def main():
fcnxs = symbol_fcnxs.get_fcn32s_symbol(numclass=2, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN32s_VGG16"
num_epoch = 1
learning_rate = 1e-3
if args.model == "fcn16s":
fcnxs = symbol_fcnxs.get_fcn16s_symbol(numclass=2, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN16s_VGG16"
learning_rate = 1e-5
elif args.model == "fcn8s":
fcnxs = symbol_fcnxs.get_fcn8s_symbol(numclass=2, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN8s_VGG16"
num_epoch = 30
learning_rate = 1e-7
arg_names = fcnxs.list_arguments()
_, fcnxs_args, fcnxs_auxs = mx.model.load_checkpoint(args.prefix, args.epoch)
if not args.retrain:
if args.init_type == "vgg16":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_vgg16(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
elif args.init_type == "fcnxs":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_fcnxs(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
train_dataiter = FileIter(
root_dir = ".",
flist_name = "train.lst",
# cut_off_size = 400,
rgb_mean = (123.68, 116.779, 103.939),
)
# val_dataiter = FileIter(
# root_dir = "/home/zw/dataset/VOC2012Segmentation/VOC2012",
# flist_name = "val.lst",
# rgb_mean = (123.68, 116.779, 103.939),
# )
model = Solver(
ctx = ctx,
symbol = fcnxs,
begin_epoch = 0,
num_epoch = num_epoch,
arg_params = fcnxs_args,
aux_params = fcnxs_auxs,
learning_rate = learning_rate,
momentum = 0.9,
wd = 0.0001)
model.fit(
train_data = train_dataiter,
# eval_data = val_dataiter,
batch_end_callback = mx.callback.Speedometer(1, 10),
epoch_end_callback = mx.callback.do_checkpoint(fcnxs_model_prefix))
def main():
fcnxs = symbol_fcnxs.get_fcn32s_symbol(numclass=21, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN32s_VGG16"
if args.model == "fcn16s":
fcnxs = symbol_fcnxs.get_fcn16s_symbol(numclass=21, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN16s_VGG16"
elif args.model == "fcn8s":
fcnxs = symbol_fcnxs.get_fcn8s_symbol(numclass=21, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN8s_VGG16"
arg_names = fcnxs.list_arguments()
_, fcnxs_args, fcnxs_auxs = mx.model.load_checkpoint(args.prefix, args.epoch)
# print arg_names
if not args.retrain:
if args.init_type == "vgg16":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_vgg16(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
elif args.init_type == "fcnxs":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_fcnxs(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
train_dataiter = FileIter(
root_dir = "./VOC2012",
flist_name = "train.lst",
# cut_off_size = 400,
rgb_mean = (123.68, 116.779, 103.939),
)
val_dataiter = FileIter(
root_dir = "./VOC2012",
flist_name = "val.lst",
rgb_mean = (123.68, 116.779, 103.939),
)
model = Solver(
ctx = ctx,
symbol = fcnxs,
begin_epoch = 0,
num_epoch = 50,
arg_params = fcnxs_args,
aux_params = fcnxs_auxs,
learning_rate = 1e-10,
momentum = 0.99,
wd = 0.0005)
model.fit(
train_data = train_dataiter,
eval_data = val_dataiter,
batch_end_callback = mx.callback.Speedometer(1, 10),
epoch_end_callback = mx.callback.do_checkpoint(fcnxs_model_prefix))
def main():
symbol = symbol_fcnxs.get_fcn8s_symbol(numclass=21, workspace_default=1536)
_, fcnxs_args, fcnxs_auxs = mx.model.load_checkpoint(model_previx, epoch)
mod = mx.mod.Module(symbol, context=ctx)
#data_shape = fcnxs_args["data"].shape
data_shape = (1, 3, 800, 800)
mod.bind(data_shapes=[('data', (1, 3, data_shape[2], data_shape[3]))])
mod.set_params(fcnxs_args, fcnxs_auxs)
start = timer()
data = get_data(img, data_shape[2], data_shape[3])
mod.forward(Batch([mx.nd.array(data)]))
output = mod.get_outputs()[0]
out_img = np.uint8(np.squeeze(output.asnumpy().argmax(axis=1)))
time_elapsed = timer() - start
out_img = Image.fromarray(out_img)
out_img.putpalette(pallete)
out_img.save(seg)
print "Detection time: {:.6f} sec".format(time_elapsed)
def main():
ctx = mx.cpu() if not args.gpu else mx.gpu(args.gpu)
fcnxs = symbol_fcnxs.get_fcn32s_symbol(numclass=21, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN32s_VGG16"
