path.append('scripts')
from test_flic_dataset import draw_joints as flic_draw_joints
from test_lsp_dataset import draw_joints as lsp_draw_joints
from test_mpii_dataset import draw_joints as mpii_draw_joints
from transform import Transform
from cmd_options import get_arguments
args = get_arguments()
print(args)
flic_swap_joints = [(2, 4), (1, 5), (0, 6)]
lsp_swap_joints = [(8, 9), (7, 10), (6, 11), (2, 3), (1, 4), (0, 5)]
mpii_swap_joints = [(12, 13), (11, 14), (10, 15), (2, 3), (1, 4), (0, 5)]
for datadir, n_joints, draw_joints, swap_joints, min_dim in [
('data/FLIC-full', 7, flic_draw_joints, flic_swap_joints, 100),
('data/lspet_dataset', 14, lsp_draw_joints, lsp_swap_joints, 50),
('data/mpii', 16, mpii_draw_joints, mpii_swap_joints, 100)]:
args.datadir = datadir
args.joint_num = n_joints
args.min_dim = min_dim
# augmentation setting
trans = Transform(args, swap_joints)
j = 0
while j != len(x_batch):
a, b = o_queue.get()
input_data[j] = a
label[j] = b
j += 1
minibatch_q.put([input_data, label])
for _ in range(args.batchsize):
x_queue.put(None)
for w in workers:
w.join()
if __name__ == '__main__':
args = cmd_options.get_arguments()
result_dir = create_result_dir(args.model, args.resume_model)
create_logger(args, result_dir)
model = get_model(args.model, args.n_joints, result_dir, args.resume_model)
model = loss.PoseEstimationError(model)
opt = get_optimizer(model, args.opt, args.lr, adam_alpha=args.adam_alpha,
adam_beta1=args.adam_beta1, adam_beta2=args.adam_beta2,
adam_eps=args.adam_eps, weight_decay=args.weight_decay,
resume_opt=args.resume_opt)
train_dataset = dataset.PoseDataset(
args.train_csv_fn, args.img_dir, args.im_size, args.fliplr,
args.rotate, args.rotate_range, args.zoom, args.base_zoom,
args.zoom_range, args.translate, args.translate_range, args.min_dim,
args.coord_normalize, args.gcn, args.n_joints, args.fname_index,
args.joint_index, args.symmetric_joints, args.ignore_label
)
