def extract(self, x, label):
conv1 = self.conv_block1(x)
conv2 = self.conv_block2(conv1)
conv3 = self.conv_block3(conv2)
conv4 = self.conv_block4(conv3)
conv5 = self.conv_block5(conv4)
fconv = self.fconv_block(conv5)
if self.use_norm_weights:
fconv_norm = l2_norm(fconv)
conv3_norm = l2_norm(conv3)
conv4_norm = l2_norm(conv4)
else:
fconv_norm = fconv
conv3_norm = conv3
conv4_norm = conv4
if self.weighted_mask:
fconv_pooled = weighted_masked_embeddings(fconv_norm.shape, label,
fconv_norm, self.n_classes)
conv3_pooled = weighted_masked_embeddings(conv3_norm.shape, label,
conv3_norm, self.n_classes)
conv4_pooled = weighted_masked_embeddings(conv4_norm.shape, label,
conv4_norm, self.n_classes)
else:
fconv_pooled = masked_embeddings(fconv_norm.shape, label, fconv_norm,
self.n_classes)
conv3_pooled = masked_embeddings(conv3_norm.shape, label, conv3_norm,
self.n_classes)
conv4_pooled = masked_embeddings(conv4_norm.shape, label, conv4_norm,
self.n_classes)
return fconv_pooled, conv4_pooled, conv3_pooled
def forward(self, x):
conv1 = self.conv_block1(x)
conv2 = self.conv_block2(conv1)
conv3 = self.conv_block3(conv2)
conv4 = self.conv_block4(conv3)
conv5 = self.conv_block5(conv4)
fconv = self.fconv_block(conv5)
if self.use_norm:
fconv = l2_norm(fconv)
if self.use_scale:
fconv = self.scale * fconv
score = self.classifier(fconv)
if self.use_norm:
conv2 = l2_norm(conv2)
conv1 = l2_norm(conv1)
if self.use_scale:
conv2 = self.scale * conv2
conv1 = self.scale * conv1
if self.multires:
score_pool4 = self.score_pool4(conv2)
score_pool3 = self.score_pool3(conv1)
score = F.upsample(score, score_pool4.size()[2:])
score += score_pool4
score = F.upsample(score, score_pool3.size()[2:])
score += score_pool3
out = F.upsample(score, x.size()[2:])
return out
def forward(self, x):
# if not self.training:
# if self.multires:
# print('MultiRes Classifier with Alpha = 0.5')
# else:
# print('Vanilla Classifier with Alpha = 0.05')
conv1 = self.conv_block1(x)
conv2 = self.conv_block2(conv1)
conv3 = self.conv_block3(conv2)
conv4 = self.conv_block4(conv3)
conv5 = self.conv_block5(conv4)
fconv = self.fconv_block(conv5)
if self.use_norm:
fconv = l2_norm(fconv)
if self.use_scale:
fconv = self.scale * fconv
score = self.classifier(fconv)
if self.learned_billinear:
upscore2 = self.upscore2(score)
score_pool4c = self.score_pool4(conv4)[:, :, 5:5+upscore2.size()[2],
5:5+upscore2.size()[3]]
upscore_pool4 = self.upscore4(upscore2 + score_pool4c)
score_pool3c = self.score_pool3(conv3)[:, :, 9:9+upscore_pool4.size()[2],
9:9+upscore_pool4.size()[3]]
out = self.upscore8(score_pool3c + upscore_pool4)[:, :, 31:31+x.size()[2],
31:31+x.size()[3]]
return out.contiguous()
else:
if self.use_norm:
conv4 = l2_norm(conv4)
conv3 = l2_norm(conv3)
if self.use_scale:
conv4 = self.scale * conv4
conv3 = self.scale * conv3
if self.multires:
score_pool4 = self.score_pool4(conv4)
score_pool3 = self.score_pool3(conv3)
score = F.upsample(score, score_pool4.size()[2:])
score += score_pool4
score = F.upsample(score, score_pool3.size()[2:])
score += score_pool3
if self.offsetting:
pad = 100
if not self.training:
target_size = [s+pad*2 for s in x.size()[2:]]
out = F.upsample(score, target_size)
out = out[:, :, pad:-pad, pad:-pad]
else:
target_size = [s+2*pad for s in x.size()[2:]]
out = F.upsample(score, target_size)
else:
out = F.upsample(score, x.size()[2:])
return out