def normalize(self, statistics=None):
# normalize the input-output pairs with optional given statistics
if statistics is None:
mean_in, std_in = nop.get_statistics_1d(self.input)
mean_out, std_out = nop.get_statistics_1d(self.output)
self.statistics = {
'mean_in': mean_in,
'mean_out': mean_out,
'std_in': std_in,
'std_out': std_out
}
else:
mean_in, std_in = statistics['mean_in'], statistics['std_in']
mean_out, std_out = statistics['mean_out'], statistics['std_out']
self.statistics = statistics
self.input = nop.normalize_1d(self.input, mean_in, std_in)
self.output = nop.normalize_1d(self.output, mean_out, std_out)
return
def lift_2d_to_3d(records, model, stats, template, cuda=True):
"""
Foward-pass of the lifter model.
"""
for path in records.keys():
data = np.concatenate(records[path]['kpts_2d_pred'], axis=0)
data = nop.normalize_1d(data, stats['mean_in'], stats['std_in'])
data = data.astype(np.float32)
data = torch.from_numpy(data)
if cuda:
data = data.cuda()
prediction = model(data)
prediction = nop.unnormalize_1d(prediction.data.cpu().numpy(),
stats['mean_out'], stats['std_out'])
records[path]['kpts_3d_pred'] = prediction.reshape(
len(prediction), -1, 3)
return records