if args.model == "fcn16s":
fcnxs = symbol_fcnxs.get_fcn16s_symbol(numclass=21, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN16s_VGG16"
elif args.model == "fcn8s":
fcnxs = symbol_fcnxs.get_fcn8s_symbol(numclass=21, workspace_default=1536)
fcnxs_model_prefix = "model_pascal/FCN8s_VGG16"
arg_names = fcnxs.list_arguments()
_, fcnxs_args, fcnxs_auxs = mx.model.load_checkpoint(args.prefix, args.epoch)
if not args.retrain:
if args.init_type == "vgg16":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_vgg16(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
elif args.init_type == "fcnxs":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_fcnxs(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
train_dataiter = FileIter(
root_dir = "./VOC2012",
flist_name = "train.lst",
# cut_off_size = 400,
rgb_mean = (123.68, 116.779, 103.939),
)
val_dataiter = FileIter(
root_dir = "./VOC2012",
flist_name = "val.lst",
rgb_mean = (123.68, 116.779, 103.939),
)
model = Solver(
ctx = ctx,
symbol = fcnxs,
begin_epoch = 0,
num_epoch = 50,
arg_params = fcnxs_args,
aux_params = fcnxs_auxs,
learning_rate = 1e-10,
momentum = 0.99,
wd = 0.0005)
model.fit(
train_data = train_dataiter,
eval_data = val_dataiter,
batch_end_callback = mx.callback.Speedometer(1, 10),
epoch_end_callback = mx.callback.do_checkpoint(fcnxs_model_prefix))
def main():
symbol = symbol_fcnxs.get_fcn8s_symbol(numclass=21, workspace_default=1536)
_, fcnxs_args, fcnxs_auxs = mx.model.load_checkpoint(model_previx, epoch)
mod = mx.mod.Module(symbol, context=ctx)
#data_shape = fcnxs_args["data"].shape
#data_shape = (1, 3, 800, 800)
# get size of video capture
video = cv2.VideoCapture(0)
width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH) + 0.5)
height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT) + 0.5)
if width<800 and height<800:
data_shape = (height, width)
elif width<height:
data_shape = (800, int(800.0*width/height))
else:
data_shape = (int(800.0*height/width), 800)
print data_shape
mod.bind(data_shapes=[('data', (1,3,data_shape[0], data_shape[1]))])
mod.set_params(fcnxs_args, fcnxs_auxs)
count = 0
while count<1000:
_, img = video.read()
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
start = timer()
data = get_data(img, data_shape[0], data_shape[1])
mod.forward(Batch([mx.nd.array(data)]))
output = mod.get_outputs()[0]
out_img = np.uint8(np.squeeze(output.asnumpy().argmax(axis=1)))
time_elapsed = timer() - start
out_img = Image.fromarray(out_img)
out_img.putpalette(pallete)
fname = 'data/'+str(count).zfill(7)+'.png'
out_img.save(fname)
print "Detection time: {:.6f} sec".format(time_elapsed)
count += 1
def main():
devs = mx.cpu() if args.gpus is None or args.gpus is '' else [
mx.gpu(int(i)) for i in args.gpus.split(',')
]
carvn_root = ''
num_classes = 2
cutoff = None if args.cutoff == 0 else args.cutoff
epochs = [74, 2, 1, 1]
if not os.path.exists(args.model_dir):
os.mkdir(args.model_dir)
model_prefixes = [
'VGG_FC_ILSVRC_16_layers', args.model_dir + "/FCN32s_VGG16",
args.model_dir + "/FCN16s_VGG16", args.model_dir + "/FCN8s_VGG16"
]
if args.model == "fcn16s":
fcnxs = symbol_fcnxs.get_fcn16s_symbol(numclass=num_classes,
workspace_default=1536)
fcnxs_model_prefix = model_prefixes[2]
load_prefix = model_prefixes[1]
lr = 1e-5
run_epochs = epochs[2]
load_epoch = epochs[1]
elif args.model == "fcn8s":
fcnxs = symbol_fcnxs.get_fcn8s_symbol(numclass=num_classes,
workspace_default=1536)
fcnxs_model_prefix = model_prefixes[3]
load_prefix = model_prefixes[2]
lr = 1e-6
run_epochs = epochs[3]
load_epoch = epochs[2]
else:
fcnxs = symbol_fcnxs.get_fcn32s_symbol(numclass=num_classes,
workspace_default=1536)
fcnxs_model_prefix = model_prefixes[1]
load_prefix = model_prefixes[0]
lr = 1e-4
run_epochs = epochs[1]
load_epoch = epochs[0]
arg_names = fcnxs.list_arguments()
print('loading', load_prefix, load_epoch)
print('lr', lr)
print('model_prefix', fcnxs_model_prefix)
