def predict(self, prefix, target_length):
"""
Predict a sequence using the given prefix.
"""
assert target_length > 0
with torch.no_grad():
prefix = prefix.reshape(prefix.shape[1], -1, 4)
prefix = qeuler_np(prefix, 'zyx')
prefix = qfix(euler_to_quaternion(prefix, 'zyx'))
inputs = torch.from_numpy(
prefix.reshape(1, prefix.shape[0], -1).astype('float32'))
if self.use_cuda:
inputs = inputs.cuda()
predicted, hidden = self.model(inputs)
frames = [predicted]
for i in range(1, target_length):
predicted, hidden = self.model(predicted, hidden)
frames.append(predicted)
result = torch.cat(frames, dim=1)
return result.view(result.shape[0], result.shape[1], -1,
4).cpu().numpy()
def predict(self, prefix, target_length):
"""
Predict a sequence using the given prefix.
"""
assert target_length > 0
with torch.no_grad():
prefix = prefix.reshape(prefix.shape[1], -1, 4)
prefix = qeuler_np(prefix, 'zyx')
prefix = qfix(euler_to_quaternion(prefix, 'zyx'))
inputs = torch.from_numpy(
prefix.reshape(1, prefix.shape[0], -1).astype('float32'))
if self.use_cuda:
inputs = inputs.cuda()
#print("HELLOLLLEOOELLLEOO", inputs.shape, target_length)
predicted, terms = self.model(
inputs, target_length,
0) #teacher forcing ratio = 0 for testing
frames = [predicted]
#result = torch.cat(frames, dim=1)
result = torch.cat(predicted, dim=1)
return result.view(result.shape[0], result.shape[1], -1,
4).cpu().numpy()
def evaluate(model, test_data):
errors = []
for d in test_data:
source = np.concatenate((d[0], d[1][:1]), axis=0).reshape(-1, 32 * 4)
target = d[2].reshape(-1, 32 * 4)
if model is None:
target_predicted = np.tile(source[-1],
target.shape[0]).reshape(-1, 32 * 4)
else:
target_predicted = model.predict(
np.expand_dims(source, 0),
target_length=np.max(frame_targets) + 1).reshape(-1, 32 * 4)
target = qeuler_np(target[:target_predicted.shape[0]].reshape(-1, 4),
'zyx').reshape(-1, 96)
target_predicted = qeuler_np(target_predicted.reshape(-1, 4),
'zyx').reshape(-1, 96)
e = np.sqrt(
np.sum((target_predicted[:, 3:] - target[:, 3:])**2, axis=1))
errors.append(e)
errors = np.mean(np.array(errors), axis=0)
return errors
def compute_euler_angles(self, order):
for subject in self._data.values():
for action in subject.values():
action['rotations_euler'] = qeuler_np(action['rotations'],
order,
use_gpu=self._use_gpu)