def test_deeplabv3_1p():
start_time = time.time()
epoch_size = 100
args_opt = argparse.Namespace(base_size=513, crop_size=513, batch_size=2)
args_opt.base_size = config.crop_size
args_opt.crop_size = config.crop_size
args_opt.batch_size = config.batch_size
train_dataset = create_dataset(args_opt, data_url, 1, config.batch_size,
usage="eval")
dataset_size = train_dataset.get_dataset_size()
callback = LossCallBack(dataset_size)
net = deeplabv3_resnet50(config.seg_num_classes, [config.batch_size, 3, args_opt.crop_size, args_opt.crop_size],
infer_scale_sizes=config.eval_scales, atrous_rates=config.atrous_rates,
decoder_output_stride=config.decoder_output_stride, output_stride=config.output_stride,
fine_tune_batch_norm=config.fine_tune_batch_norm, image_pyramid=config.image_pyramid)
net.set_train()
model_fine_tune(net, 'layer')
loss = OhemLoss(config.seg_num_classes, config.ignore_label)
opt = Momentum(filter(lambda x: 'beta' not in x.name and 'gamma' not in x.name and 'depth' not in x.name and 'bias' not in x.name, net.trainable_params()), learning_rate=config.learning_rate, momentum=config.momentum, weight_decay=config.weight_decay)
model = Model(net, loss, opt)
model.train(epoch_size, train_dataset, callback)
print(time.time() - start_time)
print("expect loss: ", callback.loss)
print("expect time: ", callback.time)
expect_loss = 0.92
expect_time = 40
assert callback.loss.asnumpy() <= expect_loss
assert callback.time <= expect_time
def model_fine_tune(flags, train_net, fix_weight_layer):
checkpoint_path = flags.checkpoint_url
if checkpoint_path is None:
return
param_dict = load_checkpoint(checkpoint_path)
load_param_into_net(train_net, param_dict)
for para in train_net.trainable_params():
if fix_weight_layer in para.name:
para.requires_grad = False
if __name__ == "__main__":
start_time = time.time()
epoch_size = 3
args_opt.base_size = config.crop_size
args_opt.crop_size = config.crop_size
train_dataset = create_dataset(args_opt, args_opt.data_url, 1, config.batch_size,
usage="train", shuffle=False)
dataset_size = train_dataset.get_dataset_size()
callback = LossCallBack(dataset_size)
net = deeplabv3_resnet50(config.seg_num_classes, [config.batch_size, 3, args_opt.crop_size, args_opt.crop_size],
infer_scale_sizes=config.eval_scales, atrous_rates=config.atrous_rates,
decoder_output_stride=config.decoder_output_stride, output_stride=config.output_stride,
fine_tune_batch_norm=config.fine_tune_batch_norm, image_pyramid=config.image_pyramid)
net.set_train()
model_fine_tune(args_opt, net, 'layer')
loss = OhemLoss(config.seg_num_classes, config.ignore_label)
opt = Momentum(filter(lambda x: 'beta' not in x.name and 'gamma' not in x.name and 'depth' not in x.name and 'bias' not in x.name, net.trainable_params()), learning_rate=config.learning_rate, momentum=config.momentum, weight_decay=config.weight_decay)
model = Model(net, loss, opt)
model.train(epoch_size, train_dataset, callback)
print(time.time() - start_time)
print("expect loss: ", callback.loss / 3)
print("expect time: ", callback.time)
from src.config import config
parser = argparse.ArgumentParser(description="Deeplabv3 evaluation")
parser.add_argument('--epoch_size', type=int, default=2, help='Epoch size.')
parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.")
parser.add_argument('--batch_size', type=int, default=2, help='Batch size.')
parser.add_argument('--data_url', required=True, default=None, help='Evaluation data url')
parser.add_argument('--checkpoint_url', default=None, help='Checkpoint path')
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=args_opt.device_id)
print(args_opt)
if __name__ == "__main__":
args_opt.crop_size = config.crop_size
args_opt.base_size = config.crop_size
eval_dataset = create_dataset(args_opt, args_opt.data_url, args_opt.epoch_size, args_opt.batch_size, usage="eval")
net = deeplabv3_resnet50(config.seg_num_classes, [args_opt.batch_size, 3, args_opt.crop_size, args_opt.crop_size],
infer_scale_sizes=config.eval_scales, atrous_rates=config.atrous_rates,
decoder_output_stride=config.decoder_output_stride, output_stride=config.output_stride,
fine_tune_batch_norm=config.fine_tune_batch_norm, image_pyramid=config.image_pyramid)
param_dict = load_checkpoint(args_opt.checkpoint_url)
load_param_into_net(net, param_dict)
mIou = MiouPrecision(config.seg_num_classes)
metrics = {'mIou': mIou}
loss = OhemLoss(config.seg_num_classes, config.ignore_label)
model = Model(net, loss, metrics=metrics)
model.eval(eval_dataset)
load_param_into_net(train_net, param_dict)
for para in train_net.trainable_params():
if fix_weight_layer in para.name:
para.requires_grad = False
if __name__ == "__main__":
if args_opt.distribute == "true":
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True)
init()
args_opt.base_size = config.crop_size
args_opt.crop_size = config.crop_size
train_dataset = create_dataset(args_opt,
args_opt.data_url,
config.epoch_size,
config.batch_size,
usage="train")
dataset_size = train_dataset.get_dataset_size()
time_cb = TimeMonitor(data_size=dataset_size)
callback = [time_cb, LossCallBack()]
if config.enable_save_ckpt:
config_ck = CheckpointConfig(
save_checkpoint_steps=config.save_checkpoint_steps,
keep_checkpoint_max=config.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(prefix='checkpoint_deeplabv3',
config=config_ck)
callback.append(ckpoint_cb)
net = deeplabv3_resnet50(
config.seg_num_classes,
[config.batch_size, 3, args_opt.crop_size, args_opt.crop_size],
if args_opt.distribute == "true":
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True)
init()
args_opt.base_size = config.crop_size
args_opt.crop_size = config.crop_size
import moxing as mox
mox.file.copy_parallel(src_url=args_opt.data_url, dst_url='voc2012/')
mox.file.copy_parallel(src_url=args_opt.checkpoint_url,
dst_url='checkpoint/')
# train
train_dataset = create_dataset(args_opt,
data_path,
config.epoch_size,
config.batch_size,
usage="train")
dataset_size = train_dataset.get_dataset_size()
time_cb = TimeMonitor(data_size=dataset_size)
callback = [time_cb, LossCallBack()]
if config.enable_save_ckpt:
config_ck = CheckpointConfig(
save_checkpoint_steps=config.save_checkpoint_steps,
keep_checkpoint_max=config.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(prefix='checkpoint_deeplabv3',
config=config_ck)
callback.append(ckpoint_cb)
net = deeplabv3_resnet50(
config.seg_num_classes,
[config.batch_size, 3, args_opt.crop_size, args_opt.crop_size],