def __init__(self,
used_concepts,
input_dims,
hidden_dims,
parameter_resolution='deterministic',
vse_attribute_agnostic=False,
args=None):
super().__init__()
self.used_concepts = used_concepts
self.input_dims = input_dims
self.hidden_dims = hidden_dims
self.parameter_resolution = parameter_resolution
self.args = args
#pdb.set_trace()
for i, nr_vars in enumerate(
['attribute', 'relation', 'temporal', 'time']):
if nr_vars not in self.used_concepts:
continue
setattr(self, 'embedding_' + nr_vars,
concept_embedding.ConceptEmbedding(vse_attribute_agnostic))
tax = getattr(self, 'embedding_' + nr_vars)
rec = self.used_concepts[nr_vars]
for a in rec['attributes']:
tax.init_attribute(a, self.input_dims[1 + i],
self.hidden_dims[1 + i])
for (v, b) in rec['concepts']:
tax.init_concept(v, self.hidden_dims[1 + i], known_belong=b)
if nr_vars == 'time':
tax.exist_object = jacnn.LinearLayer(self.input_dims[1 + i],
1,
activation=None)
# TODO more complicated filter_in and out function
tax.filter_in = jacnn.LinearLayer(self.input_dims[1 + i],
128,
activation=None)
tax.filter_out = jacnn.LinearLayer(self.input_dims[1 + i],
128,
activation=None)
def __init__(self, input_dim, output_dim):
super().__init__()
self.input_dim = input_dim
self.output_dim = output_dim
self.map = jacnn.LinearLayer(input_dim, output_dim, activation=None)