def get_item(self, index, shift=None):
meta = {}
meta['do_not_collate'] = True
datas = self.data['datas'][index]
if shift is None:
path = self.data['image_paths'][index]
meta['time'] = shift = datas['time']
else:
n = datas['n']
shift = int(shift * (n - 1))
base = datas['base']
path = self.get_jpg_path(base, datas['id'], shift)
meta['time'] = shift
target = torch.IntTensor(self.num_classes).zero_()
for x in datas['labels']:
if x['start'] < shift / float(self.fps) < x['end']:
target[self.cls2int[x['class']]] = 1
meta['id'] = datas['id']
img = default_loader(path)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target, meta
def get_item(self, index, shift=None):
ims, tars, meta = [], [], {}
meta['do_not_collate'] = True
fps = 24
n = self.data['datas'][index]['n']
if shift is None:
shift = np.random.randint(n - self.train_gap - 2)
else:
shift = int(shift * (n - self.train_gap - 2))
resize = transforms.Resize(int(256. / 224 * self.input_size))
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
spacing = np.arange(shift, shift + self.train_gap)
for loc in spacing:
ii = int(np.floor(loc))
path = '{}{:06d}.jpg'.format(self.data['datas'][index]['base'],
ii + 1)
try:
# ============ Temp ===================
timer = Timer()
img = default_loader(path)
# ============ Temp ===================
load_img_cost = timer.thetime() - timer.end
timer.tic()
print(
'Load image from disk: {0:.3f} sec'.format(load_img_cost))
except Exception as e:
print('failed to load image {}'.format(path))
print(e)
raise
img = resize(img)
img = transforms.ToTensor()(img)
# ============ Temp ===================
# totensor_cost = timer.thetime() - timer.end
# timer.tic()
# print('From PIL to tensor: {0:.3f} sec'.format(totensor_cost))
#img = 2*img - 1
img = normalize(img)
ims.append(img)
target = torch.IntTensor(self.num_classes).zero_()
for x in self.data['datas'][index]['labels']:
if x['start'] < ii / float(fps) < x['end']:
target[self.cls2int(x['class'])] = 1
tars.append(target)
meta['id'] = self.data['datas'][index]['id']
meta['time'] = shift
img = torch.stack(ims).permute(0, 2, 3, 1).numpy()
target = torch.stack(tars)
if self.transform is not None:
img = self.transform(img)
# ============ Temp ===================
# transform_cost = timer.thetime() - timer.end
# timer.tic()
# print('Image transform per mini-batch: {0:.3f} sec'.format(transform_cost))
if self.target_transform is not None:
target = self.target_transform(target)
# batch will be b x n x h x w x c
# target will be b x n x nc
return img, target, meta
def __getitem__(self, index):
ims = []
personims = []
tars = []
meta = {}
fps = 24
n = self.data['datas'][index]['n']
if hasattr(self.data['datas'][index], 'shift'):
print('using shift')
shift = self.data['datas'][index]['shift']
else:
shift = np.random.randint(n-self.train_gap-2)
resize = transforms.Resize((int(256./224*self.input_size), int(256./224*self.input_size)))
spacing = np.arange(shift, shift+self.train_gap)
if len(self.data['datas'][index]['persontimes']) > 0:
closest = np.argmin([abs(x-(shift+self.train_gap/2)) for x in self.data['datas'][index]['persontimes']])
closest = self.data['datas'][index]['persontimes'][closest]
with open(self.data['datas'][index]['personbase']+'{:06d}.txt'.format(closest), 'r') as f:
box = [float(x) for x in f.readline().strip().split(' ')]
else:
print('no box, skipping cropping: {}'.format(self.data['datas'][index]['personbase']))
box = []
for loc in spacing:
ii = int(np.floor(loc))
path = '{}{:06d}.jpg'.format(self.data['datas'][index]['base'], ii+1)
try:
img = default_loader(path)
except Exception as e:
print('failed to load image {}'.format(path))
print(e)
raise
if len(box) > 0:
try:
personimg = _cropimg(img, *box[1:]).copy()
personimg = resize(personimg)
personimg = transforms.ToTensor()(personimg)
personimg = 2*personimg - 1
personims.append(personimg)
except Exception, e:
print(e)
box = []
img = resize(img)
img = transforms.ToTensor()(img)
img = 2*img - 1
ims.append(img)
if not len(box) > 0:
personims.append(img)
target = torch.IntTensor(self.num_classes).zero_()
for x in self.data['datas'][index]['labels']:
if x['start'] < ii/float(fps) < x['end']:
target[self.cls2int(x['class'])] = 1
tars.append(target)
