def experiment(args):
track_local_dir = os.path.join(args.logroot, args.experimentname)
if args.remote:
track_remote_dir = os.path.join(args.remote, args.projectname,
args.experimentname)
else:
track_remote_dir = None
with track.trial(track_local_dir, track_remote_dir, param_map=vars(args)):
track.debug("Starting trial")
do_training(args)
raise Exception('num_gpu should be positive divisor of batch_size')
if mode == "predict":
data_train, args = load_data(args)
elif mode == "train" or mode == "load":
data_train, data_val, args = load_data(args)
# log current config
config_logger = ConfigLogger(log)
config_logger(args.config)
model_loaded, model_num_epoch = load_model(args, contexts, data_train)
# if mode is 'train', it trains the model
if mode == 'train':
data_names = [x[0] for x in data_train.provide_data]
label_names = [x[0] for x in data_train.provide_label]
module = mx.mod.Module(model_loaded, context=contexts, data_names=data_names, label_names=label_names)
do_training(args=args, module=module, data_train=data_train, data_val=data_val)
# if mode is 'load', it loads model from the checkpoint and continues the training.
elif mode == 'load':
do_training(args=args, module=model_loaded, data_train=data_train, data_val=data_val, begin_epoch=model_num_epoch+1)
# if mode is 'predict', it predict label from the input by the input model
elif mode == 'predict':
# predict through data
model_loaded.bind(for_training=False, data_shapes=data_train.provide_data,
label_shapes=data_train.provide_label)
max_t_count = args.config.getint('arch', 'max_t_count')
eval_metric = STTMetric(batch_size=batch_size, num_gpu=num_gpu, seq_length=max_t_count)
model_loaded.score(eval_data=data_train, num_batch=None, eval_metric=eval_metric, reset=True)
def train_it(self, fuck_off=None):
train.do_training()
return "trained!"