def forward(self, x, tc):
#def forward(self, x):
tc = F.dropout3d(self.bn_0tc(F.relu(self.c0tc(tc))), p=0.1)
x = F.dropout3d(self.bn_0(F.relu(self.c0(x))), p=0.1)
x = F.dropout3d(self.bn_02(F.relu(self.c02(x))), p=0.2)
x = x + tc
x = F.dropout3d(self.bn_1(F.relu(self.conv_1(x))), p=0.2)
"""tc = F.dropout3d(self.bn_0tc(F.relu(self.c0tc(tc))), p=0.1)
tc = F.dropout3d(self.bn_02tc(F.relu(self.c02tc(tc))), p=0.2)
tc = F.dropout3d(self.bn_1tc(F.relu(self.conv_1tc(tc))), p=0.2)"""
s1 = x
x = F.dropout3d(self.bn_2(F.relu(self.conv_2(x))), p=0.3)
s2 = x
x = F.dropout3d(self.bn_3(F.relu(self.conv_3(x))), p=0.3)
s3 = x
#x = F.dropout2d(self.bn_4(F.relu(self.conv_4(x))), p=0.3)
s3 = F.dropout3d(self.bn_5(F.relu(self.upsample_1(s3))), p=0.3)
s3 = s3 + s2
s3 = F.dropout3d(self.bn_6(F.relu(self.upsample_2(s3))), p=0.3)
s3 = s3 + s1
s3 = F.dropout3d(self.bn_7(F.relu(self.upsample_3(s3))), p=0.3)
s3 = F.dropout3d(self.bn_8(F.sigmoid(self.upsample_4(s3))), p=0.2)
return s3
def forward(self, input):
n, c, h, w, d = input.shape
x_volumes = x2d_to_volumes(input)
out_2d, fea_2d = self.denseunet_2d(x_volumes)
out_3d, fea_3d = dim_tran(out_2d) * 250, dim_tran(fea_2d)
x_3d = torch.cat((input, out_3d), 1)
x_3d = F.relu(self.rn(x_3d))
x_3d = self.up1(x_3d, self.sfs[3].features)
x_3d = self.up2(x_3d, self.sfs[2].features)
x_3d = self.up3(x_3d, self.sfs[1].features)
x_3d = self.up4(x_3d, self.sfs[0].features)
x_out = F.upsample(x_3d, size=(h, w, d), mode='trilinear')
x_out = self.conv1(x_out)
x_out = F.dropout3d(x_out, p=0.3)
x_out = F.relu(self.bn1(x_out))
x_out = x_out + fea_3d
x_out = self.conv2(x_out)
x_out_dropout = F.dropout3d(x_out, p=0.1)
x_out_bn = F.relu(self.bn2(x_out_dropout))
final_result = self.conv3(x_out_bn)
return final_result
def forward(self, x, y):
if self.drop_out:
x = F.dropout3d(x)
x = F.relu(self.up_bn(self.up_conv(x)))
y = F.dropout3d(y)
x = torch.cat([x, y], dim=1)
a = x
for i in self.layer:
x = i(x)
x = F.relu(torch.add(x, a))
return x
def forward(self, x):
# print(self.training)
out = F.dropout3d(self.bn1(F.leaky_relu(self.conv1(x))),
p=0.4,
training=self.training)
out = F.dropout3d(self.bn2(F.leaky_relu(self.conv2(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn3(F.leaky_relu(self.conv3(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn4(F.leaky_relu(self.conv4(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn5(F.leaky_relu(self.conv5(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn6(F.leaky_relu(self.conv6(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn7(F.leaky_relu(self.conv7(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn8(F.leaky_relu(self.conv8(out))),
p=0.3,
training=self.training)
out = F.sigmoid(self.conv9(out))
return out
def decode(self, input_s, down_inputs):
for d, i in zip(self.up, down_inputs[::-1]):
d.to(self.device)
# Remember that pooling is optional
if self.pooling:
input_s = F.dropout3d(
d(
torch.cat(
(F.interpolate(input_s, size=i.size()[2:]), i),
dim=1)), self.dropout, self.training)
else:
input_s = F.dropout3d(d(torch.cat((input_s, i), dim=1)),
self.dropout, self.training)
return input_s
def forward(self, input):
out = self.conv1(self.relu1(self.batch_norm1(input)))
out = self.conv2(self.relu2(self.batch_norm2(out)))
if self.dropout > 0:
out = F.dropout3d(out, p=self.dropout, training=self.training)
out = torch.cat([input, out], dim=1)
return out
def forward(self, x):
new_feature = super(DenseLayer, self).forward(x)
if (self.drop_rate > 0):
new_feature = F.dropout3d(new_feature,
p=self.drop_rate,
training=self.training)
return torch.cat([x, new_feature], 1) #每一层denselayer的输出为该层的输入和输出的堆叠
def forward(self, inputs):
if self.training and self.p > 0.:
if self.method == "uniform":
