def power_loss(self, wav_dict):
feat_fn = self.stft_feat_fn
pred_wav = wav_dict['x']
orig_wav = wav_dict['wav']
pred_len = pred_wav.get_shape().as_list()[1]
orig_len = orig_wav.get_shape().as_list()[1]
# crop longer wave
if pred_len > orig_len:
pred_wav = self._trim(pred_wav, pred_len - orig_len)
elif pred_len < orig_len:
orig_wav = self._trim(orig_wav, orig_len - pred_len)
orig_stft = mel_extractor._tf_stft(orig_wav)
pred_stft = mel_extractor._tf_stft(pred_wav)
orig_feat = PWNHelper.norm_or_not_fn(self, feat_fn(orig_stft))
pred_feat = PWNHelper.norm_or_not_fn(self, feat_fn(pred_stft))
diff = PWNHelper.diff_fn(orig_feat, pred_feat)
avg_loss = PWNHelper.avg_loss_fn(diff)
return {'power_loss': avg_loss}
def spec_feat_mean_std(train_path, feat_fn=lambda x: tf.pow(tf.abs(x), 2.0)):
local_graph = tf.Graph()
with local_graph.as_default():
input_vals = get_init_batch(train_path,
batch_size=4096,
seq_len=7680,
first_n=10000)['wav']
ph = tf.placeholder(dtype=np.float32, shape=[4096, 7680])
feat = feat_fn(mel_extractor._tf_stft(ph))
tf.logging.info('Calculating mean and std for stft feat.')
config = tf.ConfigProto(device_count={'GPU': 0})
sess = tf.Session(config=config, graph=local_graph)
feat_val = sess.run(feat, feed_dict={ph: input_vals})
mean_val = np.mean(feat_val, axis=(0, 1))
std_val = np.std(feat_val, axis=(0, 1))
tf.logging.info('Done calculating mean and std for stft feat.')
return mean_val, std_val