def __init__(self, args):
self.opt = Bunch(json.load(open(args.experiment, 'r')))
self.opt.resume = args.resume
self.global_step = 0
self.epoch = 0
self.best_score = 0.0
self.start_time = time()
# LOGGING
self.result_path = os.path.join(self.opt.exp_dir,
self.opt.result_folder,
self.opt.run_name)
util.create_folder(self.result_path)
util.create_folder(os.path.join(self.result_path, 'checkpoints'))
self.logger = TensorboardLogger(
os.path.join(self.opt.exp_dir, self.opt.result_folder,
"tensorboard/" + self.opt.run_name))
self.std_logger = util.get_std_logger(
'results', os.path.join(self.result_path, 'stdout.log'))
self.visualizer = QGenVisualizer(
os.path.join(self.result_path,
"{}.html".format(self.opt.run_name)),
os.path.join(self.opt.exp_dir, "Utils/css/question.css"))
self.get_data_loaders()
# MODEL
self.model = QuestionGenerator(self.opt).to(device)
self.model.apply(util.weights_init)
self.vqa = load_model(self.opt.vqa_path, AttentionVQA)
self.captioner = load_model(self.opt.cap_path, AttentionCaptioner)
self.vqa.eval()
if self.opt.cap_eval:
self.captioner.eval()
# OPTIMIZER
self.optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
self.model.parameters()),
lr=self.opt.lr)
if self.opt.resume is not None:
self.resume(self.opt.resume)
self.arang_vector = torch.arange(self.opt.batch_size,
dtype=torch.long).to(device)
def __init__(self, args):
__metaclass__ = ABCMeta
self.opt = Bunch(json.load(open(args.experiment, 'r')))
self.opt.resume = args.resume
self.set_seed(self.opt.seed)
self.chunk, self.collection_epoch, self.cap_epoch, self.collection_steps, self.cap_steps = 0, 0, 0, 0, 0
self.start_time = time()
# LOGGING
self.result_path = os.path.join(self.opt.exp_dir,
self.opt.result_folder,
self.opt.run_name)
util.create_folder(self.result_path)
self.logger = TensorboardLogger(
os.path.join(self.opt.exp_dir, self.opt.result_folder,
"tensorboard/" + self.opt.run_name))
self.std_logger = util.get_std_logger(
'results', os.path.join(self.result_path, 'stdout.log'))
self.Cvisualizer = CaptionVisualizer(
os.path.join(self.result_path,
"{}_cap.html".format(self.opt.run_name)),
os.path.join(self.opt.exp_dir, "Utils/css/caption.css"))
self.init_files_dirs()
# Auto-eval scorers
self.scorers = get_scorers(self.opt.cached_tokens)
self.get_data_loaders()
self.captioner = load_model(self.opt.cap_path, AttentionCaptioner)
self.c_optimizer = torch.optim.Adam([{
'params':
filter(lambda p: p.requires_grad, self.captioner.parameters()),
'lr':
self.opt.c_lr,
'mod_name':
'cap'
}])
if self.opt.resume is not None:
self.resume(self.opt.resume)
def __init__(self, args):
self.opt = Bunch(json.load(open(args.experiment, 'r')))
self.opt.resume = args.resume
self.global_step = 0
self.epoch = 0
self.best_score = 0.0
self.start_time = time()
# LOGGING
self.result_path = os.path.join(self.opt.exp_dir,
self.opt.result_folder,
self.opt.run_name)
util.create_folder(self.result_path)
util.create_folder(os.path.join(self.result_path, 'checkpoints'))
self.logger = TensorboardLogger(
os.path.join(self.opt.exp_dir, self.opt.result_folder,
"tensorboard/" + self.opt.run_name))
self.std_logger = util.get_std_logger(
'results', os.path.join(self.result_path, 'stdout.log'))
self.visualizer = CaptionVisualizer(
os.path.join(self.result_path,
"{}.html".format(self.opt.run_name)),
os.path.join(self.opt.exp_dir, "Utils/css/caption.css"))
self.get_data_loaders()
# MODEL
self.model = AttentionCaptioner(self.opt).to(device)
self.model.apply(util.weights_init)
# OPTIMIZER
self.optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
self.model.parameters()),
lr=self.opt.lr)
# Auto-eval scorers
self.scorers = get_scorers(self.opt.cached_tokens)
if self.opt.resume is not None:
self.resume(self.opt.resume)
def init_files_dirs(self):
base = self.result_path
self.lifelong_data_files = self.opt.train_files.split(',')
self.chunks = len(self.lifelong_data_files)
self.cap_checkpoint_path, self.collection_path, self.dm_checkpoint_path \
= [base+x for x in ["/ccheckpoints", "/collected_data", "/mcheckpoints"]]
for d in [
self.cap_checkpoint_path, self.collection_path,
self.dm_checkpoint_path
]:
util.create_folder(d)
self.collected_data_paths = [
os.path.join(self.collection_path,
'CollectedDataRound{}.p'.format(x + 1))
for x in range(self.chunks)
]
self.caption_data_files = [self.opt.warmup_file]
self.caption_data_files.extend([
os.path.join(self.collection_path,
'AccumDataRound{}.p'.format(x + 1))
for x in range(self.chunks)
])
self.chunks = len(self.lifelong_data_files)
self.caption_model_files = [
os.path.join(self.cap_checkpoint_path,
'chunk{}_best.pth'.format(x + 1))
for x in range(self.chunks)
]
self.decision_maker_model_files = [
os.path.join(self.dm_checkpoint_path,
'chunk{}_best.pth'.format(x + 1))
for x in range(self.chunks)
]