def forward(self, x, m):
"""
NOTE: this method works only with subclasses, which initialize main_net
"""
x = m * self.main_net(x)
if isinstance(self, BaseClassifier):
out = Sigmoid()(x)
out = out.view(-1, 1)
return None, out
def forward(self, x, m):
x = self.main_net(x)
x_det = torch.transpose(m * self.detector(x), 2, 1)
w = Softmax(dim=2)(x_det)
w = WeightsDropout(p=self.dropout)(w)
x = torch.bmm(w, x)
out = self.estimator(x)
if isinstance(self, BaseClassifier):
out = Sigmoid()(out)
out = out.view(-1, 1)
return w, out
def forward(self, x: torch.Tensor,
m: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Feed forward input data.
Parameters
----------
x: torch.Tensor
m: torch.Tensor
Returns
--------
Tuple with weights of conformers and tensor
of shape Nmol*1, where Nmol is the number of molecules. The tensor is final output y, but it needs to be passed
to sigmoid to obtain final class probabilities in case of classification (recall, this classs shouldnt be called directly,
call regressor/classifier subclass to obtain final y).
Examples
--------
>> > import torch
>> > import numpy as np
>> > from torch import randn
>> > from miqsar.estimators.attention_nets import AttentionNet
>> > x_train = randn((3, 3, 3))
>> > at_net = AttentionNet(ndim=(x_train[0].shape[-1], 4, 6, 4), det_ndim = (4,4), init_cuda=False)
>> > _, m = at_net.add_padding(x_train)
>> > m = torch.from_numpy(m.astype('float32'))
>> > _ = at_net.forward(x_train, m) # (assign result to a variable to supress std output)
"""
x = self.main_net(x)
x_det = torch.transpose(m * self.detector(x), 2, 1)
w = nn.functional.gumbel_softmax(x_det,
tau=self.instance_dropout,
dim=2)
x = torch.bmm(w, x)
out = self.estimator(x)
if isinstance(self, BaseClassifier):
out = Sigmoid()(out)
out = out.view(-1, 1)
return w, out
def forward(self, x, m):
x = m * self.main_net(x)
if isinstance(self, BaseClassifier):
out = Sigmoid()(x)
out = out.view(-1, 1)
return None, out