'stride': [1, 1],
'padding': [2, 4],
'activation_fn': ['ReLU', 'Tanh']
}
net_params_2 = {
'filter': 2,
'channel': [10, 1],
'kernel_size': [3, 5],
'stride': [1, 1],
'padding': [2, 4],
'activation_fn': ['ReLU6', 'Tanh']
}
model_1 = ModelFocusCNN((84, 84), net_params=net_params_1)
model_2 = ModelFocusCNN((84, 84), net_params=net_params_2)
model_boost = ModelFocusBoost(
LinearCPD(np.pi / 4),
model_1,
model_2,
train_flags=[True, False],
)
# parameters
print(model_boost.count_parameters())
ones = np.arange(model_boost.count_parameters())
model_boost.set_parameters(ones)
print(list(model_boost.parameters()))
# forward
out = model_boost.forward(torch.rand([100, 1, 84, 84]))
print(out, out.shape)
dataset = DatasetAtari(
'BreakoutNoFrameskip-v4', # atari game name
actor, # mock actor
n_state=args.n_state, # set max number of states
save_path='results', # save path for gym
binarize=args.binarize, # binarize image to 0 and 1
)
"""
Changepoint Detector
- specify changepoint detector which fits the dynamic of the object
"""
if args.champ:
cpd = CHAMPDetector('premise->object', CHAMP_params)
else:
logger.info('using simple linear changepoint detector')
cpd = LinearCPD(np.pi / 4.0)
"""
Model Template & Constructor
"""
model = ModelCollectionDAG()
net_params = json.loads(open(args.net).read())
if args.model_type == 'focus':
train_model = ModelFocusCNN(
image_shape=dataset.frame_shape,
net_params=net_params,
use_prior=args.prior,
argmax_mode=args.argmax_mode,
)
elif args.model_type == 'attn':
train_model = ModelAttentionCNN(
image_shape=dataset.frame_shape,
def __init__(self, *micp_losses, **kwargs):
self.micp_losses = list(micp_losses)
self.agg_fn = kwargs.get('agg_fn', sum)
self.cp_detector = kwargs.get('cp_detector', LinearCPD(np.pi / 4))