# For step training when p decay to mini value.
if self.start_p > self.p:
self.start_p = self.p
self.init_value.uniform_(self.start_p, self.p)
else:
# Only take (0, self.p) of the gaussian curve.
self.init_value = self.init_value.normal_(
self.mean, self.std).clamp(min=0., max=self.p)
if self.dim == 1:
outputs = F.dropout(inputs,
self.init_value,
inplace=self.inplace)
elif self.dim == 2:
outputs = F.dropout2d(inputs,
self.init_value,
inplace=self.inplace)
else:
outputs = F.dropout3d(inputs,
self.init_value,
inplace=self.inplace)
return outputs
else:
return inputs
def forward(self, x):
x = self.conv1(x)
x = self.bn_op_1(x)
x = F.leaky_relu(x)
if self.dropout is True:
x = F.dropout3d(x, p=0.3)
x = self.conv2(x)
x = self.bn_op_2(x)
x = F.leaky_relu(x)
if self.dropout is True:
x = F.dropout3d(x, p=0.3)
return x
def forward(self, input):
feature = super(DenseASPPBlock, self).forward(input)
if self.drop_rate > 0:
feature = F.dropout3d(feature, p=self.drop_rate, training=self.training)
return feature
def encode(self, input_s):
# We need to keep track of the convolutional outputs, for the skip
# connections.
down_inputs = []
for c in self.down:
c.to(self.device)
input_s = F.dropout3d(c(input_s), self.dropout, self.training)
down_inputs.append(input_s)
# Remember that pooling is optional
if self.pooling:
input_s = F.max_pool3d(input_s, 2)
self.u.to(self.device)
input_s = F.dropout3d(self.u(input_s), self.dropout, self.training)
return down_inputs, input_s
def forward(self, im, features):
for c, p in zip(self.base_model.convlist, self.base_model.pooling):
c.to(self.device)
p.to(self.device)
im = p(c(im))
self.base_model.dropout = self.dropout
im = self.base_model.midconv(im)
drop = F.dropout3d(im, p=self.dropout, training=self.drop)
self.global_pooling.to(self.device)
x = self.global_pooling(drop).view(im.shape[:2])
for l in self.linear:
l.to(self.device)
x = l(x)
x = F.dropout(x, p=self.dropout, training=self.drop)
x = torch.cat((x,features.type_as(x)), dim=1)
self.out.to(self.device)
output = self.out(x)
if self.dropout <= 0.5:
output = F.relu(output)
else:
output = F.leaky_relu(output)
return output
def _add_noise(self, x: torch.Tensor) -> torch.Tensor:
if self.dropout_p > 0:
x = F.dropout3d(x, self.dropout_p, training=self.enable_dropout) if self.is_3d else \
F.dropout2d(x, self.dropout_p, training=self.enable_dropout)
if self.noise_lvl > 0:
x.add_(torch.randn_like(x.detach()) * self.noise_lvl)
return x
def _add_noise(self, x: torch.Tensor) -> torch.Tensor:
if self.dropout_prob > 0:
x = F.dropout3d(x, self.dropout_prob, training=self.enable_dropout, inplace=self.inplace) if self.dim == 3 else \
F.dropout2d(x, self.dropout_prob, training=self.enable_dropout, inplace=self.inplace) if self.dim == 2 else \
F.dropout(x, self.dropout_prob, training=self.enable_dropout, inplace=self.inplace)
if self.noise_lvl > 0:
