def get_model(args):
""" Create base model, and wrap it with an optional wrapper, useful for extending models
"""
model = generic_load(args.arch, args.pretrained, args.pretrained_weights,
args)
print("arch is ", args.arch)
if args.replace_last_layer:
print('replacing last layer')
model = replace_last_layer(model, args)
for module in model.modules():
if args.dropout != 0 and isinstance(module, torch.nn.modules.Dropout):
print('setting Dropout p to {}'.format(args.dropout))
module.p = args.dropout
wrapper = case_getattr(import_module('models.wrappers.' + args.wrapper),
args.wrapper)
model = wrapper(model, args)
model = set_distributed_backend(model, args)
# define loss function
criterion = case_getattr(
import_module('models.criteria.' + args.criterion), args.criterion)
criterion = criterion(args).cuda()
return model, criterion
def run(cls, model, criterion, epoch, args):
model = ActorObserverClassifierWrapper(model, args)
model = set_distributed_backend(model, args)
criterion = DefaultCriterion(args)
task = cls(model, epoch, args)
loader, = get_dataset(args, splits=('val_video', ), dataset=args.actor_observer_classification_task_dataset)
return task.validate_video(loader, model, criterion, epoch, args)
def alignment(self, loader, model, epoch, args, task=best_one_sec_moment):
timer = Timer()
abssec = MedianMeter()
abssec0 = MedianMeter()
randsec = MedianMeter()
model = ActorObserverFC7Wrapper(model, args)
model = set_distributed_backend(model, args)
# switch to evaluate mode
model.eval()
def fc7_generator():
for i, (inputs, target, meta) in enumerate(loader):
if not args.cpu:
target = target.cuda(async=True)
first_fc7, third_fc7, w_x, w_y = model(*inputs)
timer.tic()
if i % args.print_freq == 0:
print('Alignment: [{0}/{1}]\t'
'Time {timer.val:.3f} ({timer.avg:.3f})'.format(
i, len(loader), timer=timer))
for vid, o1, o2 in zip(meta['id'], first_fc7, third_fc7):
yield vid, (o1.cpu().numpy(), o2.cpu().numpy())
for key, grp in groupby(fc7_generator(), key=lambda x: x[0]):
print('processing id: {}'.format(key))
_, mat = fc7list2mat(grp)
_, _, _, j, gt = task(mat, winsize=3)
_, _, _, j0, gt0 = task(mat, winsize=0)
_, _, _, jr, gtr = task(np.random.randn(*mat.shape), winsize=3)
abssec.update(abs(j - gt))
abssec0.update(abs(j0 - gt0))
randsec.update(abs(jr - gtr))
print(self.name)
print(' abs3: {abs3.val:.3f} ({abs3.avg:.3f}) [{abs3.med:.3f}]'
' abs0: {abs0.val:.3f} ({abs0.avg:.3f}) [{abs0.med:.3f}]'
'\n'
' absr: {absr.val:.3f} ({absr.avg:.3f}) [{absr.med:.3f}]'.
format(abs3=abssec, abs0=abssec0, absr=randsec))
scores = {
self.name + '_1sec': abssec0.med,
self.name + '_1sec_smooth': abssec.med,
self.name + '_1sec_random': randsec.med
}
return scores
def run(cls, model, criterion, epoch, args):
model = ActorObserverClassifierWrapper(model, args)
model = set_distributed_backend(model, args)
criterion = DefaultCriterion(args)
task = cls(model, epoch, args)
newargs = copy.deepcopy(args)
if ';' in args.train_file:
vars(newargs).update({
'train_file': args.train_file.split(';')[1],
'val_file': args.val_file.split(';')[1],
'data': args.data.split(';')[1]
})
if '3d' in args.arch:
loader, = get_dataset(newargs,
splits=('val_video', ),
dataset='charades_video')
else:
loader, = get_dataset(newargs,
splits=('val_video', ),
dataset='charades')
return task.validate_video(loader, model, criterion, epoch, args)