net_params=net_params,
binarize=0.000)
paddle_model.set_parameters(params)
# ball_model_net_params_path = 'ObjectRecognition/net_params/attn_base.json'
ball_model_net_params_path = 'ObjectRecognition/net_params/attn_softmax.json'
net_params = json.loads(open(ball_model_net_params_path).read())
params = load_param('results/cmaes_soln/focus_self/ball.pth')
# params = load_param('ObjectRecognition/models/atari/42531_2_smooth_3_2.pth')
ball_model = ModelFocusCNN(image_shape=(84, 84),
net_params=net_params,
binarize=0.0)
ball_model.set_parameters(params)
model = ModelCollectionDAG()
model.add_model('Paddle',
paddle_model, [],
augment_fn=util.RemoveMeanMemory(nb_size=(8, 8)))
f1 = util.LowIntensityFiltering(6.0)
f2 = util.JumpFiltering(3, 0.05)
def f(x, y):
return f2(x, f1(x, y))
# model.add_model('train', r_model, ['premise'], augment_pt=f)
model.add_model('Ball', ball_model, ['Paddle'],
augment_pt=f) #,augment_pt=util.JumpFiltering(2, 0.05))
####
if args.true_environment:
model = None
print(args.true_environment, args.env)
if args.env == 'SelfPusher':
if args.true_environment:
params = load_param('ObjectRecognition/models/paddle_bin_long_smooth_2.pth')
paddle_model = ModelFocusCNN(
image_shape=(84, 84),
net_params=net_params,
)
paddle_model.set_parameters(params)
ball_model_net_params_path = 'ObjectRecognition/net_params/two_layer.json'
net_params = json.loads(open(ball_model_net_params_path).read())
params = load_param('ObjectRecognition/models/ball.npy')
ball_model = ModelFocusCNN(
image_shape=(84, 84),
net_params=net_params,
)
ball_model.set_parameters(params)
model = ModelCollectionDAG()
model.add_model('Paddle', paddle_model, [], augment_fn=util.RemoveMeanMemory(nb_size=(3, 9)))
model.add_model('Ball', ball_model, ['Paddle'])
####
true_environment = FocusEnvironment(model)
dataset_path = args.record_rollouts
changepoint_path = args.changepoint_dir
option_chain = OptionChain(true_environment, args.changepoint_dir, args.train_edge, args)
reward_paths = glob.glob(os.path.join(option_chain.save_dir, "*rwd.pkl"))
print(reward_paths)
reward_paths.sort(key=lambda x: int(x.split("__")[2]))
head, tail = get_edge(args.train_edge)
reward_classes = [load_from_pickle(pth) for pth in reward_paths]
# train_models = MultiOption(1, BasicModel)