x = x + (torch.randn_like(x.detach()) * self.noise_lvl)
return x
def forward(self, x):
if not self.first:
x = self.maxpool(x)
x = self.bn1(self.conv1(x))
y = self.relu(self.bn2(self.conv2(x)))
if self.dropout > 0:
y = F.dropout3d(y, self.dropout)
y = self.bn3(self.conv3(x))
return self.relu(x + y)
def forward(self, x):
out = self.conv(x)
if self.drop_rate > 0:
out = F.dropout3d(out,
self.drop_rate,
training=self.training,
inplace=True)
out = self.nonlin(self.norm(out))
return torch.cat([x, out], dim=1)
def forward(self, *prev_features):
bn_function = _bn_function_factory(self.norm1, self.relu1, self.conv1)
if self.efficient and any(prev_feature.requires_grad for prev_feature in prev_features):
bottleneck_output = cp.checkpoint(bn_function, *prev_features)
else:
bottleneck_output = bn_function(*prev_features)
new_features = self.conv2(self.relu2(self.norm2(bottleneck_output)))
if self.drop_rate > 0:
new_features = F.dropout3d(new_features, p=self.drop_rate, training=self.training)
return new_features
def forward(self, x, device):
# print('1. ', x.shape)
h = self.conv1(x, device)
# print('2. ', h.shape)
h = relu(h)
h = self.pool1(h)
# print('3. ', h.shape)
h = self.conv2(h, device)
# print('4. ', h.shape)
h = relu(h)
h = self.pool2(h)
# print('5. ', h.shape)
h = self.conv3(h, device)
# print('6. ', h.shape)
h = relu(h)
h = self.conv4(h, device)
# print('7. ', h.shape)
h = relu(h)
h = self.conv5(h, device)
# print('8. ', h.shape)
h = relu(h)
h = self.pool3(h)
# print('9. ', h.shape)
h = self.pool4(h)
# print('10. ', h.shape)
_shape = h.shape
h = h.view(-1, _shape[1] * _shape[2] * _shape[3] * _shape[4])
# print('11. ', h.shape)
h = dropout3d(h, p=0.2)
h = self.fc1(h)
# print('12. ', h.shape)
h = relu(h)
h = dropout3d(h, p=0.2)
h = self.fc2(h)
# print('13. ', h.shape)
h = relu(h)
y = self.fc3(h)
# print('14. ', h.shape)
return y
def forward(self, x):
a = F.relu(self.down_bn(self.down_conv(x)))
if self.drop_out:
x = F.dropout3d(a)
else:
x = a
for i in self.layer:
x = i(x)
x = F.relu(torch.add(x, a))
return x
def forward(self, x):
x = F.dropout3d(self.bn_0(F.relu(self.c0(x))), p=0.1)
x = F.dropout3d(self.bn_01(F.relu(self.c01(x))), p=0.2)
x = F.dropout3d(self.bn_02(F.relu(self.c02(x))), p=0.2)
x = F.dropout3d(self.bn_1(F.relu(self.conv_1(x))), p=0.2)
s1 = x
x = F.dropout3d(self.bn_2(F.relu(self.conv_2(x))), p=0.3)
s2 = x
x = F.dropout3d(self.bn_3(F.relu(self.conv_3(x))), p=0.3)
s3 = x
#x = F.dropout2d(self.bn_4(F.relu(self.conv_4(x))), p=0.3)
s3 = F.dropout3d(self.bn_5(F.relu(self.upsample_1(s3))), p=0.3)
s3 = s3 + s2
s3 = F.dropout3d(self.bn_6(F.relu(self.upsample_2(s3))), p=0.3)
s3 = s3 + s1
s3 = F.dropout3d(self.bn_7(F.relu(self.upsample_3(s3))), p=0.3)
s3 = F.dropout3d(self.bn_8((self.upsample_4(s3))), p=0.2)
return s3
def forward(self):
a = torch.randn(8, 4)
b = torch.randn(8, 4, 4, 4)
