def make_omega_pred(self, use_optcam=False):
return OmegasPred(
config=self.config,
smpl=self.smpl,
use_optcam=use_optcam,
vis_max_batch=self.batch_size,
is_training=False,
)
def build_test_model(self):
B, T = self.batch_size, self.sequence_length
I_t = tf.reshape(self.images_pl,
(B * T, self.img_size, self.img_size, 3))
img_feat, phi_var_scope = self.f_image_enc(
I_t,
is_training=False,
reuse=False,
)
img_feat_full = tf.reshape(img_feat, (B, T, -1))
omega_mean = tf.tile(self.theta_mean, (self.sequence_length, 1))
if self.pred_mode == 'pred':
movie_strips = self.f_temporal_enc(
is_training=False,
net=img_feat_full,
num_conv_layers=self.num_conv_layers,
)
elif self.pred_mode == 'hal':
movie_strips = self.f_hal(img_feat_full)
else:
raise Exception('Pred mode {} not recognized'.format(
self.pred_mode))
omegas_raw, deltas_pred = batch_pred_omega(
input_features=movie_strips,
batch_size=self.batch_size,
sequence_length=self.sequence_length,
num_output=self.num_output,
is_training=False,
omega_mean=omega_mean,
scope='single_view_ief',
predict_delta_keys=self.omegas_pred.keys(),
use_optcam=self.use_optcam,
use_delta_from_pred=True,
)
self.omegas_pred[0].append_batched(omegas_raw)
for k in deltas_pred.keys():
self.omegas_pred[k].append_batched(deltas_pred[k])
if self.use_optcam:
self.omegas_pred[k].set_cams(self.omegas_pred[0].get_cams())
OmegasPred.compute_all_smpl()
self.update_encoder_vars()