using seed audio

[materialvision]

Thanks for keeping the Samplernn alive!
Using generate.py with seed audio I get this error message:

Traceback (most recent call last):
File "generate.py", line 227, in
main()
File "generate.py", line 222, in main
generate(args.output_path, args.checkpoint_path, config, args.num_seqs, args
.dur,
File "generate.py", line 190, in generate
init_samples[:, :model.big_frame_size, :] = quantize(seed_audio, q_type, q_l
evels)
TypeError: 'tensorflow.python.framework.ops.EagerTensor' object does not support
item assignment

Any hint on what I can do to get the seed audio function working?

[materialvision]

Sorry I see now the similar question in this issue: #6

[relativeflux]

@materialvision Hi,yes - apologies, this is a known bug I was going to fix a while back but didn't get around to - will sort it ASAP.

Thanks!

[fdb]

Hi @relativeflux, would love to see if this can be fixed!

I'm currently generating without any seed which works; does the algorithm then use noise as its seed?

[ezrababski]

Echoing the above comments—getting the same error (but very much appreciating the code)!

[relativeflux]

Thanks. @fdb Actually it's using silence (128 at q_levels=256). @ezrababski I do have a fix for this now, I will push it out later today or tomorrow. Looking to improve that code, as it doesn't really do much to modify the generated samples... WaveNet uses a much larger seed size, so I'm looking in that direction.

[fdb]

This is the code I currently use as a workaround (basically converting to numpy and back):

    if seed is not None:
        seed_audio = load_seed_audio(seed, seed_offset, model.big_frame_size)
        np_quantized = quantize(seed_audio, q_type, q_levels).numpy()
        np_init_samples = init_samples.numpy()
        np_init_samples[:, :model.big_frame_size, :] = np_quantized
        init_samples = tf.convert_to_tensor(np_init_samples)

[fdb]

Hi @relativeflux , is the fix I posted any help in fixing this issue?

[relativeflux]

Hi @fdb, I've just pushed a fix, similar to your own. Apologies for the delay. Seeding works again now, at least it doesn't error - it still doesn't seem to actually make any real difference to the end result, however. I'm looking into this.

[fdb]

Thanks! Looks good!