def from_config(config, load_stored_params=True, model_param_file=None):
"""
Load a network from a config file
:param load_stored_params: whether or not to load stored params, if so there should be
a "path_to_network" entry in the config
:type load_stored_params: bool
:param config: Dict specifying details of the network architecture
e.g.
path_to_network: /home/manuelli/code/dense_correspondence/recipes/trained_models/10_drill_long_3d
parameter_file: dense_resnet_34_8s_03505.pth
descriptor_dimensionality: 3
image_width: 640
image_height: 480
:return: DenseCorrespondenceNetwork
:rtype:
"""
# fcn = resnet_dilated.Resnet34_8s(num_classes=config['descriptor_dimension'])
fcn = resnet_dilated.Resnet34_8s(
num_classes=config['descriptor_dimension'])
if 'normalize' in config:
normalize = config['normalize']
else:
normalize = False
dcn = DenseCorrespondenceNetwork(fcn,
config['descriptor_dimension'],
image_width=config['image_width'],
image_height=config['image_height'],
normalize=normalize)
if load_stored_params:
assert model_param_file is not None
config[
'model_param_file'] = model_param_file # should be an absolute path
try:
dcn.load_state_dict(
torch.load(model_param_file, map_location='cpu'))
except:
logging.info(
"loading params with the new style failed, falling back to dcn.fcn.load_state_dict"
)
dcn.fcn.load_state_dict(torch.load(model_param_file))
#dcn.cuda()
dcn.train()
dcn.config = config
return dcn
def from_model_folder(model_folder, load_stored_params=True, model_param_file=None,
iteration=None):
"""
Loads a DenseCorrespondenceNetwork from a model folder
:param model_folder: the path to the folder where the model is stored. This direction contains
files like
- 003500.pth
- training.yaml
:type model_folder:
:return: a DenseCorrespondenceNetwork objecc t
:rtype:
"""
model_folder = utils.convert_to_absolute_path(model_folder)
if model_param_file is None:
model_param_file, _, _ = utils.get_model_param_file_from_directory(model_folder, iteration=iteration)
model_param_file = utils.convert_to_absolute_path(model_param_file)
training_config_filename = os.path.join(model_folder, "training.yaml")
training_config = utils.getDictFromYamlFilename(training_config_filename)
config = training_config["dense_correspondence_network"]
config["path_to_network_params_folder"] = model_folder
fcn = resnet_dilated.Resnet34_8s(num_classes=config['descriptor_dimension'])
dcn = DenseCorrespondenceNetwork(fcn, config['descriptor_dimension'],
image_width=config['image_width'],
image_height=config['image_height'])
# load the stored params
if load_stored_params:
# old syntax
try:
dcn.load_state_dict(torch.load(model_param_file))
except:
logging.info("loading params with the new style failed, falling back to dcn.fcn.load_state_dict")
dcn.fcn.load_state_dict(torch.load(model_param_file))
# this is the new format
#
dcn.cuda()
dcn.train()
dcn.config = config
return dcn
def from_config(config, load_stored_params=True):
"""
Load a network from a config file
:param load_stored_params: whether or not to load stored params, if so there should be
a "path_to_network" entry in the config
:type load_stored_params: bool
:param config: Dict specifying details of the network architecture
e.g.
path_to_network: /home/manuelli/code/dense_correspondence/recipes/trained_models/10_drill_long_3d
parameter_file: dense_resnet_34_8s_03505.pth
descriptor_dimensionality: 3
image_width: 640
image_height: 480
:return: DenseCorrespondenceNetwork
:rtype:
"""
fcn = resnet_dilated.Resnet34_8s(num_classes=config['descriptor_dimension'])
if load_stored_params:
path_to_network_params = utils.convert_to_absolute_path(config['path_to_network_params'])
config['path_to_network_params_folder'] = os.path.dirname(config['path_to_network_params'])
fcn.load_state_dict(torch.load(path_to_network_params))
dcn = DenseCorrespondenceNetwork(fcn, config['descriptor_dimension'],
image_width=config['image_width'],
image_height=config['image_height'])
dcn.cuda()
dcn.train()
dcn.config = config
return dcn
# Initialization
model_path = pwd + '/our_data/resnet_34_8s_68.pth'
source_video_path = pwd + '/our_data/ariel.mp4'
target_video_path = pwd + '/our_data/ariel_mask.avi' # OpenCV must have avi as output. https://github.com/ContinuumIO/anaconda-issues/issues/223#issuecomment-285523938
skip_frames = 0
# Create and load weights to pre-trained Resnet-34 model:
from pathlib import Path
import urllib.request
my_file = Path(model_path)
if not my_file.is_file():
# If weights file doesn't exist, download it from dropbox
print("Weights file doesn't exist. Downloading...")
url = "https://www.dropbox.com/s/91wcu6bpqezu4br/resnet_34_8s_68.pth?dl=1"
u = urllib.request.urlopen(url)
data = u.read()
u.close()
with open(model_path, "wb") as f:
f.write(data)
##
fcn = resnet_dilated.Resnet34_8s(num_classes=21)
fcn.load_state_dict(
torch.load(model_path, map_location=lambda storage, loc: storage))
create_masked_video(fcn,
source_video_path,
target_video_path,
skipframes=skip_frames)