def train(train_cfg, model_cfg, common_cfg, dataflow):
epochs = train_cfg.epochs
assert epochs, epochs
epoch_size = train_cfg.epoch_size
assert epoch_size, epoch_size
config = TrainConfig(
model=ModelWrapSingle(train_cfg, model_cfg, common_cfg),
dataflow=dataflow,
#data=my_inputsource, # alternatively, use a customized InputSource
#callbacks=[...], # some default callbacks are automatically applied
# some default monitors are automatically applied
steps_per_epoch=
epoch_size, # default to the size of your InputSource/DataFlow
max_epoch=epochs)
print("Create trainer")
trainer = SimpleTrainer()
print("Run train")
launch_train_with_config(config, trainer)
screen_dims=IMAGE_SIZE,
spacing=SPACING)
NUM_ACTIONS = init_player.action_space.n
num_validation_files = init_player.files.num_files
if args.task != 'train':
assert args.load is not None
pred = OfflinePredictor(PredictConfig(
model=Model(),
session_init=get_model_loader(args.load),
input_names=['state'],
output_names=['Qvalue']))
if args.task == 'play':
t0 = time.time()
play_n_episodes(get_player(directory=data_dir,
files_list=test_list, viz=0.01,
saveGif=args.saveGif,
saveVideo=args.saveVideo),
pred, num_validation_files)
t1 = time.time()
print(t1-t0)
elif args.task == 'eval':
eval_model_multithread(pred, EVAL_EPISODE, get_player)
else:
logger.set_logger_dir(logger_dir) # todo: variable log dir
config = get_config()
if args.load:
config.session_init = get_model_loader(args.load)
launch_train_with_config(config, SimpleTrainer())
# prevent learning in the first epoch
# MemInitHyperParamSetter('learning_rate_mask',(0,1)),
# controls learning rate as a function of epoch
HyperParamSetterWithFunc('learning_rate', learning_rate_fun),
# GraphProfiler()
# PeakMemoryTracker()
# GPUUtilizationTracker(),
],
steps_per_epoch=steps_per_epoch,
max_epoch=200000,
# first time load model from checkpoint and reset GRU state
session_init=ChainInit([TryResumeTraining()]), #,ResetInit(model)])
# session_config=tf.ConfigProto(log_device_placement=True) #config_gpus(1)
)
trainer = SimpleTrainer()
# with tf.contrib.tfprof.ProfileContext(logger.get_logger_dir()) as pctx:
launch_train_with_config(traincfg, trainer)
################# JUNK ###############
# config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
# workaround
# with tf.Session(config=config):
# pass
# Pin = tf.placeholder(tf.float32, shape=P_py.shape)
# backp = Backproj(vlen, 10.0, 'c3')
# l = backp.score(Pin)
# class Invert(Can):