def __init__(self, latent_dim):
super(LogicEncoder, self).__init__()
self.latent_dim = latent_dim
self.char_embedding = Parameter(
torch.Tensor(len(LIST_PREDICATES) + len(LIST_OP), latent_dim))
def new_gate():
lh = nn.Linear(self.latent_dim, self.latent_dim)
rh = nn.Linear(self.latent_dim, self.latent_dim)
return lh, rh
self.ilh, self.irh = new_gate()
self.lflh, self.lfrh = new_gate()
self.rflh, self.rfrh = new_gate()
self.ulh, self.urh = new_gate()
self.ix = nn.Linear(self.latent_dim, self.latent_dim)
self.fx = nn.Linear(self.latent_dim, self.latent_dim)
self.ux = nn.Linear(self.latent_dim, self.latent_dim)
self.and_transform_c, self.and_transform_h = new_gate()
self.or_transform_c, self.or_transform_h = new_gate()
self.cx, self.ox = new_gate()
self.oend = nn.Linear(self.latent_dim * 3, self.latent_dim)
weights_init(self)
def __init__(self, latent_dim):
super(GeneralDecoder, self).__init__(latent_dim)
self.start = "p"
self.ruleset = RULESET
self.top_act_w = Parameter(
torch.Tensor(
sum([len(self.ruleset[rule]) for rule in self.ruleset]),
latent_dim))
self.char_embedding = nn.Embedding(
sum([len(self.ruleset[rule]) for rule in self.ruleset]),
latent_dim)
self.tree_embed = LogicEncoder(self.latent_dim)
weights_init(self)
self.root = genExprTree(self.ruleset, "p")
def __init__(self, latent_dim):
super(IDecoder, self).__init__()
self.latent_dim = latent_dim
self.decision = nn.Linear(latent_dim, 3)
self.state_gru = nn.GRUCell(latent_dim, latent_dim)
self.and_embedding = Parameter(torch.Tensor(1, latent_dim))
self.or_embedding = Parameter(torch.Tensor(1, latent_dim))
self.value_pred_w1 = nn.Linear(self.latent_dim, self.latent_dim)
self.value_pred_w2 = nn.Linear(self.latent_dim, 1)
if cmd_args.attention:
self.first_att = nn.Linear(self.latent_dim, 1)
weights_init(self)
def __init__(self, latent_dim, num_node_feats, max_lv = 3):
super(EmbedMeanField, self).__init__()
self.latent_dim = latent_dim
self.num_node_feats = num_node_feats
self.max_lv = max_lv
self.w_n2l = nn.Linear(num_node_feats, latent_dim)
self.conv_param_list = []
self.merge_param_list = []
for i in range(self.max_lv):
self.conv_param_list.append(nn.Linear(latent_dim, NUM_EDGE_TYPES * latent_dim))
self.merge_param_list.append( nn.Linear(NUM_EDGE_TYPES * latent_dim, latent_dim) )
self.conv_param_list = nn.ModuleList(self.conv_param_list)
self.merge_param_list = nn.ModuleList(self.merge_param_list)
self.state_gru = nn.GRUCell(latent_dim, latent_dim)
weights_init(self)
def __init__(self, latent_dim):
super(LogicEncoder, self).__init__()
self.latent_dim = latent_dim
if RULESET:
self.char_embedding = Parameter(
torch.Tensor(sum([len(RULESET[rule]) for rule in RULESET]),
latent_dim))
def new_gate():
lh = nn.Linear(self.latent_dim, self.latent_dim)
rh = nn.Linear(self.latent_dim, self.latent_dim)
return lh, rh
self.ilh, self.irh = new_gate()
self.lflh, self.lfrh = new_gate()
self.rflh, self.rfrh = new_gate()
self.ulh, self.urh = new_gate()
self.i_gates = [
nn.Linear(self.latent_dim, self.latent_dim)
for _ in range(MAX_CHILD)
]
self.f_gates = [[
nn.Linear(self.latent_dim, self.latent_dim)
for _ in range(MAX_CHILD)
] for _ in range(MAX_CHILD)]
self.u_gates = [
nn.Linear(self.latent_dim, self.latent_dim)
for _ in range(MAX_CHILD)
]
self.ix = nn.Linear(self.latent_dim, self.latent_dim)
self.fx = nn.Linear(self.latent_dim, self.latent_dim)
self.ux = nn.Linear(self.latent_dim, self.latent_dim)
self.cx, self.ox = new_gate()
self.oend = nn.Linear(self.latent_dim * 3, self.latent_dim)
weights_init(self)
def __init__(self, latent_dim):
super(AssertAwareDecoder, self).__init__(latent_dim)
self.tree_grow_decision = nn.Linear(latent_dim, 2)
self.top_act_w = Parameter( torch.Tensor(3, latent_dim) )
weights_init(self)
def __init__(self, latent_dim):
super(AssertAwareDecoder, self).__init__(latent_dim)
self.tree_grow_decision = nn.Linear(latent_dim, 2)
weights_init(self)