CODEWITHSUNDEEP
pythontensorflow
Aya
Aya

Reputation: 120

tensor flow and short time Fourier transform

I am using tensorflow version 2.0

now I have a wav file which I load here

audio_file = tf.io.read_file(wav_path)
wav_array, sample_rate = tf.audio.decode_wav(
audio_file,
desired_channels=1)
wav_array = np.array(wav_array)
wav_array = wav_array.reshape(wav_array.shape[0])

then I get the short time Fourier transf using

spec = tf.signal.stft(wav_array, window_fn=tf.signal.hann_window, 
frame_length=WIN_LEN, frame_step=HOP_LEN,fft_length=FFT_SIZE)

when I print the dims it is

Spectrogram shape: (748, 257)

Then I try to get the inverse using

inv = tf.signal.inverse_stft(spec,frame_length=WIN_LEN, 
frame_step=HOP_LEN,fft_length=FFT_SIZE, 
window_fn=tf.signal.hann_window)

but i get this error

AttributeError: 'int' object has no attribute 'value'

/usr/local/lib/python3.6/dist- 
packages/tensorflow/python/ops/signal/spectral_ops.py in 
inverse_stft(stfts, frame_length, frame_step, fft_length, window_fn, 
name)

if (frame_length_static is None or
real_frames.shape.ndims is None or
--> real_frames.shape[-1].value is None):
real_frames = real_frames[..., :frame_length]
real_frames_rank = array_ops.rank(real_frames)

Upvotes: 2

Views: 943

Answers (1)

Sharky
Sharky

Reputation: 4543

Diving deeper into the problem, I found out a workaround. TensorFlow requires fft_length parameter to be integer for both inverse_stft and stft functions. If it's no provided, it will use compute it as described in docs.

For some reason, file spectral_ops.py contains code, that first defines 

real_frames = fft_ops.irfft(stfts, [fft_length])  #line 235

and then tries to access .value of an integer multiple times

real_frames.shape[-1].value #lines 244, 255, 257

which throws

AttributeError: 'int' object has no attribute 'value'

to no surprise.

Removing call to .value makes tf.signal.inverse_stft produce expected result. It's arguably seems like a github issue.

Upvotes: 1

Related Questions