c = torch.randn(8, 4, 4, 4, 4)
return len(
F.dropout(a),
F.dropout2d(b),
F.dropout3d(c),
F.alpha_dropout(a),
F.feature_alpha_dropout(c),
)
def forward(self, x):
down_list = []
for c, p in zip(self.convlist, self.pooling):
c.to(self.device)
down = c(x)
drop = F.dropout3d(down, p=self.dropout, training=self.drop)
down_list.append(drop)
p.to(self.device)
x = p(drop)
x = self.midconv(x)
x = F.dropout3d(x, p=self.dropout, training=self.drop)
for d, prev in zip(self.deconvlist, down_list[::-1]):
interp = F.interpolate(x, size=prev.shape[2:])
d.to(self.device)
x = d(torch.cat((prev, interp), dim=1))
x = F.dropout3d(x, p=self.dropout, training=self.drop)
self.out.to(self.device)
output = self.out(x)
return output
def forward(self, x):
# print(x.shape)
x = x.permute(0, 1, 4, 2, 3)
if not self.first:
x = self.maxpool(x.squeeze(0))
x = x.unsqueeze(0)
x = x.permute(0, 1, 3, 4, 2)
x = self.bn1(self.conv1(x))
y = self.relu(self.bn2(self.conv2(x)))
if self.dropout > 0:
y = F.dropout3d(y, self.dropout)
y = self.bn3(self.conv3(x))
return self.relu(x + y)
def forward(self, x):
#print("Down---forward", self.meta_loss, self.meta_step_size, self.stop_gradient)
if not self.first:
x = maxpool3D(x, kernel_size=2)
#layer 1 conv, bn
#print(">>>" ,x, self.conv1.weight, self.conv1.bias)
x = conv3d(x,
self.conv1.weight,
self.conv1.bias,
meta_loss=self.meta_loss,
meta_step_size=self.meta_step_size,
stop_gradient=self.stop_gradient)
x = self.bn1(x)
#layer 2 conv, bn, relu
y = conv3d(x,
self.conv2.weight,
self.conv2.bias,
meta_loss=self.meta_loss,
meta_step_size=self.meta_step_size,
stop_gradient=self.stop_gradient)
y = self.bn2(y)
y = relu(y)
#droupout if required
if self.dropout > 0:
y = F.dropout3d(y, self.dropout)
#layer 3 conv, bn
z = conv3d(y,
self.conv3.weight,
self.conv3.bias,
meta_loss=self.meta_loss,
meta_step_size=self.meta_step_size,
stop_gradient=self.stop_gradient)
z = self.bn3(z)
z = relu(z) #was not there
#final relu
#k = relu(x + z)
return z
def forward(self, x):
# print(self.training)
out1 = F.dropout3d(self.bn1(F.leaky_relu(self.conv1(x))),
p=0.4,
training=self.training)
# print('out1 ', out1.shape)
out2 = F.dropout3d(self.bn2(F.leaky_relu(self.conv2(out1))),
p=0.3,
training=self.training)
# print('out2 ', out2.shape)
out3 = F.dropout3d(self.bn3(F.leaky_relu(self.conv3(out2))),
p=0.3,
training=self.training)
# print('out3 ', out3.shape)
out4 = F.dropout3d(self.bn4(F.leaky_relu(self.conv4(out3))),
p=0.3,
training=self.training)
# print('out4 ', out4.shape)
out5 = F.dropout3d(self.bn5(F.leaky_relu(self.conv5(out4))),
p=0.3,
training=self.training)
# print('out5 ', out5.shape)
out6 = F.dropout3d(self.bn6(F.leaky_relu(self.conv6(out5))),
p=0.3,
training=self.training)
# print('out6 ', out6.shape)
out6 = torch.cat([out4, out6], dim=1)
out7 = F.dropout3d(self.bn7(F.leaky_relu(self.conv7(out6))),
p=0.3,