print('running epochs', run_epochs)
print('cutoff size', cutoff)
args.batch_size = len(devs)
_, fcnxs_args, fcnxs_auxs = mx.model.load_checkpoint(
load_prefix, load_epoch)
ctx = mx.cpu()
if not args.retrain:
if args.model == "fcn16s" or args.model == "fcn8s":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_fcnxs(
ctx, fcnxs, fcnxs_args, fcnxs_auxs)
else:
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_vgg16(
ctx, fcnxs, fcnxs_args, fcnxs_auxs)
train_dataiter = BatchFileIter(
path_imglist="../data/train.lst",
cut_off_size=cutoff,
rgb_mean=(123.68, 116.779, 103.939),
batch_size=args.batch_size,
)
val_dataiter = BatchFileIter(
path_imglist="../data/val.lst",
cut_off_size=cutoff,
rgb_mean=(123.68, 116.779, 103.939),
batch_size=args.batch_size,
)
# learning rate
kv = mx.kvstore.create('local')
# create model
model = mx.mod.Module(
context=devs,
symbol=fcnxs,
#label_names = ['softmax_label', 'softmax2_label']
)
optimizer_params = {'learning_rate': lr, 'momentum': 0.99, 'wd': 0.0005}
_dice = DiceMetric()
_acc = AccMetric()
eval_metrics = [mx.metric.create(_dice), mx.metric.create(_acc)]
initializer = mx.init.Xavier(rnd_type='gaussian',
factor_type="in",
magnitude=2)
model.fit(
train_dataiter,
begin_epoch=0,
num_epoch=run_epochs,
#eval_data = val_dataiter,
eval_metric=eval_metrics,
kvstore=kv,
optimizer='sgd',
optimizer_params=optimizer_params,
initializer=initializer,
arg_params=fcnxs_args,
aux_params=fcnxs_auxs,
batch_end_callback=mx.callback.Speedometer(args.batch_size, 10),
epoch_end_callback=mx.callback.do_checkpoint(fcnxs_model_prefix),
allow_missing=True)
def main():
gpu_list = []
_gpus = args.gpu.split(',')
for _gpu in _gpus:
_gpu = _gpu.strip()
if len(_gpu)==0:
continue
gpu_list.append(int(_gpu))
assert len(gpu_list)>0
ctx = mx.gpu(gpu_list[0])
carvn_root = ''
num_classes = 2
cutoff = None if args.cutoff==0 else args.cutoff
epochs = [74,31,27,19]
model_prefixes = ['VGG_FC_ILSVRC_16_layers', args.model_dir+"/FCN32s_VGG16", args.model_dir+"/FCN16s_VGG16", args.model_dir+"/FCN8s_VGG16"]
if args.model == "fcn16s":
fcnxs = symbol_fcnxs.get_fcn16s_symbol(numclass=num_classes, workspace_default=1536)
fcnxs_model_prefix = model_prefixes[2]
load_prefix = model_prefixes[1]
lr = 1e-5
run_epochs = epochs[2]
load_epoch = epochs[1]
elif args.model == "fcn8s":
fcnxs = symbol_fcnxs.get_fcn8s_symbol(numclass=num_classes, workspace_default=1536)
fcnxs_model_prefix = model_prefixes[3]
load_prefix = model_prefixes[2]
lr = 1e-6
run_epochs = epochs[3]
load_epoch = epochs[2]
else:
fcnxs = symbol_fcnxs.get_fcn32s_symbol(numclass=num_classes, workspace_default=1536)
fcnxs_model_prefix = model_prefixes[1]
load_prefix = model_prefixes[0]
lr = 1e-4
run_epochs = epochs[1]
load_epoch = epochs[0]
arg_names = fcnxs.list_arguments()
print('loading', load_prefix, load_epoch)
print('lr', lr)
print('model_prefix', fcnxs_model_prefix)
print('running epochs', run_epochs)
_, fcnxs_args, fcnxs_auxs = mx.model.load_checkpoint(load_prefix, load_epoch)
if not args.retrain:
if args.model == "fcn16s" or args.model == "fcn8s":
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_fcnxs(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
else:
fcnxs_args, fcnxs_auxs = init_fcnxs.init_from_vgg16(ctx, fcnxs, fcnxs_args, fcnxs_auxs)
train_dataiter = FileIter(
root_dir = carvn_root,
flist_name = "../data/train.lst",
cut_off_size = cutoff,
rgb_mean = (123.68, 116.779, 103.939),
)
val_dataiter = FileIter(
root_dir = carvn_root,
flist_name = "../data/val.lst",
cut_off_size = cutoff,
rgb_mean = (123.68, 116.779, 103.939),
)
model = Solver(
ctx = ctx,
symbol = fcnxs,
begin_epoch = 0,
num_epoch = run_epochs,
arg_params = fcnxs_args,
aux_params = fcnxs_auxs,
learning_rate = lr,
momentum = 0.99,
wd = 0.0005)
_metric = DiceMetric()
model.fit(
train_data = train_dataiter,
eval_data = val_dataiter,
eval_metric = _metric,
batch_end_callback = mx.callback.Speedometer(1, 10),
epoch_end_callback = mx.callback.do_checkpoint(fcnxs_model_prefix))