def __init__(self, mNet, y_activ_regression, mBGsub, depth):
n_leaves = 2**depth
x_activ = mNet.activation_volume
size_last_convolution = mNet.nb_activations
H_ds, W_ds = x_activ.shape[2], x_activ.shape[3]
batch_size = x_activ.shape[0]
#get foreground/background proba
p_fb = mBGsub.p_fb
p_foreground = p_fb[:, 0, :, :].reshape((batch_size, 1, H_ds, W_ds))
#Decision Flat 2 layers
#Decision tree
self.params_regression = RegFcts.init_all_tree_params(
depth, size_last_convolution) #[w_1,b_1]
#defining all decision nodes, breadth first, storing all the probas tensors in a list
tree_probas_list = RegFcts.create_tree_probas_list(
x_activ, p_foreground, self.params_regression, depth, batch_size)
# Put Output in the leaves
p_leaves = []
for l in range(0, n_leaves):
p_leaves.append(tree_probas_list[pow(2, depth) - 1 + l])
#Define objects
self.p_leaves = p_leaves
self.n_leaves = n_leaves
self.size_last_convolution = size_last_convolution
self.depth = depth
def get_random_regression_params(self):
random_params_regression = RegFcts.random_FC3_params(
self.size_last_convolution,
self.n_leaves,
hidden_0_size=300,
hidden_1_size=80)
return random_params_regression
def init_regression_params(self):
self.params_regression = RegFcts.init_FC3_params(
self.size_last_convolution,
self.n_leaves,
hidden_0_size=300,
hidden_1_size=80)
return self.params_regression
def __init__(self, mNet, y_activ_regression, mBGsub, n_leaves, p_drop=0.0):
x_activ = mNet.activation_volume
size_last_convolution = mNet.nb_activations
H_ds, W_ds = x_activ.shape[2], x_activ.shape[3]
batch_size = x_activ.shape[0]
#get foreground/background proba
p_fb = mBGsub.p_fb
p_foreground = p_fb[:, 0, :, :].reshape((batch_size, 1, H_ds, W_ds))
#Decision Flat 2 hidden layers
self.params_regression = RegFcts.init_FC3_params(size_last_convolution,
n_leaves,
hidden_0_size=300,
hidden_1_size=80)
p_fin = RegFcts.decision_node_flat3(x_activ,
self.params_regression[0],
self.params_regression[1],
self.params_regression[2],
self.params_regression[3],
self.params_regression[4],
self.params_regression[5],
batch_size,
p_drop=p_drop)
p_fin = p_fin * T.repeat(p_foreground, n_leaves, axis=1)
# Put Output in the leaves
p_leaves = []
for l in range(0, n_leaves):
p_leaves.append(p_fin[:, l:l + 1, :, :])
#Define objects
self.p_leaves = p_leaves
self.n_leaves = n_leaves
self.size_last_convolution = size_last_convolution
def init_regression_params(self):
self.params_regression = RegFcts.init_FC_params(
self.size_last_convolution, self.n_leaves)
return self.params_regression
def init_regression_params(self):
self.params_regression = RegFcts.init_all_tree_params(
self.depth, self.size_last_convolution)
return self.params_regression