weight_decay=weight_decay)
scheduler = LambdaLR(optimizer, lr_lambda=[lr_decay])
# Creation of the typegroup classifier
tgc = TypegroupsClassifier({'handwritten': 0, 'printed': 1}, net)
if os.path.exists(
os.path.join('ocrd_typegroups_classifier', 'models',
'classifier.tgc')):
tgc = TypegroupsClassifier.load(
os.path.join('ocrd_typegroups_classifier', 'models', 'classifier.tgc'))
else:
print('Could not load a model to evaluate')
quit(1)
validation = ImageFolder('lines/validation', transform=None)
validation.target_transform = tgc.classMap.get_target_transform(
validation.class_to_idx)
good = 0
bad = 0
with torch.no_grad():
tgc.network.eval()
for idx in tqdm(range(validation.__len__()), desc='Evaluation'):
sample, target = validation.__getitem__(idx)
path, _ = validation.samples[idx]
if target == -1:
continue
result = tgc.classify(sample, 224, 64, True)
highscore = max(result)
label = tgc.classMap.cl2id[result[highscore]]
if target == label:
good += 1
else:
trans = transforms.Compose([
transforms.RandomAffine(4, shear=10),
transforms.RandomAffine(4, shear=10),
transforms.RandomAffine(4, shear=10),
transforms.RandomCrop(224),
transforms.ColorJitter(brightness=0.5,
contrast=0.5,
saturation=0.1,
hue=0.05),
QLoss(min_q=5, max_q=80),
transforms.ToTensor()
])
#training = ImageFolder('/cluster/seuret/patches/all', transform=trans)
# TODO : replace by correct path
training = ImageFolder('../extracted/samples', transform=trans)
training.target_transform = tgc.classMap.get_target_transform(
training.class_to_idx)
data_loader = torch.utils.data.DataLoader(training,
batch_size=64,
shuffle=True,
num_workers=4)
# Iterating over the data
print('Starting the training - grab a coffee and a good book!')
for epoch in range(200):
# Modify learning rate
scheduler.step()
# Iterate over the data
lossSum = 0
good = 0
"folder",
help="path to the folder containing the datasets (subfolders with images)",
type=str)
args = parser.parse_args()
sampling = transforms.Compose(
[transforms.RandomCrop(224),
transforms.ToTensor()])
inv = transforms.ToPILImage()
test = ImageFolder(args.folder, transform=sampling)
dev = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
tgc = TypegroupsClassifier.load(args.classifier)
test.target_transform = tgc.classMap.get_target_transform(test.class_to_idx)
nb_classes = 1 + max(tgc.classMap.cl2id.values())
nb_outputs = tgc.network.fc.out_features
print(nb_classes, 'classes to consider')
print(nb_outputs, 'outputs to process')
sm = torch.nn.Softmax()
nb_good = 0
nb_bad = 0
tgc.network.eval()
imgn = 0
with torch.no_grad():
f = open("foo.html", "w")
f.write('<html><head></head><body><table>')