def load_pretrained_model(self):
base_dir = self.config.pretrain_model.exp_dir
checkpoint_name = self.config.pretrain_model.checkpoint_name
config_path = os.path.join(base_dir, 'config.json')
config_json = load_json(config_path)
config = DotMap(config_json)
SystemClass = globals()[config.system]
system = SystemClass(config)
checkpoint_file = os.path.join(base_dir, 'checkpoints',
checkpoint_name)
checkpoint = torch.load(checkpoint_file, map_location=self.device)
system.load_state_dict(checkpoint['state_dict'])
encoder = system.model.eval()
viewmaker = system.view.eval()
for param in encoder.parameters():
param.requires_grad = False
for param in viewmaker.parameters():
param.requires_grad = False
return encoder, viewmaker, system, system.config
def get_system(model_path, device):
config_dir = '/'.join(
model_path.split('/')[:-2]) # Remove "checkpoints/epoch=X.ckpt"
config_json = load_json(config_dir + '/config.json')
config = DotMap(config_json)
system = TransferExpertSystem(config)
checkpoint = torch.load(model_path, map_location=device)
system.load_state_dict(checkpoint['state_dict'], strict=False)
return system
def load_pretrained_model(self):
base_dir = self.config.pretrain_model.exp_dir
checkpoint_name = self.config.pretrain_model.checkpoint_name
config_path = os.path.join(base_dir, 'config.json')
config_json = utils.load_json(config_path)
config = DotMap(config_json)
# overwrite GPU to load on same as current agent
config.gpu_device = self.config.gpu_device
SystemClass = globals()[config.system]
system = SystemClass(config)
checkpoint_file = os.path.join(base_dir, 'checkpoints', checkpoint_name)
checkpoint = torch.load(checkpoint_file)
system.load_state_dict(checkpoint['state_dict'])
encoder = system.model.eval()
return encoder
def load_pretrained_model(self):
base_dir = self.config.pretrain_model.exp_dir
checkpoint_name = self.config.pretrain_model.checkpoint_name
config_path = os.path.join(base_dir, 'config.json')
config_json = utils.load_json(config_path)
config = DotMap(config_json)
if self.config.model_params.resnet_small:
config.model_params.resnet_small = self.config.model_params.resnet_small
SystemClass = globals()[config.system]
system = SystemClass(config)
checkpoint_file = os.path.join(base_dir, 'checkpoints', checkpoint_name)
checkpoint = torch.load(checkpoint_file, map_location=self.device)
system.load_state_dict(checkpoint['state_dict'], strict=False)
encoder = system.model.eval()
return encoder, config
def process_config(config_path, override_dotmap=None):
config_json = load_json(config_path)
return _process_config(config_json, override_dotmap=override_dotmap)
# this will load both the weights and memory bank
agent.load_checkpoint('checkpoint.pth.tar',
pre_checkpoint_dir,
load_memory_bank=True,
load_model=True,
load_optim=True,
load_epoch=True)
try:
agent.run()
agent.finalise()
except KeyboardInterrupt:
pass
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('config', type=str, default='path to config file')
args = parser.parse_args()
config_json = load_json(args.config)
pre_checkpoint_dir = None
if config_json['pretrained_exp_dir'] is not None:
print("NOTE: found pretrained model...continue training")
pre_checkpoint_dir = os.path.join(config_json['pretrained_exp_dir'],
'checkpoints')
run(args.config, pre_checkpoint_dir)
def load_samples(samples_file):
return load_json(samples_file)
args = parser.parse_args()
OUT_IMG_DIR = f"/mnt/fs5/wumike/reference/pretrain/{args.dataset}/vaegan_image"
OUT_TXT_DIR = f"/mnt/fs5/wumike/reference/pretrain/{args.dataset}/vaegan_text"
if not os.path.isdir(OUT_IMG_DIR):
os.makedirs(OUT_IMG_DIR)
if not os.path.isdir(OUT_TXT_DIR):
os.makedirs(OUT_TXT_DIR)
config_path = os.path.join(MODEL_DIR, 'config.json')
checkpoint_dir = os.path.join(MODEL_DIR, 'checkpoints')
assert os.path.isfile(os.path.join(checkpoint_dir, 'model_best.pth.tar'))
config = load_json(config_path)
config['gpu_device'] = args.gpu_device
config = DotMap(config)
AgentClass = globals()[config.agent]
mvae = AgentClass(config)
mvae.load_checkpoint('model_best.pth.tar')
mvae._set_models_to_eval()
gpu_device = mvae.config.gpu_device
override_vocab = copy.deepcopy(mvae.train_dataset.vocab)
image_transforms = transforms.Compose([
transforms.Resize(32),
transforms.CenterCrop(32),
transforms.ToTensor(),
def load_config(config_file):
config = load_json(config_file)
return config
def process_config(config_path, override_dotmap=None, exp_name_suffix=None):
config_json = load_json(config_path)
return _process_config(config_json, override_dotmap=override_dotmap,
exp_name_suffix=exp_name_suffix)
