def __getitem__(self, index):
## Label
filen = 'synset_words.txt'
indicestosynsets, synsetstoindices, synsetstoclassdescriptions = gc.parsesynsetwords(
filen)
label, firstname = gc.parseclasslabel('val/' + self.labellist[index],
synsetstoindices)
## Image
img_as_np = io.imread('imagespart/' + self.imglist[index])
img_as_img = Image.fromarray(img_as_np)
# img_as_img = img_as_img.convert('L')
## Resize smaller size to 224
a, b = img_as_img.size[0], img_as_img.size[1]
if a < 224 or b < 224:
return None
if a < b:
percent = 224 / float(a)
hsize = int((float(b) * float(percent)))
img_as_img = img_as_img.resize((224, hsize), Image.BILINEAR)
else:
percent = 224 / float(b)
hsize = int((float(a) * float(percent)))
img_as_img = img_as_img.resize((hsize, 224), Image.BILINEAR)
## Read to numpy array
img_as_np = np.array(img_as_img)
if len(img_as_np.shape) != 3:
return None
## Normalize [0, 1]
img_as_np = (img_as_np - np.min(img_as_np)) / (np.max(img_as_np) -
np.min(img_as_np))
# print(img_as_np[0:])
## swap axes
img_as_np = np.swapaxes(img_as_np, 0, 2)
img_as_np = np.swapaxes(img_as_np, 1, 2)
# print(img_as_np)
## input normalization
img_as_np[0:] = (img_as_np[0:] - 0.485) / 0.229
img_as_np[1:] = (img_as_np[1:] - 0.456) / 0.224
img_as_np[2:] = (img_as_np[2:] - 0.406) / 0.225
## center crop
a = img_as_np.shape[1]
b = img_as_np.shape[2]
if a < b:
x = int((b - 224) / 2)
img_as_np = img_as_np[:, :, x:(x + 224)]
# print(img_as_np.shape)
else:
x = int((a - 224) / 2)
img_as_np = img_as_np[:, x:(x + 224), :]
img_as_tensor = torch.tensor(img_as_np)
# print(img_as_np.shape)
if self.transforms is not None:
img_as_tensor = self.transforms(img_as_img)
return (img_as_tensor.type('torch.FloatTensor'), label)
def __init__(self, root_dir, file_prefix='ILSVRC2012_val_',
img_ext='.JPEG', val_ext='.xml', synset='synset_words.txt',
five_crop=False, data_limit=0, selector=None):
'''
NOTE: set up your root_dir directory to consist of
2 directories:
- imagespart: where the images are
- val: where the xml's are (for class values etc)
'''
self.meta = {'root_dir':root_dir,
'file_prefix':file_prefix,
'synset':synset,
'img_ext':img_ext, 'val_ext':val_ext,
'five_crop': five_crop}
# assertion for the mentioned assumption
assert os.path.exists(os.path.join(root_dir, 'imagespart'))
assert os.path.exists(os.path.join(root_dir, 'val'))
assert os.path.exists(os.path.join(root_dir, synset))
# metadata
self.classes = ginc.get_classes()
_, s2i, s2d = ginc.parsesynsetwords(os.path.join(root_dir, synset))
self.dataset = [filename[len(file_prefix):-len(img_ext)]
for filename in os.listdir(os.path.join(root_dir, 'imagespart'))]
if data_limit > 0:
self.dataset = self.dataset[:data_limit]
if selector is not None:
self.dataset = [d for d, s in zip(self.dataset, selector) if s]
self._rev_dataset = s2i
self.data_description = s2d
def __getitem__(self, index):
## Label
filen = 'synset_words.txt'
indicestosynsets, synsetstoindices, synsetstoclassdescriptions = gc.parsesynsetwords(
filen)
label, firstname = gc.parseclasslabel('val/' + self.labellist[index],
synsetstoindices)
## Image
img_as_np = io.imread('imagespart/' + self.imglist[index])
img_as_img = Image.fromarray(img_as_np)
# img_as_img = img_as_img.convert('L')
if self.transforms is not None:
img_as_tensor = self.transforms(img_as_img)
return (img_as_tensor, label)
def __init__(self,
root_dir,
xmllabeldir,
synsetfile,
maxnum,
transform=None):
"""
Args:
root_dir (string): Directory with all the images.
transform (callable, optional): Optional transform to be applied
on a sample.
"""
self.root_dir = root_dir
self.xmllabeldir = xmllabeldir
self.transform = transform
self.imgfilenames = []
self.labels = []
self.ending = ".JPEG"
self.clsdict = get_classes()
indicestosynsets, self.synsetstoindices, synsetstoclassdescr = parsesynsetwords(
synsetfile)
for root, dirs, files in os.walk(self.root_dir):
for ct, name in enumerate(files):
nm = os.path.join(root, name)
#print(nm)
if (maxnum > 0) and ct >= (maxnum):
break
self.imgfilenames.append(nm)
label, firstname = parseclasslabel(self.filenametoxml(nm),
self.synsetstoindices)
self.labels.append(label)
def get_label(path_words_txt, path_xml):
indicestosynsets, synsetstoindices, synsetstoclassdescr = parsesynsetwords(
path_words_txt)
idx, label_id = parseclasslabel(path_xml, synsetstoindices)
name = synsetstoclassdescr[indicestosynsets[idx]]
return dict(label_idx=idx, label_name=name, label_id=label_id)
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
inv_transform = transforms.Compose([
transforms.Normalize(mean=[0., 0., 0.],
std=[1 / 0.229, 1 / 0.224, 1 / 0.225]),
transforms.Normalize(mean=[-0.485, -0.456, -0.406], std=[1., 1., 1.]),
transforms.ToPILImage(),
])
resize_transform = transforms.Compose(
[transforms.Resize(224),
transforms.CenterCrop(224)])
i2s, s2i, s2d = ginc.parsesynsetwords('synset_words.txt')
def get_descript(class_no):
s = i2s[class_no]
desc = s2d[s]
return desc
def get_trained_resnet(use_gpu=True):
model_ft = models.resnet18(pretrained=True)
if use_gpu:
model_ft = model_ft.try_cuda()
return model_ft