Dataset creation problem about wavegan HOT 13 CLOSED

Change ext to 'wav' not 'wave'

I need more information if that doesn't fix it

from wavegan.

I have used "wav" in the command but still the problem persists
i have attaches the screenshots below

this is the command i used

Thanks in advance

from wavegan.

I believe I have fixed it with the most recent update. The issue occurred if the number of WAV files in the directory was smaller than the number of shards. Should be fixed now, please give it another try!

from wavegan.

Hi,
sorry to bother you again,
but that didn't fix my problem
I would attach a wav file from my data .
could you please check with them .
test.zip
Thanks in advance

from wavegan.

I was able to successfully make tfrecords from those two files using my script. Not sure what the issue is.

What version of tensorflow are you using? Could you print out npershard in the script to see if that's the issue? Also, can you print _slices.shape after line 114? This information might help me debug the issue.

I suspect it might be an issue with FFmpeg

from wavegan.

Hello. Just a note to @Varshithpolu, I had the same problem with "empty dataset" errors. I solved it by using a longer wav file (several minutes instead of 10 seconds pieces).

from wavegan.

I have the same issue, in which the *tfrecord is empty. I downloaded the code 2 days ago. After I tested line by line, I figured out the issue from FFmpeg '_UnavailableError: FFmpeg must be installed to run this op. FFmpeg can be found at http://www.ffmpeg.org. [[Node: DecodeAudioV2 = DecodeAudioV2[stream="", device="/job:localhost/replica:0/task:0/device:CPU:0"](ReadFile, DecodeAudioV2/file_format, DecodeAudioV2/samples_per_second, DecodeAudioV2/channel_count)]]'

In my Ubuntu system, the issue was gone after I installed ffmpeg. However, another issue shows up '_example = tf.train.Example(features=tf.train.Features(feature={
#'id': tf.train.Feature(bytes_list=tf.train.BytesList(value=audio_id)),
'label': tf.train.Feature(bytes_list=tf.train.BytesList(value=audio_label)),
'slice': tf.train.Feature(int64_list=tf.train.Int64List(value=[j])),
'samples': tf.train.Feature(float_list=tf.train.FloatList(value=slice))
}))' (the line I commented)
The error information is 'File "", line 1, in
tf.train.Feature(bytes_list=tf.train.BytesList(value=audio_id))

TypeError: 'S' has type str, but expected one of: bytes'

But anyway, the array is generated from ' tf.train.Feature(float_list=tf.train.FloatList(value=_slice))' and the tfrecord file is not empty now.

from wavegan.

@LiangqunLu I think that issue arises from using Python 3 which differentiates between str and bytes. You can always delete this line to avoid this issue. The labels are only added to the records for conditional GAN experiments

from wavegan.

I was also having issue with the tf_records creation,
I was using the following command for creation

python data/make_tfrecord.py
./new_data/sc09/train
./new_data/train
--name train
--ext wav
--fs 16000
--nshards 128
--slice_len 1.5 \

The files created are empty (0 bytes)

But when I changed slice_len to 1sec, I got all sorted out.
I also added --labels to pre append labels (train,test,valid) to the file names.
python data/make_tfrecord.py
./new_data/sc09/train
./new_data/train/
--name train --labels
--ext wav
--fs 16000
--nshards 128
--slice_len 1 \

[SOLVED]

from wavegan.

If anyone is having trouble with making datasets (empty TFrecord files), make sure FFmpeg is installed. It is required by the make_tfrecords.py file. On ubuntu 16: sudo apt install ffmpeg

from wavegan.

@Varshithpolu What method did you use to solve this problem? I also encountered this problem when I tried to build Dataset for SC09. Please help me！
I've already installed ffmpeg. But it still didn't work .
My tensorflow-gpu version is 1.4

this is the command i used：
python data/make_tfrecord.py ./sc09/train ./train --ext wav --fs 16000 --nshards 128 --slice_len 1.5

from wavegan.

You shouldn't have to build the dataset for SC09 from WAV files. You can get it prebuilt here:

https://drive.google.com/open?id=1qRdAWmjfWwfWIu-Qk7u9KQKGINC52ZwB

from wavegan.

@chrisdonahue Thank you very much for taking note of my questions！I have found a solution to this problem, hoping to help others.
1.Try not to use Windows system. Make sure your system is Ubantu 16.04.
2. Make sure you have installed ffmpeg;
3.【--slice_len 1】The meaning of this command is the length of time per slice. So make sure that this length is less than the audio length of your dataset. eg：SC09 dateset length is about 1.2s，so you should set 【--slice_len】 to a value shorter than 1.2 seconds. If you set 【--slice_len】 to 1.5, you will get some empty files.
4. Check your Python version. If your Python version is Python 3，you may encounter some errors when you adjust the length of the slice before running it.
Like this：TypeError: 'S' has type str, but expected one of: bytes'
The solutions to this problem are listed below in @LiangqunLu question. @chrisdonahue has offered a solution.

That's all.

from wavegan.