target_net.load_state_dict(policy_net.state_dict())
# create optimizer
#opt = optim.RMSprop(policy_net.parameters(),
# lr=info["RMS_LEARNING_RATE"],
# momentum=info["RMS_MOMENTUM"],
# eps=info["RMS_EPSILON"],
# centered=info["RMS_CENTERED"],
# alpha=info["RMS_DECAY"])
opt = optim.Adam(policy_net.parameters(), lr=info['ADAM_LEARNING_RATE'])
if args.model_loadpath is not '':
# what about random states - they will be wrong now???
# TODO - what about target net update cnt
target_net.load_state_dict(model_dict['target_net_state_dict'])
policy_net.load_state_dict(model_dict['policy_net_state_dict'])
opt.load_state_dict(model_dict['optimizer'])
print("loaded model state_dicts")
if args.buffer_loadpath == '':
args.buffer_loadpath = args.model_loadpath.replace(
'.pkl', '_train_buffer.npz')
print("auto loading buffer from:%s" % args.buffer_loadpath)
try:
replay_memory.load_buffer(args.buffer_loadpath)
except Exception as e:
print(e)
print(
'not able to load from buffer: %s. exit() to continue with empty buffer'
% args.buffer_loadpath)
train_sim(start_step_number, start_last_save)
print("using randomized prior")
policy_net = NetWithPrior(policy_net, prior_net, info['PRIOR_SCALE'])
target_net = NetWithPrior(target_net, prior_net, info['PRIOR_SCALE'])
target_net.load_state_dict(policy_net.state_dict())
opt = optim.Adam(policy_net.parameters(), lr=info['ADAM_LEARNING_RATE'])
if load_model:
# what about random states - they will be wrong now???
# TODO - what about target net update cnt (TODO from johana)
target_net.load_state_dict(model_dict['target_net_state_dict'])
policy_net.load_state_dict(model_dict['policy_net_state_dict'])
opt.load_state_dict(model_dict['optimizer'])
print("loaded model state_dicts")
buffer_loadpath = info['model_loadpath'].replace('.pkl', '_train_buffer.npz')
print("auto loading buffer from:%s" %buffer_loadpath)
try:
replay_memory.load_buffer(buffer_loadpath)
except Exception as e:
print(e)
print('not able to load from buffer: %s. exit() to continue with empty buffer' %buffer_loadpath)
advice_net=None
if info['advice_flg']:
print('loading advice model from: %s' %info['advicemodel_loadpath'])
model_dict = torch.load(info['advicemodel_loadpath'])
advice_net = EnsembleNet(n_ensemble=info['N_ENSEMBLE'],
n_actions=env.num_actions,
network_output_size=info['NETWORK_INPUT_SIZE'][0],
num_channels=info['HISTORY_SIZE'], dueling=info['DUELING']).to(info['DEVICE'])
if info['PRIOR']:
advice_net = NetWithPrior(advice_net, prior_net, info['PRIOR_SCALE'])
action_getter = ActionGetter(n_actions=env.num_actions,
policy_net=policy_net,