def _get_one_image(self, index, shift):
fps = 24
path = '{}{:06d}.jpg'.format(self.data['datas'][index]['base'], shift+1)
target = torch.IntTensor(self.num_classes).zero_()
for x in self.data['datas'][index]['labels']:
if x['start'] < shift/float(fps) < x['end']:
target[self.cls2int(x['class'])] = 1
meta = {}
meta['id'] = self.data['datas'][index]['id']
meta['time'] = shift
img = default_loader(path)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target, meta
def get_item(self, index, shift=None):
meta = {}
if shift is None:
path = self.data['image_paths'][index]
shift = meta['time'] = self.data['times'][index]
else:
n = self.data['ns'][index]
shift = meta['time'] = int(shift * (n - 1))
base = self.data['image_paths'][index][:-10]
path = '{}{:06d}.jpg'.format(base, shift + 1)
target = torch.IntTensor(self.num_classes).zero_()
for x in self.data['labels'][index]:
if x['start'] < shift / float(self.fps) < x['end']:
target[self.cls2int(x['class'])] = 1
meta['id'] = self.data['ids'][index]
img = default_loader(path)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target, meta
def get_item(self, index, shift=None):
meta = self.data['meta'][index]
n = meta['n']
n_ego = meta['n_ego']
if shift is None:
shift = meta['thirdtime']
else:
shift = int(shift * (n - 1))
imbase, egobase = self.data['image_paths'][index]
egoii = to_ego_time(shift, n, n_ego)
negii = get_neg_time(egoii, n_ego, self.deltaneg)
if negii is None:
raise Exception('deltaneg too big for video')
meta['firsttime_pos'] = egoii
meta['thirdtime'] = shift
meta['firsttime_neg'] = negii
egopath_pos = '{}{:06d}.jpg'.format(egobase, egoii + 1)
impath = '{}{:06d}.jpg'.format(imbase, shift + 1)
egopath_neg = '{}{:06d}.jpg'.format(egobase, negii + 1)
impaths = (egopath_pos, impath, egopath_neg)
target = self.data['targets'][index]
classtarget = torch.IntTensor(self.num_classes).zero_()
for x in self.data['labels'][index]:
if x['start'] < shift / float(self.fps) < x['end']:
classtarget[self.cls2int(x['class'])] = 1
meta['class_target'] = classtarget
meta['id'] = self.data['ids'][index]
ims = [default_loader(im) for im in impaths]
if self.transform is not None:
ims = [self.transform(im) for im in ims]
if self.target_transform is not None:
target = self.target_transform(target)
#if random() > 0.5: # TODO DEBUG
# ims[2], ims[0] = ims[0], ims[2]
# target = -1
return ims, target, meta
def _process_stack(self, base, shift, data):
ims, tars, meta = [], [], {}
meta['do_not_collate'] = True
if self.split == 'train' and np.random.random() > 0.5:
resize = transforms.Resize(int(320. / 224 * self.input_size))
else:
resize = transforms.Resize(int(256. / 224 * self.input_size))
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
spacing = np.arange(shift, shift + self.train_gap)
for loc in spacing:
ii = int(np.floor(loc))
path = '{}{:06d}.jpg'.format(base, ii + 1)
try:
img = default_loader(path)
except Exception as e:
print('failed to load image {}'.format(path))
print(e)
raise
img = resize(img)
img = transforms.ToTensor()(img)
img = normalize(img)
ims.append(img)
target = torch.IntTensor(self.num_classes).zero_()
for x in data['labels']:
if x['start'] < ii / float(self.fps) < x['end']:
target[self.cls2int(x['class'])] = 1
tars.append(target)
meta['id'] = data['id']
meta['time'] = shift
img = torch.stack(ims).permute(0, 2, 3, 1).numpy() # n, h, w, c
target = torch.stack(tars)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target, meta
def read_one_iamge(path):
# path = '/home/nfs/x.chang/Datasets/Kinetics400/train_frms/lunge/Ku56XQb3N40_000004_000014/00000000.jpg'
return default_loader(path)
""" Video loader for the Charades dataset """
from datasets import utils
from misc_utils.utils import Timer
# path = '/home/SERILOCAL/xiatian.zhu/Data/test_video/ZZXQF-000002.jpg'
path = '/home/nfs/x.chang/Datasets/Charades/Charades/Charades_v1_rgb/ZZXQF/ZZXQF-000002.jpg'
for i in range(10):
try:
# ============ Temp ===================
timer = Timer()
img = utils.default_loader(path)
# ============ Temp ===================
load_img_cost = timer.thetime() - timer.end
timer.tic()
print('Load image from disk: {0:.3f} sec'.format(load_img_cost))
except Exception as e:
print('failed to load image {}'.format(path))
print(e)
raise
