modelFile_handle = open(modelFile, 'w')
with torch.no_grad():
# classifier = FontClasifier_1_4(nClasses, classifyFonts)
# if opt.cuda:
# classifier.cuda()
# classifier.load_state_dict(torch.load(fontClassifierModelPath))
# classifier.eval()
# For the feature matching loss
classifier = FontClasifier_1_4(nClasses, classifyFonts)
classifier.load_state_dict(torch.load(fontClassifierModelPath))
feature_extractor = KfirFeatureExtractor(classifier, opt)
print(feature_extractor)
cont_enc = _EncoderNoa(opt.imageSize, opt.contentEnc_nz)
stl_enc = _EncoderNoa(opt.imageSize, opt.styleEnc_nz)
if useConcat:
dec = _DecoderNoa(opt.imageSize, opt.contentEnc_nz + opt.styleEnc_nz)
else:
dec = _DecoderNoa(opt.imageSize, opt.contentEnc_nz)
writeModel(modelFile_handle, feature_extractor, "feature_extractor")
writeModel(modelFile_handle, classifier, "classifier")
writeModel(modelFile_handle, cont_enc, "cont_enc")
writeModel(modelFile_handle, stl_enc, "stl_enc")
writeModel(modelFile_handle, dec, "dec")
modelFile_handle.close()
#decr = _DecoderNoa(opt.imageSize)
mse_criterion = nn.MSELoss()
modelFile_handle = open(modelFile, 'w')
with torch.no_grad():
# classifier = FontClasifier_1_4(nClasses, classifyFonts)
# if opt.cuda:
# classifier.cuda()
# classifier.load_state_dict(torch.load(fontClassifierModelPath))
# classifier.eval()
# For the feature matching loss
classifier = FontClasifier_1_4(nClasses, classifyFonts)
classifier.load_state_dict(torch.load(fontClassifierModelPath))
feature_extractor = KfirFeatureExtractor(classifier, opt)
print(feature_extractor)
cont_enc = _EncoderNoa(opt.imageSize)
stl_enc = _EncoderNoa(opt.imageSize)
dec = _DecoderNoa(opt.imageSize, useConcat)
writeModel(modelFile_handle, feature_extractor, "feature_extractor")
writeModel(modelFile_handle, classifier, "classifier")
writeModel(modelFile_handle, cont_enc, "cont_enc")
writeModel(modelFile_handle, stl_enc, "stl_enc")
writeModel(modelFile_handle, dec, "dec")
modelFile_handle.close()
#decr = _DecoderNoa(opt.imageSize)
mse_criterion = nn.MSELoss()
L1criterion = nn.L1Loss()
class_criterion = nn.CrossEntropyLoss()
#criterion = nn.BCELoss()
def runMethodTest(destPath, opt, contEncPath, styleEncPath, decPath,
useConcat):
if not os.path.exists(destPath):
os.makedirs(destPath)
data_loader = CreateDataLoader(opt)
dataloader = data_loader.load_data()
dataset_size = len(data_loader)
print('#testing images = %d' % dataset_size)
cont_enc = _EncoderNoa(opt.imageSize)
stl_enc = _EncoderNoa(opt.imageSize)
dec = _DecoderNoa(opt.imageSize, useConcat)
cont_enc.load_state_dict(torch.load(contEncPath))
stl_enc.load_state_dict(torch.load(styleEncPath))
dec.load_state_dict(torch.load(decPath))
if opt.cuda:
# feature_extractor.cuda()
cont_enc.cuda()
stl_enc.cuda()
dec.cuda()
torch.manual_seed(0)
for i, data in enumerate(dataloader, 0):
img1 = data['A']
img2 = data['B']
img12 = data['A2']
img21 = data['B2']
if opt.cuda:
img1 = img1.cuda()
img2 = img2.cuda()
img12 = img12.cuda()
img21 = img21.cuda()
stl1 = stl_enc(img1)
stl2 = stl_enc(img2)
cont1 = cont_enc(img1)
cont2 = cont_enc(img2)
# stl12 = stl_enc(img12)
# stl21 = stl_enc(img21)
# cont12 = cont_enc(img12)
# cont21 = cont_enc(img21)
if (useConcat):
stl1cont2 = torch.cat((stl1, cont2), 1)
stl2cont1 = torch.cat((stl2, cont1), 1)
# stl1cont1 = torch.cat((stl1, cont1), 1)
# stl2cont2 = torch.cat((stl2, cont2), 1)
else:
stl1cont2 = stl1 + cont2
stl2cont1 = stl2 + cont1
# stl1cont1 = stl1 + cont1
# stl2cont2 = stl2 + cont2
dec12 = dec(stl1cont2)
dec21 = dec(stl2cont1)
# dec11 = dec(stl1cont1)
# dec22 = dec(stl2cont2)
if i % 10 == 0:
im1 = util.tensor2im(img1[0])
im2 = util.tensor2im(img2[0])
oim12 = util.tensor2im(img12[0])
oim21 = util.tensor2im(img21[0])
im12 = util.tensor2im(dec12[0])
im21 = util.tensor2im(dec21[0])
imageio.imwrite(
os.path.join(destPath, '%d_style_1_cont_2.png' % (i)), im12)
imageio.imwrite(
os.path.join(destPath, '%d_style_2_cont_1.png' % (i)), im21)
imageio.imwrite(
os.path.join(destPath, '%d_style_1_cont_1_orig.png' % (i)),
im1)
imageio.imwrite(
os.path.join(destPath, '%d_style_2_cont_2_orig.png' % (i)),
im2)
imageio.imwrite(
os.path.join(destPath, '%d_style_1_cont_2_orig.png' % (i)),
oim12)
imageio.imwrite(
os.path.join(destPath, '%d_style_2_cont_1_orig.png' % (i)),
oim21)