f90 / wave-u-net-pytorch Goto Github PK

Traceback (most recent call last):
File "F:/1Python/1_code_xunlian_from_githubOrOther/2_wave_u_net/Wave-U-Net-Pytorch-fuben/train.py", line 243, in
main(args)
File "F:/1Python/1_code_xunlian_from_githubOrOther/2_wave_u_net/Wave-U-Net-Pytorch-fuben/train.py", line 100, in main
utils.set_cyclic_lr(optimizer, example_num, len(train_data) // args.batch_size, args.cycles, args.min_lr, args.lr)
File "F:\1Python\1_code_xunlian_from_githubOrOther\2_wave_u_net\Wave-U-Net-Pytorch-fuben\utils.py", line 16, in set_cyclic_lr
curr_cycle = min(it // cycle_length, cycles-1)
ZeroDivisionError: integer division or modulo by zero

Thanks for such excellent work! The code is great and well-organized. But I met this problem during training.

I tried evaluating the SDR for both pretrained model and the model I trained with default setting, and the best overall SDR was around 2. That was 1. away from the accuracy claimed in the paper.

Is there any important setting I miss? Should I use the original tensorflow implement code to try it again?

Changing the depth to anything other than 1 will result in this error

RuntimeError: Given groups=1, weight of size [512, 512, 5], expected input[1, 1024, 357] to have 512 channels, but got 1024 channels instead

I am pretty certain this is the cause of said error

line 24 implies that the next modules will not take the shortcut (n_outputs as opposed to n_outputs+n_shortcut)
while line 38 adds the shortcut after the first iteration.

It should be either all post shortcut convs to take the added shortcut, or the shortcut is only added once. I think the latter makes more sense, though.

I have download the weights^ but it is not .zip of 7z file, so how can I extract the pre-trained model weights?

As my device does not have an NVIDIA GPU (it is equipped with a Radeon), I had to disable CUDA. But when removing --cuda, I get an exception:

Using valid convolutions with 97961 inputs and 88409 outputs
Loading model from checkpoint /code/checkpoints/waveunet/model
Traceback (most recent call last):
  File "/code/predict.py", line 73, in <module>
    main(args)
  File "/code/predict.py", line 23, in main
    state = utils.load_model(model, None, args.load_model)
  File "/code/utils.py", line 140, in load_model
    checkpoint = torch.load(path)
  File "/usr/local/lib/python3.6/site-packages/torch/serialization.py", line 529, in load
    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
  File "/usr/local/lib/python3.6/site-packages/torch/serialization.py", line 702, in _legacy_load
    result = unpickler.load()
  File "/usr/local/lib/python3.6/site-packages/torch/serialization.py", line 665, in persistent_load
    deserialized_objects[root_key] = restore_location(obj, location)
  File "/usr/local/lib/python3.6/site-packages/torch/serialization.py", line 156, in default_restore_location
    result = fn(storage, location)
  File "/usr/local/lib/python3.6/site-packages/torch/serialization.py", line 132, in _cuda_deserialize
    device = validate_cuda_device(location)
  File "/usr/local/lib/python3.6/site-packages/torch/serialization.py", line 116, in validate_cuda_device
    raise RuntimeError('Attempting to deserialize object on a CUDA '
RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

To solve this, I had to change Line 140 of utils.py from

    checkpoint = torch.load(path)

to

    checkpoint = torch.load(path, map_location='cpu')

This should be done automatically when removing the --cuda flag.

Could you share a pre-trained model? I train several times and still not get a applicable model, or should I assign a random seed? Many thanks!

In this function,when the transpose is False and padding is reflect,the output_size is equal to the input_size.But in get_input_size function under the same conditions:input_size=(output_size - 1)*self.stride + 1 (stride!=1).I think the output_size should equal to the (input_size-1)/self.stride+1 in the DS block using the get_output_size function.

Traceback (most recent call last):
File "train.py", line 19, in
from data.dataset import SeparationDataset
File "/root/autodl-tmp/wave-u-net/data/dataset.py", line 9, in
from data.utils import load
File "/root/autodl-tmp/wave-u-net/data/utils.py", line 3, in
import soundfile
File "/root/miniconda3/lib/python3.8/site-packages/soundfile.py", line 142, in
raise OSError('sndfile library not found')
OSError: sndfile library not found

When I start training, I will report this error

I use IDM to download, but it keeps showing that the download task connection has timed out and I can’t download，what should I do？

Doesn't difference output layer exist in pytorch version?
In tensorflow version I could find 'output_type' parameter, but not in pytorch version.
Is there any reason for not including that layer?

FileNotFoundError: [Errno 2] No such file or directory: 'E:\xxx\Dataset\Dataset\musdb18\train\A Classic Education - NightOwl.stem.mp4\bass.wav'

How can I solve it T_T

Please add support for the MPS backend as you do for cuda:

if torch.backends.mps.is_available():
    mps = torch.device("mps")
    model = model_utils.DataParallel(model)
    model.to(mps)

# ... and so on...
x = x.to(mps)

Here's a problem；

RuntimeError: ffmpeg or ffprobe could not be found! Please install them before using stempeg

But I have both libraries installed

could someone help me? thankyou