def forward(self, lrinput, refinput, showmode=True):
lr_feature = self.relu(self.lrfeature_ext1(lrinput))
lr_feature = self.relu(self.lrfeature_ext2(lr_feature))
lr_feature = self.feature_extraction_network(lr_feature)
ref_feature = self.downsample_module(refinput)
ref_feature = self.relu(ref_feature)
ref_feature = self.feature_extraction_network(ref_feature)
if showmode:
showpatch(lr_feature,
foldername="extracted_features_lr_image",
modelname="EDSR")
showpatch(ref_feature,
foldername="extracted_features_ref_image",
modelname="EDSR")
deform_feature = self.ssen_show(lr_batch=lr_feature,
init_hr_batch=ref_feature,
showmode=showmode,
modelname="EDST")
concated_feature = torch.cat((lr_feature, deform_feature), dim=1)
reconstruct_image = self.EDSR_show(x=concated_feature,
lr_feature=lr_feature)
return reconstruct_image
def forward(self, x, lr_feature):
x = self.mhead(x)
res = self.body(x)
res = torch.add(res, lr_feature)
showpatch(res,
foldername="features_after_reconstruction_blocks",
modelname="EDSR")
x = self.tail(res)
return x
def forward(self, lr_features, hr_features, modelname, showmode=False, num_block=None):
input_offset = torch.cat((lr_features, hr_features), dim=1)
estimated_offset = self.offset_estimator(input_offset)
if showmode:
showpatch(estimated_offset, foldername="DOE_output{}".format(num_block), modelname=modelname)
showpatch(input_offset, foldername="offsetinput{}".format(num_block), modelname=modelname)
estimated_offset = self.offset_conv(estimated_offset)
if showmode:
# showpatch(estimated_offset,foldername="extracted_offset_deformconv{}".format(num_block))
saveoffset(estimated_offset, modelname=modelname, foldername="resultoffset_deformconv{}".format(num_block),
istensor=True)
output = self.deformconv(x=hr_features, offset=estimated_offset)
return output
def forward(self, input_lr, ref_input, showmode=False):
ref_input = self.downsampling_network(ref_input)
input_lr = self.lrfeature_scaler1(input_lr)
input_lr = self.lrfeature_scaler2(input_lr)
lr_feature_out = self.feature_extractor(input_lr)
ref_feature_out = self.feature_extractor(ref_input)
SSEN_out = self.SSEN_Network(lr_batch=lr_feature_out,
init_hr_batch=ref_feature_out,
showmode=showmode)
residual_input = torch.cat((lr_feature_out, SSEN_out), dim=1)
residual_input_scaled = self.preprocessing_residual_block(
residual_input)
out = self.residual_blocks(residual_input_scaled)
out = torch.add(out, lr_feature_out)
if showmode:
showpatch(lr_feature_out, foldername="extracted_features_lr_image")
showpatch(ref_feature_out,
foldername="extracted_features_ref_image")
showpatch(out, foldername="features_after_reconstruction_blocks")
out = self.upscaling_4x(out)
out = self.outconv(out)
return out
def forward(self, lr_batch, init_hr_batch, modelname, showmode=False):
hr_out1 = self.deformblock1(lr_batch,
init_hr_batch,
showmode=showmode,
num_block=1,
modelname=modelname)
hr_out2 = self.deformblock2(lr_batch,
hr_out1,
showmode=showmode,
num_block=2,
modelname=modelname)
hr_out3 = self.deformblock3(lr_batch,
hr_out2,
showmode=showmode,
num_block=3,
modelname=modelname)
if showmode:
showpatch(hr_out1,
foldername="extracted_features_by_deformconv1",
modelname=modelname)
showpatch(hr_out2,
foldername="extracted_features_by_deformconv2",
modelname=modelname)
showpatch(hr_out3,
foldername="extracted_features_by_deformconv3",
modelname=modelname)
return hr_out3