def forward(self, input):
assert len(input) == self.n_directions
prepared_batches = None
if all(SLayer.is_prepared_batch(b) for b in input):
prepared_batches = input
elif all(SLayer.is_list_of_tensors(b) for b in input):
prepared_batches = [SLayer.prepare_batch(input_i, self.point_dim) for input_i in input]
else:
raise ValueError('Unrecognized input format! Expected list of Tensors or list of SLayer.prepare_batch outputs!')
batch_size = prepared_batches[0][0].size()[0]
assert all(prep_b[0].size()[0] == batch_size for prep_b in prepared_batches)
output = []
for i, sl in enumerate(self.slayers):
i_th_output = []
i_th_output.append(sl(prepared_batches[i]))
for j in range(1, self.n_neighbor_directions + 1):
i_th_output.append(sl(prepared_batches[i - j]))
i_th_output.append(sl(prepared_batches[(i + j) % self.n_directions]))
if self.n_directions > 0:
i_th_output = torch.stack(i_th_output, 1)
else:
i_th_output = output[0]
output.append(i_th_output)
return output
def __init__(self, subscripted_views):
super(MyModel, self).__init__()
self.subscripted_views = subscripted_views
self.transform = UpperDiagonalThresholdedLogTransform(0.1)
def get_init(n_elements):
transform = UpperDiagonalThresholdedLogTransform(0.1)
return transform(pers_dgm_center_init(n_elements))
self.dim_0 = SLayer(150, 2, get_init(150), torch.ones(150, 2) * 3)
self.dim_0_ess = SLayer(50, 1)
self.dim_1_ess = SLayer(50, 1)
self.slayers = [self.dim_0,
self.dim_0_ess,
self.dim_1_ess
]
self.stage_1 = []
stage_1_outs = [75, 25, 25]
for i, (n_in, n_out) in enumerate(zip([150, 50, 50], stage_1_outs)):
seq = nn.Sequential()
seq.add_module('linear_1', nn.Linear(n_in, n_out))
seq.add_module('batch_norm', nn.BatchNorm1d(n_out))
seq.add_module('drop_out_1', nn.Dropout(0.1))
seq.add_module('linear_2', nn.Linear(n_out, n_out))
seq.add_module('relu', nn.ReLU())
seq.add_module('drop_out_2', nn.Dropout(0.1))
self.stage_1.append(seq)
self.add_module('stage_1_{}'.format(i), seq)
linear_1 = nn.Sequential()
linear_1.add_module('linear_1', nn.Linear(sum(stage_1_outs), 200))
linear_1.add_module('batchnorm_1', torch.nn.BatchNorm1d(200))
linear_1.add_module('relu_1', nn.ReLU())
linear_1.add_module('linear_2', nn.Linear(200, 100))
linear_1.add_module('batchnorm_2', torch.nn.BatchNorm1d(100))
linear_1.add_module('drop_out_2', torch.nn.Dropout(0.1))
linear_1.add_module('relu_2', nn.ReLU())
linear_1.add_module('linear_3', nn.Linear(100, 50))
linear_1.add_module('batchnorm_3', nn.BatchNorm1d(50))
linear_1.add_module('relu_3', nn.ReLU())
linear_1.add_module('linear_4', nn.Linear(50, 5))
linear_1.add_module('batchnorm_4', nn.BatchNorm1d(5))
self.linear_1 = linear_1
def __init__(self,
n_directions,
n_elements,
point_dim,
n_neighbor_directions=0,
center_init=None,
sharpness_init=None):
super(SLayerPHT, self).__init__()
self.n_directions = n_directions
self.n_elements = n_elements
self.point_dim = point_dim
self.n_neighbor_directions = n_neighbor_directions
self.slayers = [SLayer(n_elements, point_dim, center_init, sharpness_init)
for i in range(n_directions)]
for i, l in enumerate(self.slayers):
self.add_module('sl_{}'.format(i), l)