training=self.training)
out7 = torch.cat([out3, out7], dim=1)
# print('out7 ', out7.shape)
out8 = F.dropout3d(self.bn8(F.leaky_relu(self.conv8(out7))),
p=0.3,
training=self.training)
out8 = torch.cat([out2, out8], dim=1)
# print('out8 ', out8.shape )
out = F.sigmoid(self.conv9(out8))
return out
def forward(self, x):
# print(self.training)
out = F.dropout3d(self.bn1(F.leaky_relu(self.conv1(x))),
p=0.4,
training=self.training)
out = F.dropout3d(self.bn2(F.leaky_relu(self.conv2(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn3(F.leaky_relu(self.conv3(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn4(F.leaky_relu(self.conv4(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn5(F.leaky_relu(self.conv5(out))),
p=0.3,
training=self.training)
# out2 = self.stn(out2)
# out2 = F.dropout3d(self.m_bn6(F.leaky_relu(self.m_conv6(out))), p=0.3, training=self.training)
# out2 = F.dropout3d(self.m_bn7(F.leaky_relu(self.m_conv7(out2))), p=0.3, training=self.training)
# out2 = F.dropout3d(self.m_bn8(F.leaky_relu(self.m_conv8(out2))), p=0.3, training=self.training)
# out_rois = F.sigmoid(self.conv9(out2))
out = F.dropout3d(self.bn6(F.leaky_relu(self.conv6(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn7(F.leaky_relu(self.conv7(out))),
p=0.3,
training=self.training)
out = F.dropout3d(self.bn8(F.leaky_relu(self.conv8(out))),
p=0.3,
training=self.training)
out_frames = F.sigmoid(self.conv9(out))
print('out_frames.shape', out_frames.shape)
out_rois = self.stn(out_frames)
return out_frames, out_rois
def forward(self, x, meta_loss, meta_step_size, stop_gradient):
if not self.first:
x = maxpool3D(x, kernel_size=2)
#layer 1 conv, bn
x = conv3d(x,
self.conv1.weight,
self.conv1.bias,
meta_loss=meta_loss,
meta_step_size=meta_step_size,
stop_gradient=stop_gradient)
x = self.bn1(x)
#layer 2 conv, bn, relu
y = conv3d(x,
self.conv2.weight,
self.conv2.bias,
meta_loss=meta_loss,
meta_step_size=meta_step_size,
stop_gradient=stop_gradient)
y = self.bn2(y)
y = relu(y)
#droupout if required
if self.dropout > 0:
y = F.dropout3d(y, self.dropout)
#layer 3 conv, bn
z = conv3d(y,
self.conv3.weight,
self.conv3.bias,
meta_loss=meta_loss,
meta_step_size=meta_step_size,
stop_gradient=stop_gradient)
z = self.bn3(z)
z = relu(z) #was not there
return z
def test_dropout3d(self):
inp = torch.randn(16, 8, 32, 64, 64, device='cuda', dtype=self.dtype)
output = F.dropout3d(inp, p=0.5, training=True, inplace=False)
def forward(self, input):
inputShape = input.shape
return F.dropout3d(
input.reshape((inputShape[0], -1, 1, 1, inputShape[-1])), self.p,
self.training, self.inplace).reshape(inputShape)
def forward(self, x):
result = f.dropout3d(
x.reshape((x.shape[0], x.shape[1] * x.shape[2], x.shape[3], 1, 1,
x.shape[4])), self.p, self.training, self.inplace)
return result.reshape(
(result.shape[0], x.shape[1], x.shape[2], x.shape[3], x.shape[4]))
