Using ddsp_prepare_tfrecord with different 'example_secs' values than default about ddsp HOT 9 CLOSED

You'll need to change set the example_secs appropriately in the TFRecordProvider for this to work.
For example, you can use a flag like --gin_param='TFRecordProvider.example_secs = 8' or add that setting to your gin config file.

from ddsp.

Thanks Adarob, do you mean the use of of a fixed and same length for train and test files is a must for this experiment?

from ddsp.

I believe it is possible to use a different length at inference time. @jesseengel to confirm. As for training, a fixed length is required by the data pipeline, but could probably be relaxed if necessary.

from ddsp.

That's exactly my problem, I use fixed length (example 4 secs) during training. No problem in re-synthesizing train chunks of 4 seconds. Then I have a bunch of test files(from the same instrument) at various lengths. I would like to feed them to the autoencoder forward process (without running the feedback loop for training). This way I would like to check the analysis-synthesis quality of the trained network comparing audio files by listening to them (original versus resynthesized). Thanks

from ddsp.

This is possible at the moment, although not the most elegant.

For a single case we do this in the timbre_transfer demo, by just reseting the values with gin.

# Ensure dimensions and sampling rates are equal
time_steps_train = gin.query_parameter('DefaultPreprocessor.time_steps')
n_samples_train = gin.query_parameter('Additive.n_samples')
hop_size = int(n_samples_train / time_steps_train)

time_steps = int(audio.shape[1] / hop_size)
n_samples = time_steps * hop_size

gin_params = [
    'Additive.n_samples = {}'.format(n_samples),
    'FilteredNoise.n_samples = {}'.format(n_samples),
    'DefaultPreprocessor.time_steps = {}'.format(time_steps),
]

with gin.unlock_config():
  gin.parse_config(gin_params)

However I would recommend creating a new dataset out of your test examples (cut into overlapping fixed lengths) and then reassembling them after the fact to listen to the whole sample together. This is due to a known issue with long samples (10+ seconds) where float32 errors accumulate in the sinusoidal oscillators.

from ddsp.

Thanks Jesse. I thought using a long enough fixed length for both train and test could be a simpler solution (rather than rejoining segments, one could crop a longer).

So I was trying to have everything at 8 seconds length. I observed there exists another parameter window_secs set default as 4 in prepare_tfrecord_lib.py (line 110) and it seems a user set example_secs value (other than 4) would not be reflected there. Could there be a missing link between window_secs and example_secs?

Has anyone trained the auto_encoder with segment lengths other than 4 seconds?

from ddsp.

With 2 seconds it seems to be working fine. My training and test data were already sliced to 2 seconds though.

from ddsp.

@barisbozkurt , the examples_secs in prepare_tfrecord_lib.py should be set by the flag, so as long as you make the dataset with 8 second long examples --example_secs=8 and then change your gin parameters / file too you should be good to go.

https://github.com/magenta/ddsp/blob/master/ddsp/training/data_preparation/prepare_tfrecord.py#L79

from ddsp.

Thanks for the comments Adarob, Jesse and Voodoohop. I have not completely solved my problem but I guess that's something to be done on my side

from ddsp.