if __name__ == '__main__':
with open(pred_path,'r') as predf, open(target_path, 'r') as tarf:
print(pred_path)
print(target_path)
pred_data = json.load(predf)
tar_data = json.load(tarf)
if show_ground_truth:
draw(frames=np.array(pred_data['skeletons']),video_path=pred_video, tempo_path=tempo_path, target_frames=np.array(tar_data['skeletons']))
else:
draw(frames=np.array(pred_data['skeletons']),video_path=pred_video, tempo_path=tempo_path)
if to_video:
start, end = load_start_end_frame_num(config_path)
gr_duration,_ = load_skeleton(target_path)
pred_duration,_ = load_skeleton(pred_path)
audio = AudioFileClip(audio_path)
start = start + gr_duration -pred_duration
end = start + pred_duration
sub = audio.subclip(start/fps, end/fps)
print('Analyzed the audio, found a period of %.02f seconds' % sub.duration)
video = VideoFileClip(pred_video, audio=False)
video = video.set_audio(sub)
video.write_videofile(pred_video2)
pass
# with open(input_path2,'r') as fin:
# data2 = json.load(fin)
#
# draw(np.array(data2['skeletons']), center=np.array(data2['center']), avi_path=file_path2, tempo_path=tempo_path)
def test(test_sample_dir,
acoustic_features_scaler,
motion_features_scaler,
tempo_scaler=None):
print("\n........testing........\n")
test_acoustic_features, temporal_indexes, test_motion_features = load_features_from_dir(
test_sample_dir)
if tempo_scaler is not None:
temporal_indexes = tempo_scaler.transform(temporal_indexes)
test_acoustic_features = acoustic_features_scaler.transform(
test_acoustic_features)
test_motion_features = motion_features_scaler.transform(
test_motion_features)
if with_tempo_features:
test_acoustic_features = np.hstack(
(test_acoustic_features, temporal_indexes))
# num_test_seq = int(len(test_acoustic_features) / seq_len)
# test_acoustic_features = test_acoustic_features[:num_test_seq * seq_len, :].reshape(num_test_seq, seq_len, -1)
# test_motion_features = test_motion_features[:num_test_seq * seq_len, :].reshape(num_test_seq, seq_len, -1)
print("shape:{0}".format(test_acoustic_features.shape))
# model_path = os.path.join(ck_dir, model_name + "_epoch_%d.pth"%last_epoch)
model_path = os.path.join(ck_dir, model_name + "_latest.pth")
print("model load from %s" % model_path)
checkpoint = torch.load(model_path)
model = create_model(model_name)
model.load_state_dict(checkpoint['net'])
model.eval()
# test_dataset = TensorDataset(torch.from_numpy(test_acoustic_features[np.newaxis,:]), torch.from_numpy(test_motion_features[np.newaxis,:]))
test_dataset = TensorDataset(
torch.from_numpy(test_acoustic_features[np.newaxis, ]),
torch.from_numpy(test_motion_features[np.newaxis, ]))
test_dataloader = DataLoader(dataset=test_dataset,
batch_size=test_batch_size,
shuffle=False,
drop_last=False,
num_workers=num_workers)
criterion = nn.MSELoss()
predict_motion_features = np.empty([0, motion_size])
with torch.no_grad():
for i, (batch_x, batch_y) in enumerate(test_dataloader):
batch_x, batch_y = Variable(batch_x), Variable(batch_y)
batch_x, batch_y = batch_x.to(device), batch_y.to(device)
output = model(batch_x)
if model_name == 'LSTM-AE'or model_name=='LSTM-VAE'\
:
loss = criterion(output[1], batch_y)
output = np.reshape(output[1].detach().cpu().numpy(),
newshape=[-1, motion_size])
batch_y = np.reshape(batch_y.detach().cpu().numpy(),
newshape=[-1, motion_size])
else:
# loss = criterion(output, batch_y)
output = np.reshape(output.detach().cpu().numpy(),
newshape=[-1, motion_size])
batch_y = np.reshape(batch_y.detach().cpu().numpy(),
newshape=[-1, motion_size])
# output = output.detach().cpu().numpy()
# loss_data = loss.detach().cpu().numpy()
predict_motion_features = np.append(predict_motion_features,
output,
axis=0)
predict_real_motion_features = motion_features_scaler.inverse_transform(
predict_motion_features)
predict_real_motion_features = np.reshape(
predict_real_motion_features, newshape=[-1, motion_size // 3, 3])
center = load_skeleton(
os.path.join(test_sample_dir,
'skeletons.json'))[1][:len(predict_real_motion_features)]
if with_centering:
for i in range(len(predict_real_motion_features)):
for j in range(len(predict_real_motion_features[i])):
predict_real_motion_features[i][j] += center[i]
data = dict()
output_json_fn = os.path.join(test_sample_dir, output_json)
data['length'] = len(predict_real_motion_features)
data['center'] = center
data['skeletons'] = predict_real_motion_features.tolist()
with open(output_json_fn, 'w') as f:
json.dump(data, f)
print("saved as %s" % output_json_fn)
pass